Models

GPT-6.1 Sol is OpenAI's latest model in the GPT-6 series, specifically designed for complex programming and agent workflows. GPT-6.1 Sol's performance in complex programming, computer operations, and professional work tasks is close to Astra, but at a lower cost. You can compare the two on your own tasks to weigh quality against cost.

Gemini Nano Banana 2.1 (gemini-nano-banana-2.1) is Google's latest high-efficiency image generation and conversational editing model. As an upgraded version of Nano Banana 2 (Gemini 3.1 Flash Image), it significantly improves visual quality, prompt adherence, role consistency in multi-turn conversations, and text rendering while maintaining Flash-level speed and cost efficiency. It can be regarded as a more efficient counterpart to Gemini 3 Pro Image (Nano Banana Pro).

Claude Haiku 5.5 is Anthropic’s next-generation product, designed for high-throughput, cost-sensitive tasks. It can reliably handle fast and repetitive workloads such as summarization, context compression, database queries, and classification requests. It is also well suited to be used as a sub-agent model in programming workflows alongside Opus 5.5 and Sonnet 5.5.

Claude Opus 5.5 is Anthropic’s latest flagship model. In Anthropic’s comprehensive alignment testing, Opus 5.5 is the strongest-performing model tested to date. It also includes the safety safeguards Anthropic has developed for its most capable models.

Claude Sonnet 5.5 is the next-generation product in Anthropic’s Sonnet model series; Claude Sonnet 5.5 offers the best balance between speed and intelligence. The following five major changes will affect code currently running on Claude Sonnet 5: 1. Using "between_tools" can disable up-front thinking. 2. Forcing the use of tools will result in an error. 3. Thinking content blocks are bound to the model and the conversation. 4. On the Claude API and Google Cloud, the older "computer_20251124" computer tool usage is no longer accepted. 5. The advisor tool will refuse to accept Claude Opus 4.8, Claude Opus 4.7, and Claude Sonnet 5 as advisor models. Use between_tools to disable up-front thinking. Forcing tool use will return an error. Thinking blocks are bound to the model and the conversation. On the Claude API and Google Cloud, early computer_20251124 computer tool usage is no longer accepted. The advisor tool no longer accepts Claude Opus 4.8, Claude Opus 4.7, and Claude Sonnet 5 as advisors.

GPT-6 Luna is OpenAI's latest and most efficient model, designed for focused, high-throughput tasks.

GPT-6 Sol is OpenAI's latest model in the GPT-6 series, specifically designed for complex programming and agent workflows.

Mistral Large 4 is Mistral AI’s frontier multimodal model with text and image input, designed for complex reasoning, coding, and agentic workflows. It supports tool calling and structured outputs, with a 512K-token context window and up to 256K output tokens, making it well suited for long-context and multi-step tasks. Compared with earlier Mistral large models, Mistral Large 4 places greater emphasis on multimodal understanding, extended context, and agentic execution.

Gemini 3.8 Flash-Lite TTS (gemini-3.8-flash-lite-tts) is Google's fast, cost-efficient workhorse text-to-speech model, built to replace gemini-3.1-flash-tts-preview for high-throughput production workloads.

Gemini 3.8 Flash TTS (gemini-3.8-flash-tts) is Google's 3.8 Flash text-to-speech audio model, designed to deliver studio-grade sound fidelity, expressive performance, and long-form stability.

Grok 4.7 is xAI’s flagship model for coding, agentic tasks, and professional knowledge work. Compared with Grok 4.6, it uses a larger base model and longer reinforcement-learning training, improving long-horizon task execution, self-verification, context management, and document and presentation creation.

mimo-v2.6-flash is Xiaomi's latest model series. A fully multimodal, highly intelligent, low-cost efficient inference model, it is the best-balanced choice for high-frequency invocations and large-scale tasks in professional office settings.

MiMo-V2.6-Pro is Xiaomi’s trillion-parameter, natively omni-modal flagship reasoning model for complex projects, long-horizon agentic tasks, coding, research, and professional knowledge work. It supports a 1M-token context window and up to 128K output tokens. Compared with V2.5, it delivers stronger reasoning and long-task performance while retaining the same API pricing.

MiMo-V2.6-Pro-UltraSpeed combines the flagship performance of MiMo-V2.6-Pro and offers up to 20× output speed to serve highly real-time interactive and latency-sensitive production scenarios.

Qwen3.8 Omni Flash is Alibaba Cloud Qwen's next-generation native multimodal model, supporting 1M-length sequences and able to directly accept text, images, audio, and video as input. It is based on the Qwen3.8-Flash-Next architecture. The model is designed for agent capabilities in real productivity scenarios: while offering agentic abilities such as programming, text knowledge work, and GUI operation, it also achieves significant results in audio/video-centered agentic applications—such as video editing, music video creation, film production and narration, audio/video-to-text-and-image summarization, and audio/video dialogue—that require integrated processing of text, images, audio, and video. It supports 2-channel and 4-channel spatial audio parsing, is compatible with DashScope and OpenAI protocols, and it is recommended to install the accompanying Qwen-MM-Plugins to facilitate agent frameworks' access to native multimodal capabilities.

Alibaba Cloud has launched the decision model decision-model-preview. This structured decision model is designed for high-frequency business judgments; it can concurrently perform classification, binary decisions, and scoring based on text or business state, and returns probability distributions and confidence levels. It is suitable for scenarios such as ticket routing, content moderation, agent routing, and result verification. Free for a limited time.

Jev is a structured decision model launched by TypeSafe and is also their first "System One" model. System One models can make structured decisions for software quickly, returning typed selection results rather than freely generated text. They are suitable for routing, classification, and other decision nodes in applications—where fast, predictable answers are more important than generating natural language content.

Jev is a structured decision model launched by TypeSafe and is also their first "System One" model. System One models can make structured decisions for software quickly, returning typed selection results rather than freely generated text. They are suitable for routing, classification, and other decision nodes in applications—where fast, predictable answers are more important than generating natural language content.

Step 5 Preview is StepFun's next-generation flagship base model designed for real-world tasks, with a focus on programming and professional expertise.

Doubao-Seedream-5.0-flash is a fast image-creation model released by ByteDance that balances speed and cost efficiency; it supports precise editing, layer separation, and transparent-background output, making it suitable for high-frequency, large-scale image production and rapid interactive scenarios.

GLM-5.3-FlashX is Z.AI’s high-speed inference model, designed for coding agents, real-time interactions, long-running agentic workflows, and applications that require fast token generation. It is heavily optimized at the inference infrastructure level for lower latency and higher execution efficiency, reaching generation speeds of up to 200 tokens per second. Compared with GLM-5.3-Flash, GLM-5.3-FlashX delivers up to 5× faster inference, with a stronger focus on low-latency and high-throughput workloads.

GPT-6 Astra is OpenAI's newest and most intelligent model, with industry-leading performance in computer operations, web browsing, software engineering, scientific research, and professional work. It excels at executing multi-step workflows across code, browsers, and various professional software. Astra can achieve better results with significantly fewer output tokens, making its estimated API cost per task lower.

jina-ocr-v1 is Jina AI's latest page-to-Markdown document parsing model, with 570 million active parameters. Its architecture inherits the DeepEncoder from DeepSeek-OCR; jina-ocr-v1 retains that encoder and decoder, and thus supports document parsing for approximately 100 languages.(This model is free for a limited time.)

Gemini 3.8 Flash is Google's most intelligent Flash-series model, designed for long-running software engineering tasks, autonomous agents, and complex enterprise workflows, while retaining the Flash series' fast responsiveness and cost-effectiveness.

Claude-fable-5-1 is Anthