Models

GPT-5.4 Pro is OpenAI’s most capable model, extending GPT-5.4’s unified architecture with stronger reasoning for complex, high-stakes tasks. It includes a 1M+ token context window (922K input, 128K output) and supports both text and image inputs. Tuned for step-by-step reasoning, instruction following, and high accuracy, GPT-5.4 Pro stands out in agentic coding, long-context workflows, and multi-step problem solving.

GPT-5.4 is OpenAI’s newest frontier model that merges the Codex and GPT families into a single unified system. It offers a 1M+ token context window (922K input, 128K output) and supports both text and image inputs, enabling high-context reasoning, coding, and multimodal analysis in one workflow. The model brings stronger performance in coding, document understanding, tool usage, and instruction following. It’s built to be a solid default for both general tasks and software engineering—able to produce production-ready code, synthesize information from multiple sources, and handle complex multi-step workflows with fewer iterations and better token efficiency.

GPT-5.3 Chat is an updated version of ChatGPT’s most widely used model, designed to make everyday conversations smoother, more practical, and more directly helpful. It provides more accurate answers with stronger context awareness and noticeably cuts down on unnecessary refusals, excessive caveats, and overly cautious wording that can disrupt the flow of conversation.

Gemini 3.1 Flash Lite Preview is Google’s high-efficiency model designed for high-throughput, high-volume use cases. It delivers better overall quality than Gemini 2.5 Flash Lite and comes close to Gemini 2.5 Flash performance across core capabilities. Enhancements include audio input/ASR, RAG snippet ranking, translation, data extraction, and code completion. It supports the full range of thinking levels (minimal, low, medium, high) to enable fine-grained cost/performance tuning. Pricing is set at half the cost of Gemini 3 Flash.

Nano Banana 2 (Gemini 3.1 Flash Image) is Google DeepMind’s flagship Flash image model for high-fidelity generation and fast, advanced editing at scale, optimized for price–performance. It follows complex prompts more reliably and adds configurable thinking levels (Minimal vs High/Dynamic) to balance latency and quality. Nano Banana 2 improves in-image text rendering and supports in-image localization (generate/translate text across languages directly in the image), while leveraging stronger world knowledge and web image search for more grounded, realistic outputs. It supports native aspect ratios (including 4:1, 1:4, 8:1, 1:8) and 512px/1K/2K/4K resolutions.

GPT-5.3-Codex is OpenAI’s most advanced agentic coding model. It pairs the frontier software engineering performance of GPT-5.2-Codex with the broader reasoning and professional knowledge capabilities of GPT-5.2. It delivers state-of-the-art results on SWE-Bench Pro and strong performance on Terminal-Bench 2.0 and OSWorld-Verified, highlighting better multi-language coding, terminal fluency, and real-world computer-use skills. The model is tuned for long-running, tool-driven workflows and supports interactive steering during execution, making it well-suited for complex development work, debugging, deployment, and iterative product cycles. Outside of coding, GPT-5.3-Codex also performs well on structured knowledge-work benchmarks such as GDPval, enabling tasks like drafting documents, analyzing spreadsheets, creating slides, and conducting operational research across domains. It is trained with increased cybersecurity awareness, including the ability to identify vulnerabilities, and is deployed with extra safeguards for higher-risk scenarios. Relative to earlier Codex models, it is more token-efficient and about 25% faster, aimed at end-to-end professional workflows that combine reasoning, execution, and computer interaction.

Step 3.5 Flash is our most capable open-source foundation model, designed to deliver frontier-level reasoning and agentic performance with standout efficiency. It uses a sparse Mixture of Experts (MoE) architecture that activates only 11B of its 196B parameters per token, concentrating “intelligence density” to approach top proprietary models while staying fast enough for real-time interaction. Built for rapid, deep reasoning, it’s powered by 3-way Multi-Token Prediction (MTP-3), enabling typical generation speeds of 100–300 tok/s (and up to ~350 tok/s in single-stream coding). For coding and long-horizon agent work, it integrates a scalable RL training framework that supports stable autonomous execution, reaching 74.4% on SWE-bench Verified and 51.0% on Terminal-Bench 2.0. For long-context workloads, Step 3.5 Flash offers a cost-efficient 256K context window via a hybrid attention design with a 3:1 Sliding Window Attention ratio (three SWA layers per one full-attention layer), helping maintain performance on large codebases and massive documents while reducing the compute burden typical of long-context models.

Gemini 3.1 Pro Preview is Google’s frontier reasoning model, offering stronger software engineering performance, more dependable agent behavior, and more efficient token use across demanding workflows. Built on the multimodal foundation of the Gemini 3 series, it delivers high-accuracy reasoning across text, images, video, audio, and code, supported by a 1M-token context window. The 3.1 update brings clear improvements on SWE benchmarks and in real-world coding scenarios, along with more robust autonomous task execution in structured areas like finance and spreadsheet-driven workflows. Created for advanced development and agentic systems, Gemini 3.1 Pro Preview enhances long-horizon stability and tool coordination while further improving token efficiency. It also adds a new medium thinking mode to better balance cost, speed, and quality. The model shines in agentic coding, structured planning, multimodal analysis, and workflow automation—making it a strong fit for autonomous agents, financial modeling, spreadsheet automation, and other high-context enterprise tasks.

Sonnet 4.6 is Anthropic’s strongest Sonnet-class model to date, delivering frontier-level performance in coding, agentic tasks, and professional work. It shines in iterative development, navigating complex codebases, end-to-end project management with memory, creating polished documents, and reliable computer use for web QA and workflow automation.

The Qwen3.5 series 397B-A17B native vision-language model is built on a hybrid architecture that combines a linear attention mechanism with a sparse mixture-of-experts approach, delivering improved inference efficiency. It achieves state-of-the-art results comparable to top-tier models across a broad range of tasks, including language understanding, logical reasoning, code generation, agent-based tasks, image understanding, video understanding, and graphical user interface (GUI) interaction. Thanks to its strong code-generation and agent capabilities, the model demonstrates solid generalization across a wide variety of agents.

Trinity-Large-Preview is a frontier-scale open-weight language model from Arcee, built as a 400B-parameter sparse Mixture-of-Experts with 13B active parameters per token using 4-of-256 expert routing. It delivers exceptional performance in creative writing, storytelling, role-play, and conversational scenarios—surpassing the capabilities of typical reasoning models, particularly in real-time voice assistance. Furthermore, this model introduces advanced agentic capabilities; it is specifically optimized to navigate agent frameworks like OpenCode, Cline, and Kilo Code, while seamlessly managing intricate toolchains and extensive, constraint-heavy prompts. The architecture natively supports massive context windows of up to 512k tokens, though the current Preview API is served at a 128k context using 8-bit quantization for efficient deployment. Trinity-Large-Preview is a testament to Arcee’s efficiency-first design, providing a production-grade frontier model with open weights and permissive licensing tailored for both real-world deployment and rigorous experimentation.

MiniMax-M2.5 is a state-of-the-art large language model built for real-world productivity. Trained across a wide variety of complex digital work environments, M2.5 extends the coding strengths of M2.1 into broader office tasks—becoming fluent in creating and manipulating Word, Excel, and PowerPoint files, seamlessly switching context between different software tools, and collaborating across mixed agent and human teams. It scores 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp. Compared to earlier generations, M2.5 is also more token-efficient, having been trained to plan effectively in order to optimize both its actions and outputs.

GLM-5 is the flagship open-source foundation model from Z.ai, specifically designed for advanced systems engineering and handling long-term agent workflows. Tailored for skilled developers, it provides production-level performance on extensive programming tasks, matching up to the top proprietary models. Thanks to its sophisticated agent planning, robust backend reasoning, and iterative self-improvement, GLM-5 goes beyond simple code generation to enable comprehensive system creation and autonomous operation.

Opus 4.6 is Anthropic’s most advanced model for programming and handling extensive professional tasks. It’s designed for agents that manage entire workflows rather than isolated prompts, making it particularly effective for working with large codebases, implementing complex refactoring, and managing multi-stage debugging processes that evolve over time. Compared to previous versions, this model demonstrates deeper contextual awareness, better problem analysis, and increased reliability in challenging engineering scenarios. In addition to coding, Opus 4.6 excels at sustained knowledge-intensive work. It’s capable of generating near-production-ready documents, comprehensive plans, and thorough analyses in a single go, while maintaining coherence throughout lengthy outputs and extended sessions. This makes it an excellent default choice for tasks that demand persistence, sound judgment, and consistent execution—such as technical architecture, migration strategy planning, and end-to-end project management.

Kimi K2.5 is Moonshot AI’s proprietary multimodal model, offering cutting-edge visual coding abilities and supporting a self-directed agent swarm approach. Developed from Kimi K2 and further trained on around 15 trillion mixed visual and text tokens, it achieves excellent results in general reasoning, visual coding, and autonomous tool invocation.

As a SOTA 30B-class model, GLM-4.7-Flash provides a new option that balances efficiency and performance. It has been further optimized for agentic coding scenarios, enhancing coding abilities, long-term task planning, and tool integration, and has demonstrated leading results among open-source models of its size on multiple current public benchmark leaderboards.

GPT-5.2-Codex is an enhanced version of GPT-5.1-Codex, optimized for software engineering and coding tasks. Designed for both interactive sessions and longer, independent execution, it excels at building projects, developing features, debugging, large-scale refactoring, and code review. Compared to its predecessor, 5.2-Codex follows developer instructions more closely and delivers cleaner, higher-quality code. It integrates seamlessly with developer tools—CLI, IDEs, GitHub, and cloud platforms—and adapts its reasoning effort based on task complexity, providing quick responses for simple tasks and maintaining extended performance for larger projects. The model supports structured code reviews, detects critical flaws, validates behavior against tests, and handles multimodal inputs like images for UI work. It is specifically tailored for agentic coding applications.

MiniMax-M2.1 is a cutting-edge, lightweight large language model designed for coding, agentic workflows, and modern application development. With just 10 billion activated parameters, it offers a significant boost in real-world performance while ensuring low latency, high scalability, and cost-effectiveness. Compared to the previous version, M2.1 delivers more concise, clearer outputs and quicker response times. It excels in multilingual coding, achieving 49.4% on Multi-SWE-Bench and 72.5% on SWE-Bench Multilingual, making it an adaptable engine for IDEs, coding tools, and a wide range of assistant applications.

GLM-4.7 is Z.AI’s newest flagship model, offering advancements in two main aspects: improved programming abilities and greater stability in multi-step reasoning and execution. It shows notable progress in handling complex agent tasks, while also providing more natural conversational experiences and enhanced front-end design.

MiMo-V2-Flash is an open-source foundational language model created by Xiaomi, featuring a Mixture-of-Experts architecture with 309 billion total parameters (15 billion active) and a hybrid attention mechanism. It supports hybrid-thinking, offers a 256K context window, and excels in reasoning, coding, and agent-based tasks. Ranking #1 among open-source models worldwide on SWE-bench Verified and Multilingual benchmarks, MiMo-V2-Flash matches the performance of Claude Sonnet 4.5 at just 3.5% of the cost. For best and fastest results when using agentic tools like Claude Code, Cline, or Roo Code, be sure to disable reasoning mode, as the model is extensively optimized for these scenarios.

Gemini 3 Flash Preview is a high-speed, cost-effective reasoning model built for agent-driven workflows, multi-turn conversation, and coding support. Offering near-Pro level performance in both reasoning and tool use, it stands out by delivering significantly lower latency than larger Gemini versions—making it ideal for interactive development, long-running agent loops, and collaborative programming. Compared to Gemini 2.5 Flash, it features notable improvements in reasoning ability, multimodal comprehension, and overall reliability. The model supports a 1M token context window and handles multimodal inputs—text, images, audio, video, and PDFs—with text-based output. Features like configurable reasoning levels, structured outputs, tool integration, and automatic context caching make it a strong choice for users seeking powerful agentic capabilities without the high cost or lag of more extensive models.

GPT-5.2 Chat (also known as Instant) is the fast and lightweight version of the 5.2 family, built for low-latency chatting while maintaining strong general intelligence. It leverages adaptive reasoning to focus more “thinking” on challenging queries, boosting accuracy in math, coding, and multi-step tasks without sacrificing speed in everyday conversations. The model is naturally warmer and more conversational, with improved instruction following and more stable short-form reasoning. GPT-5.2 Chat is ideal for high-throughput, interactive scenarios where quick response and consistency are more important than in-depth analysis.

GPT-5.2 Pro is OpenAI's most advanced model, featuring significant upgrades in agentic coding and long-context capabilities compared to GPT-5 Pro. It is specifically optimized for handling complex tasks that demand step-by-step reasoning, precise instruction following, and accuracy in critical scenarios. The model supports advanced test-time routing and sophisticated prompt understanding, including user cues like "think hard about this." Key improvements include reduced hallucination and sycophancy, along with stronger performance in coding, writing, and health-related tasks.

GPT-5.2 is the newest frontier-level model in the GPT-5 line, providing enhanced agentic abilities and better long-context performance than GPT-5.1. It employs adaptive reasoning to dynamically distribute computational resources, enabling quick responses to simple requests and deeper analysis for complex challenges. Designed for wide-ranging tasks, GPT-5.2 offers steady improvements in mathematics, programming, science, and tool usage, delivering more coherent long-form responses and increased reliability when using tools.

GPT-5.1-Codex-Max is OpenAI’s newest agentic coding model, created for extended, high-context software development tasks. Built on an enhanced 5.1 reasoning stack, it’s been trained with agentic workflows covering software engineering, mathematics, and research. GPT-5.1-Codex-Max offers faster performance, better reasoning abilities, and increased token efficiency throughout the development process.

Mistral Large 3 2512 is Mistral’s most capable model to date, featuring a sparse mixture-of-experts architecture with 41B active parameters (675B total). Released under the Apache 2.0 license.

DeepSeek-V3.2 is a large language model optimized for high computational efficiency and strong tool-use reasoning. It features DeepSeek Sparse Attention (DSA), a mechanism that lowers training and inference costs while maintaining quality in long-context tasks. A scalable reinforcement learning post-training framework further enhances reasoning, achieving performance comparable to GPT-5 and earning top results on the 2025 IMO and IOI. V3.2 also leverages large-scale agentic task synthesis to improve reasoning in practical tool-use scenarios, boosting its generalization and compliance in interactive environments.

Claude Opus 4.5 is Anthropic’s latest reasoning model, developed for advanced software engineering, complex agent workflows, and extended computer tasks. It features strong multimodal capabilities and performs competitively on real-world coding and reasoning benchmarks, with enhanced resilience against prompt injection. Optimized for efficiency at varying effort levels, it allows developers to balance speed, depth, and token usage according to their specific needs, thanks to a new parameter for controlling token efficiency. Opus 4.5 excels in advanced tool integration, contextual management, and multi-agent coordination, making it ideal for autonomous research, debugging, complex planning, and spreadsheet or browser manipulation. Compared to previous Opus generations, it delivers significant improvements in structured reasoning, long-duration task performance, execution reliability, and alignment, all while reducing token overhead.

Gemini 3 Pro Image Preview (Nano Banana Pro) is Google’s most advanced image generation and editing model, built on Gemini 3 Pro. Building on the original Nano Banana, it offers much improved multimodal reasoning, real-world grounding, and high-fidelity visual synthesis. The model produces context-rich visuals—from infographics and diagrams to cinematic composites—and can incorporate up-to-the-minute information through Search grounding. It leads the industry with sophisticated text rendering in images, handles consistent multi-image blending, and maintains accurate identity preservation for up to five subjects. Nano Banana Pro gives users fine-grained creative controls like localized edits, lighting and focus adjustments, camera transformations, 2K/4K output, and flexible aspect ratios. Tailored for professional design, product visualization, storyboarding, and complex compositions, it remains efficient for everyday image creation needs.

Grok 4.1 Fast Reasoning is xAI's most capable tool-calling model, engineered for production-grade agentic applications with a 2M token context window. Achieving state-of-the-art results on Berkeley Function Calling v4 and leading agentic search benchmarks like Research-Eval Reka (63.9) and FRAMES (87.6), it excels at multi-turn conversations, long-horizon planning, and autonomous task execution. Built through RL training in real-world simulated environments, Grok 4.1 Fast Reasoning delivers exceptional performance on complex enterprise scenarios like customer support and finance while cutting hallucination rates in half compared to its predecessor.

Grok 4.1 Fast Non-Reasoning is xAI's high-speed variant optimized for instant responses and straightforward queries, featuring a 2M token context window. Designed for production workflows requiring rapid inference without deep reasoning overhead, it maintains strong tool-calling capabilities and multi-turn consistency while delivering faster response times. Ideal for real-time applications, customer-facing chatbots, and scenarios where speed is critical, Grok 4.1 Fast Non-Reasoning balances performance with cost-effectiveness for efficient, production-ready agent deployments.

Gemini 3 Pro Preview is Google’s most advanced AI model, setting new records on leading benchmarks like LMArena (1501 Elo), GPQA Diamond (91.9%), and MathArena Apex (23.4%). With powerful reasoning and deep multimodal understanding across text, images, code, video, and audio, Gemini 3 Pro Preview delivers nuanced, context-aware responses and excels at complex problem-solving, scientific analysis, and creative coding tasks.

GPT-5.1-Codex-Mini is a more compact and faster variant of GPT-5.1-Codex.

GPT-5.1-Codex is a specialized version of GPT-5.1 optimized for software engineering and coding workflows. It's designed for both interactive development sessions and long, independent execution of complex engineering tasks. The model supports building projects from scratch, feature development, debugging, large-scale refactoring, and code review. Compared to GPT-5.1, Codex is more steerable, closely follows developer instructions, and produces cleaner, higher-quality code. Codex integrates into developer environments like the CLI, IDE extensions, GitHub, and cloud tasks. It adapts its reasoning dynamically—providing quick answers for small tasks and sustaining long, multi-hour runs for large projects. The model is trained for structured code reviews, identifying critical flaws by reasoning over dependencies and validating behavior against tests. It also supports multimodal inputs like images or screenshots for UI development and integrates tools for search, dependency installation, and environment setup. Codex is specifically intended for agentic coding applications.

GPT-5.1 Chat (also known as Instant) is the fast, lightweight member of the 5.1 family, optimized for low-latency chat while retaining strong general intelligence. It uses adaptive reasoning to selectively “think” on harder queries, improving accuracy on math, coding, and multi-step tasks without slowing down typical conversations. The model is warmer and more conversational by default, with better instruction following and more stable short-form reasoning. GPT-5.1 Chat is designed for high-throughput, interactive workloads where responsiveness and consistency matter more than deep deliberation.

GPT-5.1 is the newest top-tier model in the GPT-5 series, featuring enhanced general reasoning, better instruction following, and a more natural conversational tone compared to GPT-5. With adaptive reasoning, it dynamically adjusts its computational effort—responding swiftly to simple queries and diving deeper into complex tasks. Explanations are now clearer and use less jargon, making challenging topics easier to grasp. Designed for a wide range of tasks, GPT-5.1 consistently improves performance in math, coding, and structured analysis, offering more cohesive long-form responses and more reliable tool usage. Its conversation style is warmer and more intuitive, yet still precise. GPT-5.1 stands as the main, fully capable successor to GPT-5.

Kimi K2 Thinking is Moonshot AI’s most advanced open reasoning model, extending the K2 series into agentic, long-horizon reasoning. Built on a trillion-parameter Mixture-of-Experts (MoE) architecture, it activates 32 billion parameters per forward pass and supports a 256k-token context window. Optimized for persistent step-by-step thought and dynamic tool use, it enables complex reasoning workflows and stable multi-agent behavior across 200–300 tool calls, setting new open-source records on HLE, BrowseComp, SWE-Multilingual, and LiveCodeBench. With MuonClip optimization and large-scale MoE architecture, it delivers strong reasoning depth and high inference efficiency for demanding agentic and analytical tasks.

MiniMax-M2 is a compact, efficient language model with 10B active (230B total) parameters, optimized for coding and agentic workflows. It achieves near-frontier reasoning and tool use with low latency and deployment cost. The model excels in code generation, multi-file editing, compile-run-fix cycles, and automated test repair, showing strong results on SWE-Bench and Terminal-Bench. MiniMax-M2 performs well in agentic benchmarks like BrowseComp and GAIA, handling long-term planning, retrieval, and error recovery. With a small activation footprint, it delivers fast inference and high concurrency, making it ideal for developer tools, agents, and applications that demand cost-effective, responsive reasoning.

Claude Haiku 4.5 is Anthropic’s fastest and most efficient model, offering near-frontier intelligence with much lower cost and latency than larger Claude models. It matches Claude Sonnet 4’s performance in reasoning, coding, and computer-use tasks, making it ideal for real-time and large-scale applications. Haiku 4.5 introduces controllable reasoning depth, supports summarized or interleaved thought outputs, and enables tool-assisted workflows across coding, bash, web search, and computer-use tools. With over 73% on SWE-bench Verified, it stands among the top coding models while maintaining fast responsiveness for sub-agents, parallel execution, and scaled deployment.

Gemini 2.5 Flash Image, also known as "Nano Banana" is a state-of-the-art image generation model with strong contextual understanding. It supports image generation, editing, and multi-turn conversational interactions.

Venice Uncensored is a fine-tuned version of Mistral-Small-24B-Instruct-2501, created by dphn.ai in partnership with Venice.ai. This "uncensored" instruct-tuned LLM is built to give users full control over alignment, system prompts, and model behavior. Designed for advanced and unrestricted scenarios, it prioritizes steerability and transparency, removing the default safety and alignment layers present in most mainstream assistant models.

Amazon Nova Pro 1.0 is a versatile multimodal model from Amazon, designed to balance accuracy, speed, and cost across a wide range of tasks. As of December 2024, it delivers state-of-the-art results on key benchmarks like visual question answering (TextVQA) and financial document analysis. The model excels in processing both visual and textual information, though video input is currently not supported.

Amazon Nova Lite 1.0 is a low-cost multimodal model from Amazon, designed for fast processing of image, video, and text inputs to generate text output. It handles real-time customer interactions, document analysis, and visual question answering with high accuracy. With a 300K token input context, it can process multiple images or up to 30 minutes of video in a single input.

Amazon Nova Micro 1.0 is a text-only model in the Amazon Nova family, optimized for ultra-low latency and cost. With a 128K token context length, it excels at text summarization, translation, content classification, interactive chat, and brainstorming, while also offering basic mathematical reasoning and coding abilities.

Ministral 8B is an 8B parameter model with a unique interleaved sliding-window attention pattern for faster and more memory-efficient inference. Optimized for edge use cases, it supports up to 128k context length and delivers strong performance in knowledge and reasoning tasks. Exceeding other models in the sub-10B category, it is ideal for low-latency and privacy-focused applications.

Mistral-Small-3.2-24B-Instruct-2506 is an updated 24B parameter model from Mistral, optimized for instruction following, repetition reduction, and improved function calling. It supports both image and text inputs, delivers strong performance across coding, STEM, and vision benchmarks, and is designed for efficient, structured output generation.

Building upon Mistral Small 3.2 (2506), with added reasoning capabilities, undergoing SFT from Magistral Medium traces and RL on top, it's a small, efficient reasoning model with 24B parameters.

GLM-4.6 is the latest version in the GLM series, featuring a longer 200K token context window (up from 128K in GLM-4.5) for handling more complex tasks. It offers improved coding performance with higher benchmark scores and better real-world results, including visually enhanced front-end code generation. The model also delivers stronger reasoning, more effective tool use during inference, better integration within agent frameworks, and more refined, human-like writing style compared to GLM-4.5.

Claude Sonnet 4.5 is Anthropic’s most advanced Sonnet model, optimized for real-world agents and coding workflows. It achieves state-of-the-art results on coding benchmarks like SWE-bench Verified, with notable improvements in system design, code security, and following specifications. Designed for extended autonomous operation, the model maintains task continuity across sessions and offers fact-based progress tracking. Sonnet 4.5 features enhanced agentic abilities, such as improved tool orchestration, speculative parallel execution, and more efficient context and memory management. With better context tracking and awareness of token usage across tool calls, it excels in multi-context and long-running workflows. Key use cases include software engineering, cybersecurity, financial analysis, research agents, and other areas requiring sustained reasoning and tool use.

DeepSeek-V3.2-Exp is an experimental large language model from DeepSeek, serving as an intermediate step between V3.1 and future architectures. It features DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism that enhances training and inference efficiency for long-context tasks while preserving high output quality.

Qwen3 Coder Plus is Alibaba's proprietary version of the open-source Qwen3 Coder 480B A35B, designed as a powerful coding agent that excels in autonomous programming through tool use and environment interaction, blending strong coding skills with broad general-purpose capabilities.

Gemini 2.5 Flash-Lite Preview September 2025 Checkpoint is a lightweight, high-throughput model from the Gemini 2.5 family, focused on ultra-low latency and cost efficiency. It delivers even faster token generation, concise output, and improved performance on standard benchmarks compared to earlier Flash-Lite models, making it ideal for large-scale, real-time applications.

Gemini 2.5 Flash Preview September 2025 Checkpoint is Google’s high-performance model, built for advanced reasoning, code generation, mathematical tasks, and scientific applications. This version introduces faster, more efficient output and smarter tool use for complex, multi-step workflows.

Qwen3-VL-235B-A22B Thinking is a multimodal model that combines advanced text generation with visual understanding for images and video, specifically optimized for multimodal reasoning in STEM and math. It delivers robust perception, strong spatial (2D/3D) understanding, and long-form visual comprehension, showing competitive performance in public benchmarks for both perception and reasoning. Beyond analysis, Qwen3-VL supports agentic interaction, tool use, following complex instructions in multi-image dialogues, aligning text with video timelines, and automating GUI operations. The model also enables visual coding workflows, such as turning sketches into code and assisting with UI debugging, while maintaining strong text-only capabilities on par with Qwen3 language models. This makes it ideal for use cases like document AI, multilingual OCR, UI/software help, spatial reasoning, and vision-language agent research.

Qwen3-VL-235B-A22B Instruct is an open-weight multimodal model that combines strong text generation with advanced visual understanding for images and video. Designed for general vision-language tasks like VQA, document parsing, chart and table extraction, and multilingual OCR, the model emphasizes robust perception, spatial (2D/3D) understanding, and long-form visual comprehension, with competitive results on public benchmarks. Qwen3-VL also supports agentic interaction and tool use, following complex instructions in multi-image dialogues, aligning text to video timelines, operating GUIs for automation, and enabling visual coding workflows such as turning sketches into code or debugging UIs. Its strong text-only capabilities match Qwen3 language models, making it suitable for document AI, OCR, UI/software assistance, spatial reasoning, and vision-language agent research.

GPT-5-Codex is a specialized version of GPT-5 tailored for software engineering and coding tasks. It is suitable for both interactive development sessions and the independent execution of complex engineering projects. The model is capable of building projects from scratch, developing new features, debugging, performing large-scale refactoring, and conducting code reviews. Compared to the standard GPT-5, Codex offers greater steerability, follows developer instructions more closely, and delivers cleaner, higher-quality code.

DeepSeek-V3.1 Terminus is an enhanced version of DeepSeek V3.1 that retains the original model’s capabilities while resolving user-reported issues, such as language consistency and agent functionality. The update further refines the model’s performance in coding and search agent tasks. This large-scale hybrid reasoning model (671B parameters, 37B active) supports both thinking and non-thinking modes. Building on the DeepSeek-V3 foundation, it incorporates a two-phase long-context training approach, allowing for up to 128K tokens, and adopts FP8 microscaling for more efficient inference.

GLM-4.5 is the latest flagship foundation model from Z.AI, specifically designed for agent-based applications. It utilizes a Mixture-of-Experts (MoE) architecture and supports context lengths of up to 128k tokens. GLM-4.5 offers significantly improved capabilities in reasoning, code generation, and agent alignment. It features a hybrid inference mode with two options: a "thinking mode," tailored for complex reasoning and tool usage, and a "non-thinking mode," optimized for instant responses.

A blazing fast variant for instant, cost-effective answers without reasoning traces. Built on the same Grok 4 Fast backbone for unified quality and efficiency, it excels at search, summarization, Q&A, and lightweight agent use. Delivers low latency, reduced token cost, and supports the 2M token context for long inputs. Perfect for rapid and scalable information workflows.

State-of-the-art reasoning model optimized for cost-efficient, high-quality chain-of-thought. Trained end-to-end with tool use and agentic search, it matches top-tier benchmarks like AIME, HMMT, and GPQA at 40% lower token use versus Grok 4. Features a huge 2M token context and native web/X browsing. Ideal for agentic workflows, research, code, logic, and complex multi-step tasks. Offers up to 98% cheaper reasoning versus previous models.

Qwen3-Next-80B-A3B-Thinking is a reasoning-focused model that generates structured “thinking” traces by default. Suited for complex multi-step tasks like math proofs, code synthesis, logic, and agentic planning. Compared to earlier Qwen3 models, it’s more stable with long reasoning chains and scales efficiently during inference. Designed for agent frameworks, function calling, retrieval-based workflows, and benchmarks needing step-by-step solutions, it supports detailed completions and faster output through multi-token prediction. Runs only in thinking mode.

Qwen3-Next-80B-A3B-Instruct is an instruction-tuned chat model from the Qwen3-Next series, designed for quick and stable responses without “thinking” traces. It handles complex tasks like reasoning, code generation, knowledge Q&A, and multilingual applications with strong alignment and formatting. Compared to earlier Qwen3 instruct versions, it offers higher throughput and stability, even with long inputs or multi-turn conversations. Ideal for RAG, tool use, and agentic workflows, it delivers consistent and reliable answers with efficient parameter use and fast inference.

Qwen3-Max, the updated model in the Qwen3 series, brings significant advances in reasoning, instruction following, multilingual support, and knowledge coverage compared to the January 2025 version. It offers better accuracy in math, coding, logic, and science, handles complex instructions in Chinese and English more reliably, reduces hallucinations, and gives higher-quality responses in open Q&A and conversations. Supporting 100+ languages, it improves translation and commonsense reasoning, and is optimized for retrieval-augmented generation (RAG) and tool use, though it lacks a specific “thinking” mode.

Kimi K2 0905 is the September update of Kimi K2 0711, a Mixture-of-Experts (MoE) language model from Moonshot AI with 1 trillion parameters and 32 billion active per pass. The long-context window has been expanded to 256k tokens. This release brings improved agentic coding accuracy and generalization across scaffolds, as well as more aesthetic and functional frontend code for web, 3D, and similar tasks. Kimi K2 remains optimized for advanced tool use, reasoning, and code synthesis, excelling in benchmarks like LiveCodeBench, SWE-bench, ZebraLogic, GPQA, Tau2, and AceBench. Its training uses a novel stack with the MuonClip optimizer for stable large-scale MoE training.

Gemini 2.5 Flash Image Preview is a state of the art image generation model with contextual understanding. It is capable of image generation, edits, and multi-turn conversations.

Grok Code Fast 1 is a speedy and economical reasoning model that excels at agentic coding. With reasoning traces visible in the response, developers can steer Grok Code for high-quality work flows.

DeepSeek-V3.1 is a 671B-parameter hybrid reasoning model (37B active), supporting both "thinking" and "non-thinking" modes via prompt templates. It extends DeepSeek-V3 with two-phase long-context training (up to 128K tokens) and uses FP8 microscaling for efficient inference. The model excels in tool use, code generation, and reasoning, with performance comparable to DeepSeek-R1 but with faster responses. It supports structured tool calling, code agents, and search agents, making it ideal for research and agentic workflows. Successor to DeepSeek V3-0324, it delivers strong performance across diverse tasks.

Qwen2.5 72B is the latest in the Qwen large language model series, offering significant improvements in knowledge, coding, and mathematics. It features specialized expert models, improved instruction following, long-text generation (over 8K tokens), structured data understanding, and robust multilingual support for over 29 languages. The model is optimized for resilience to diverse system prompts and enhanced role-play implementation.

Llama 3.2 11B Vision is a multimodal model with 11 billion parameters, designed for tasks combining visual and textual data. It excels at image captioning and visual question answering, bridging the gap between language generation and visual reasoning. Pre-trained on a massive dataset of image-text pairs, it is ideal for content creation, AI-driven customer service, and research.

WizardLM-2 8x22B is Microsoft AI's most advanced Wizard model, demonstrating highly competitive performance compared to leading proprietary models and consistently outperforming state-of-the-art open-source models. It is an instruct finetune of Mixtral 8x22B and is optimized for complex reasoning and instruction-following tasks. For more information, see the [official release](https://wizardlm.github.io/WizardLM2/).

A compact variant of GPT-5, designed for efficient handling of lighter-weight reasoning and conversational tasks. GPT-5 Mini retains the instruction-following and safety features of its larger counterpart, but with reduced latency and cost. It is the direct successor to OpenAI’s o4-mini model, making it ideal for scalable, cost-sensitive deployments.

GPT-5 Chat is tailored for advanced, natural, and context-aware conversations in enterprise environments. It leverages the latest advancements in OpenAI’s conversational AI, supporting multimodal and dynamic dialogue with enhanced context retention and user intent understanding.

The smallest and fastest member of the GPT-5 family, optimized for developer tools, rapid user interactions, and ultra-low latency environments. While it offers limited reasoning depth compared to larger models, GPT-5-Nano preserves essential instruction-following and safety mechanisms. It is the successor to GPT-4.1-nano and is best suited for real-time, cost-sensitive, or embedded applications.

A compact variant of GPT-5, designed for efficient handling of lighter-weight reasoning and conversational tasks. GPT-5 Mini retains the instruction-following and safety features of its larger counterpart, but with reduced latency and cost. It is the direct successor to OpenAI’s o4-mini model, making it ideal for scalable, cost-sensitive deployments.

OpenAI’s most advanced large language model, engineered for high-stakes applications requiring step-by-step reasoning, precise instruction following, and robust code generation. GPT-5 introduces major improvements in factual accuracy, user intent understanding, and hallucination reduction. It supports advanced prompt routing, user-specified intent (such as "think hard about this"), and is optimized for complex workflows in coding, writing, and health-related domains.

OpenAI’s 21B-parameter open-weight Mixture-of-Experts (MoE) model, released under the Apache 2.0 license. Features 3.6B active parameters per forward pass, optimized for low-latency inference and deployability on consumer or single-GPU hardware. Trained in OpenAI’s Harmony response format, it supports reasoning level configuration, fine-tuning, and agentic capabilities such as function calling and structured outputs.

An open-weight, 117B-parameter Mixture-of-Experts (MoE) language model from OpenAI, designed for high-reasoning, agentic, and general-purpose production use cases. Activates 5.1B parameters per forward pass and is optimized for single H100 GPU deployment with native MXFP4 quantization. Supports configurable reasoning depth, full chain-of-thought access, and native tool use, including function calling, browsing, and structured output generation.

Claude Opus 4.1 is Anthropic’s flagship model, offering improved performance in coding, reasoning, and agentic tasks. It achieves high scores on SWE-bench Verified and excels in multi-file code refactoring, debugging precision, and detail-oriented reasoning. The model supports extended thinking up to 64K tokens and is optimized for research, data analysis, and tool-assisted workflows.

Qwen3-235B-A22B-Thinking-2507 is a high-performance, open-weight Mixture-of-Experts (MoE) language model optimized for complex reasoning tasks. Activates 22B of its 235B parameters per forward pass and natively supports up to 262,144 tokens of context. This "thinking-only" variant enhances structured logical reasoning, mathematics, science, and long-form generation, and is instruction-tuned for step-by-step reasoning, tool use, agentic workflows, and multilingual tasks.

Qwen3-Coder-480B-A35B-Instruct is a Mixture-of-Experts (MoE) code generation model developed by the Qwen team. Optimized for agentic coding tasks such as function calling, tool use, and long-context reasoning over repositories. Features 480 billion total parameters, with 35 billion active per forward pass (8 out of 160 experts), and supports variable pricing based on context length.

Qwen3-235B-A22B-Instruct-2507 is a multilingual, instruction-tuned mixture-of-experts language model based on the Qwen3-235B architecture, with 22B active parameters per forward pass. Optimized for general-purpose text generation, including instruction following, logical reasoning, math, code, and tool usage. Supports a native 262K context length and delivers significant gains in knowledge coverage, long-context reasoning, and coding benchmarks.

Gemini 2.5 Flash-Lite is a streamlined reasoning model from the Gemini 2.5 family, designed for extremely low latency and cost-effectiveness. It delivers higher throughput, quicker token generation, and enhanced performance on standard benchmarks compared to previous Flash models.

Gemini 2.5 Flash is Google’s high-performance workhorse model, specifically designed for advanced reasoning, coding, mathematics, and scientific tasks. Includes built-in "thinking" capabilities and is configurable through a "max tokens for reasoning" parameter for fine-tuned performance.

The April 2023 release of GPT-4 Turbo, supporting vision, JSON mode, and function calling. Trained on data up to April 2023, optimized for advanced multimodal tasks.

Kimi K2 Instruct is a large-scale Mixture-of-Experts (MoE) language model developed by Moonshot AI, featuring 1 trillion total parameters with 32 billion active per forward pass. It is optimized for agentic capabilities, including advanced tool use, reasoning, and code synthesis. Kimi K2 excels across a broad range of benchmarks, particularly in coding (LiveCodeBench, SWE-bench), reasoning (ZebraLogic, GPQA), and tool-use (Tau2, AceBench) tasks. It supports long-context inference up to 128K tokens and is designed with a novel training stack that includes the MuonClip optimizer for stable large-scale MoE training.

Grok 4 is xAI’s latest reasoning model, featuring a 256k context window and support for parallel tool calling, structured outputs, and both image and text inputs. Designed for high-throughput, complex reasoning tasks, with pricing that scales for large token requests.

Qwen3, the latest generation in the Qwen large language model series, features both dense and mixture-of-experts (MoE) architectures to excel in reasoning, multilingual support, and advanced agent tasks. Its unique ability to switch seamlessly between a thinking mode for complex reasoning and a non-thinking mode for efficient dialogue ensures versatile, high-quality performance. The Qwen3-30B-A3B variant includes 30.5 billion parameters (3.3 billion activated), 48 layers, 128 experts (8 activated per task), and supports up to 131K token contexts with YaRN, setting a new standard among open-source models.

A lightweight, high-speed model from xAI, engineered for logic-based tasks that do not require deep domain knowledge. Grok-3-mini is optimized for rapid response and efficient reasoning, making it ideal for applications where speed and concise logic are prioritized over extensive context or specialized expertise. The model exposes raw thinking traces, providing transparency into its decision-making process and enabling advanced debugging or educational use cases.

MiniMax-M1 is a large-scale, open-weight reasoning model with 456B total parameters and 45.9B active per token, leveraging a hybrid Mixture-of-Experts (MoE) architecture and a custom "lightning attention" mechanism. It supports context windows up to 1 million tokens and is optimized for long-context understanding, software engineering, agentic tool use, and mathematical reasoning. The model is trained via a custom reinforcement learning pipeline (CISPO) and demonstrates strong performance on FullStackBench, SWE-bench, MATH, GPQA, and TAU-Bench.

Grok 3 is xAI’s flagship model, excelling at enterprise use cases such as data extraction, coding, and text summarization. Possesses deep domain knowledge in finance, healthcare, law, and science, and is optimized for high-accuracy, real-world applications.

Gemini 2.0 Flash Lite is optimized for extremely fast response times and low cost, while maintaining the quality of larger models. Ideal for real-time and large-scale applications.

Gemini 2.5 Flash is Google’s high-performance workhorse model, specifically designed for advanced reasoning, coding, mathematics, and scientific tasks. Includes built-in "thinking" capabilities and is configurable through a "max tokens for reasoning" parameter for fine-tuned performance.

Gemini 2.5 Pro is Google’s state-of-the-art AI model, designed for advanced reasoning, coding, mathematics, and scientific tasks. Employs “thinking” capabilities for nuanced context handling and achieves top-tier performance on multiple benchmarks, including first-place on the LMArena leaderboard.

Mistral Small 3 is a 24B-parameter language model optimized for low-latency performance across common AI tasks. Released under the Apache 2.0 license, it features both pre-trained and instruction-tuned versions for efficient local deployment. The model achieves 81% accuracy on the MMLU benchmark and performs competitively with larger models, while operating at three times the speed on equivalent hardware.

Mistral Saba is a 24B-parameter language model specifically developed for the Middle East and South Asia. It delivers accurate and contextually relevant responses in multiple Indian-origin languages—including Tamil and Malayalam—alongside Arabic. The model is trained on curated regional datasets and is optimized for multilingual and regional applications.

Google’s latest open-source multimodal model, Gemma 3 27B, supports vision-language input and text outputs, handles context windows up to 128k tokens, and understands over 140 languages. Offers improved math, reasoning, and chat capabilities, including structured outputs and function calling.

Qwen-Turbo is a 1M context model based on Qwen2.5, designed for fast speed and low cost. It is suitable for simple tasks and applications where efficiency and affordability are prioritized over deep reasoning.

GPT-4.1, a flagship model for advanced instruction following, software engineering, and long-context reasoning. Supports a 1 million token context window and is tuned for precise code diffs, agent reliability, and high recall in large document contexts.

Llama 4 Scout 17B Instruct (16E) is a mixture-of-experts (MoE) language model from Meta, activating 17 billion parameters out of a total of 109B. It supports native multimodal input (text and image) and multilingual output (text and code) across 12 supported languages. Designed for assistant-style interaction and visual reasoning, Scout uses 16 experts per forward pass and features a context length of 10 million tokens, with a training corpus of ~40 trillion tokens. Built for high efficiency and local or commercial deployment, it is instruction-tuned for multilingual chat, captioning, and image understanding.

GPT-4o (“o” for “omni”) is OpenAI’s latest multimodal model, supporting both text and image inputs with text outputs. Delivers improved performance in non-English languages and visual understanding, while being faster and more cost-effective than previous models.

Qwen3-32B is a dense 32.8B parameter causal language model from the Qwen3 series, optimized for both complex reasoning and efficient dialogue. Supports seamless switching between a "thinking" mode for tasks like math, coding, and logical inference, and a "non-thinking" mode for faster, general-purpose conversation. Demonstrates strong performance in instruction-following, agent tool use, creative writing, and multilingual tasks across 100+ languages and dialects.

Qwen-Max is a large-scale Mixture-of-Experts (MoE) model from Qwen, based on Qwen2.5, and provides the best inference performance among Qwen models, especially for complex multi-step tasks. Pretrained on over 20 trillion tokens and further post-trained with curated Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), it is designed for high-accuracy, high-recall applications. The exact parameter count is undisclosed.

Meta’s Llama 3 70B instruct-tuned model, optimized for high-quality dialogue and demonstrating strong performance in human evaluations. Suitable for advanced conversational AI tasks.

Llama 3.2 3B is a 3-billion-parameter multilingual model optimized for advanced NLP tasks such as dialogue generation, reasoning, and summarization. It supports eight languages and is trained on 9 trillion tokens, excelling in instruction-following, complex reasoning, and tool use.

The latest GPT-4 Turbo model with vision capabilities, supporting JSON mode and function calling. Trained on data up to December 2023, it is optimized for high-throughput, multimodal applications.

Llama 3.2 1B is a 1-billion-parameter language model focused on efficient natural language tasks, including summarization, dialogue, and multilingual text analysis. Its small size allows for deployment in low-resource environments while maintaining strong performance across eight core languages.

Preview release of GPT-4, featuring improved instruction following, JSON mode, reproducible outputs, and parallel function calling. Trained on data up to December 2023. Heavily rate-limited while in preview.

The August 2024 version of GPT-4o, offering improved structured output capabilities, including support for JSON schema in responses. Maintains high intelligence and efficiency, with enhanced non-English and visual performance.

Llama 4 Maverick 17B Instruct (128E) is a high-capacity multimodal language model from Meta, built on a mixture-of-experts (MoE) architecture with 128 experts and 17 billion active parameters per forward pass (400B total). It supports multilingual text and image input, and produces multilingual text and code output across 12 supported languages. Optimized for vision-language tasks, Maverick is instruction-tuned for assistant-like behavior, image reasoning, and general-purpose multimodal interaction. Features early fusion for native multimodality and a 1 million token context window.

Qwen3-14B is a dense 14.8B parameter causal language model from the Qwen3 series, designed for both complex reasoning and efficient dialogue. Supports seamless switching between a "thinking" mode for tasks like math, programming, and logical inference, and a "non-thinking" mode for general-purpose conversation. Fine-tuned for instruction-following, agent tool use, creative writing, and multilingual tasks across 100+ languages and dialects.

DeepSeek Prover V2 is a 671B parameter model, speculated to be geared towards logic and mathematics. Likely an upgrade from DeepSeek-Prover-V1.5, but released without an official announcement or detailed documentation.

The Meta Llama 3.3 multilingual large language model (LLM) is a pretrained and instruction-tuned generative model with 70B parameters. Optimized for multilingual dialogue, it outperforms many open-source and closed chat models on industry benchmarks. Supported languages include English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.

DeepSeek V3 is a 685B-parameter, mixture-of-experts model and the latest iteration of the flagship chat model family from the DeepSeek team. Succeeds the previous DeepSeek V3 model and demonstrates strong performance across a variety of tasks.

DeepSeek-R1-0528 is a lightly upgraded release of DeepSeek R1, utilizing more compute and advanced post-training techniques to push its reasoning and inference capabilities to the level of flagship models like O3 and Gemini 2.5 Pro. Excels in math, programming, and logic leaderboards, with a distilled 8B-parameter variant that rivals much larger models on key benchmarks.

Claude 3.5 Haiku is Anthropic’s fastest model, featuring enhancements across coding, tool use, and reasoning. It is optimized for high interactivity and low latency, making it ideal for user-facing chatbots, on-the-fly code completions, data extraction, and real-time content moderation. The model does not support image inputs.

Mistral Large 2 (version mistral-large-2407) is Mistral AI’s flagship model, supporting dozens of languages—including French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese, and Korean—and over 80 coding languages. It features a long context window for precise information recall and is optimized for reasoning, code, JSON, and chat tasks.

A fine-tuned version of o4-mini, specifically optimized for use in Codex CLI. Recommended for code-related tasks, with improved performance in code generation and completion.

Pixtral Large is a 124B parameter, open-weight, multimodal model built on top of Mistral Large 2. It is capable of understanding documents, charts, and natural images, and is available under both research and commercial licenses. The model is designed for advanced document and image analysis tasks.

A well-rounded, powerful model from OpenAI, setting new standards in math, science, coding, and visual reasoning. Excels at technical writing and instruction-following, and is designed for multi-step problem solving across text, code, and images. BYOK is required for access.

Meta’s Llama 3.1 8B instruct-tuned model, designed for fast and efficient dialogue. It performs strongly in human evaluations and is ideal for applications requiring a balance of speed and quality.

Meta’s Llama 3.1 70B instruct-tuned model, optimized for high-quality dialogue use cases. It demonstrates strong performance in human evaluations and is suitable for a wide range of conversational AI applications.

xAI’s Grok 2 Vision 1212 is a next-generation vision-language model designed for advanced image-based AI applications. It features robust visual comprehension, refined instruction-following, and strong multilingual support. The model excels at object recognition, style analysis, and visual reasoning, empowering developers to build intuitive, visually aware applications. Enhanced steerability and reasoning capabilities make it a solid foundation for next-generation image solutions. For more details, see the official [xAI announcement](https://x.ai/blog/grok-1212).

Qwen3-235B-A22B is a 235B parameter mixture-of-experts (MoE) model developed by Qwen, activating 22B parameters per forward pass. Supports seamless switching between a "thinking" mode for complex reasoning, math, and code tasks, and a "non-thinking" mode for general conversational efficiency. Demonstrates strong reasoning ability, multilingual support (100+ languages and dialects), advanced instruction-following, and agent tool-calling capabilities.

Mythomax L2 13B is one of the highest performing and most popular fine-tunes of Llama 2 13B, known for its rich descriptive capabilities and roleplay performance. It is widely used in creative and narrative-driven applications.

Claude 3 Opus is Anthropic's most powerful model for highly complex tasks, offering top-level performance, intelligence, fluency, and understanding. It is optimized for advanced research, coding, and multimodal applications, and is benchmarked as a leader in the Claude 3 family.

Claude 3.5 Haiku is Anthropic’s fastest model, featuring enhancements across coding, tool use, and reasoning. It is optimized for high interactivity and low latency, making it ideal for user-facing chatbots, on-the-fly code completions, data extraction, and real-time content moderation. The model does not support image inputs.

The highly anticipated 400B class of Llama3 is here, offering a 128k context window and impressive evaluation scores. This 405B instruct-tuned version is optimized for high-quality dialogue and demonstrates strong performance compared to leading closed-source models, including GPT-4o and Claude 3.5 Sonnet.

A compact reasoning model in OpenAI’s o-series, optimized for fast, cost-efficient performance while retaining strong multimodal and agentic capabilities. Supports tool use and demonstrates competitive reasoning and coding performance across benchmarks, outperforming its predecessor o3-mini and approaching o3 in some domains. Well-suited for high-throughput scenarios where latency or cost is critical.

Llama 4 Scout 17B Instruct (16E) is a mixture-of-experts (MoE) language model from Meta, activating 17 billion parameters out of a total of 109B. It supports native multimodal input (text and image) and multilingual output (text and code) across 12 supported languages. Designed for assistant-style interaction and visual reasoning, Scout uses 16 experts per forward pass and features a context length of 10 million tokens, with a training corpus of ~40 trillion tokens. Built for high efficiency and local or commercial deployment, it is instruction-tuned for multilingual chat, captioning, and image understanding.

Claude Opus 4.1 is Anthropic’s flagship model, offering improved performance in coding, reasoning, and agentic tasks. It achieves high scores on SWE-bench Verified and excels in multi-file code refactoring, debugging precision, and detail-oriented reasoning. The model supports extended thinking up to 64K tokens and is optimized for research, data analysis, and tool-assisted workflows.

Phi-4 is a 14B-parameter model from Microsoft Research, designed for complex reasoning tasks and efficient operation in low-memory or rapid-response scenarios. Trained on a mix of high-quality synthetic and curated data, it is optimized for English language inputs and demonstrates strong instruction following and safety standards. For more details, see the [Phi-4 Technical Report](https://arxiv.org/pdf/2412.08905).

The fastest and most cost-effective model in the GPT-4.1 series, designed for tasks demanding low latency such as classification and autocompletion. Maintains a 1 million token context window and delivers exceptional performance at a small size, outperforming even some larger models on key benchmarks.

A mid-sized GPT-4.1 model delivering performance competitive with GPT-4o at substantially lower latency and cost. Retains a 1 million token context window and demonstrates strong coding ability and vision understanding, making it suitable for interactive applications with tight performance constraints.

A mid-sized GPT-4.1 model delivering performance competitive with GPT-4o at substantially lower latency and cost. Retains a 1 million token context window and demonstrates strong coding ability and vision understanding, making it suitable for interactive applications with tight performance constraints.

Mistral Medium 3 is a high-performance, enterprise-grade language model that balances state-of-the-art reasoning and multimodal capabilities with significantly reduced operational cost. It excels in coding, STEM reasoning, and enterprise adaptation, and is optimized for scalable deployments across professional and industrial use cases, including hybrid and on-prem environments.

Claude 3.5 Haiku is engineered for real-time applications, delivering quick response times and enhanced capabilities in speed, coding accuracy, and tool use. It is highly suitable for dynamic environments such as chat interactions, immediate coding suggestions, and customer service bots. The model is currently pointing to Claude 3.5 Haiku (2024-10-22).

Claude Sonnet 4 is a next-generation model from Anthropic, significantly enhancing coding and reasoning capabilities over its predecessor. It achieves state-of-the-art performance on SWE-bench, balances capability and computational efficiency, and is optimized for both routine and complex software development projects. Key features include improved codebase navigation, reduced error rates, and increased reliability in following intricate instructions.

Llama Guard 4 is a multimodal content safety classifier derived from Llama 4 Scout, fine-tuned for both prompt and response classification. It supports content moderation for English and multiple languages, including mixed text-and-image prompts. The model is aligned with the MLCommons hazards taxonomy and is integrated into the Llama Moderations API for robust safety classification in text and images.

A cost-efficient language model from OpenAI, optimized for STEM reasoning tasks, especially in science, mathematics, and coding. Supports the `reasoning_effort` parameter for adjustable thinking time and features significant improvements over its predecessor, with better performance on complex questions and lower latency and cost.

Mistral-Small-3.2-24B-Instruct-2503 is an updated 24B parameter model from Mistral, optimized for instruction following, repetition reduction, and improved function calling. It supports both image and text inputs, delivers strong performance across coding, STEM, and vision benchmarks, and is designed for efficient, structured output generation.

Command A is an open-weights 111B parameter model from Cohere, featuring a 256k context window and optimized for agentic, multilingual, and coding use cases. It delivers high performance with minimal hardware costs, excelling in business-critical workflows that require advanced reasoning, tool use, and language understanding across multiple languages.

QwQ-32B is the medium-sized reasoning model in the Qwen series, designed for advanced thinking and reasoning tasks. It achieves competitive performance against state-of-the-art models like DeepSeek-R1 and o1-mini, and is particularly strong on hard problems requiring deep analytical skills.

Codestral is Mistral’s cutting-edge language model for coding, specializing in low-latency, high-frequency tasks such as fill-in-the-middle (FIM), code correction, and test generation. It is optimized for developer productivity and supports a wide range of programming languages and code-related tasks.

Mistral Large 2 2411 is an updated release of Mistral Large 2, featuring notable improvements in long context understanding, a new system prompt, and more accurate function calling. It is designed for advanced enterprise and research applications requiring high reliability and performance.

Claude 3.7 Sonnet is an advanced large language model from Anthropic, featuring improved reasoning, coding, and problem-solving abilities. It introduces a hybrid reasoning approach, allowing users to choose between rapid responses and extended, step-by-step processing for complex tasks. The model excels in agentic workflows, front-end development, and full-stack updates, and offers an extended reasoning mode for enhanced accuracy in math, coding, and instruction-following.

OpenAI’s most advanced small model, GPT-4o mini, supports both text and image inputs with text outputs. It is highly cost-effective, achieving SOTA intelligence and outperforming larger models on key benchmarks, making it ideal for scalable, interactive applications.

The Meta Llama 3.3 multilingual large language model (LLM) is a pretrained and instruction-tuned generative model with 70B parameters. Optimized for multilingual dialogue, it outperforms many open-source and closed chat models on industry benchmarks. Supported languages include English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.

The November 2024 release of GPT-4o, featuring enhanced creative writing, more natural and engaging responses, and improved file handling. Maintains the intelligence of GPT-4 Turbo while being twice as fast and 50% more cost-effective, with better support for non-English languages and visual tasks.

Gemini Flash 2.0 offers significantly faster time to first token (TTFT) compared to previous versions, while maintaining quality on par with larger models. Introduces enhancements in multimodal understanding, coding, complex instruction following, and function calling for robust agentic experiences.

A flagship large language model from OpenAI, optimized for advanced instruction following, real-world software engineering, and long-context reasoning. Supports a 1 million token context window and outperforms GPT-4o and GPT-4.5 in coding (54.6% SWE-bench Verified), instruction compliance (87.4% IFEval), and multimodal understanding. Tuned for precise code diffs, agent reliability, and high recall in large document contexts, making it ideal for agents, IDE tooling, and enterprise knowledge retrieval.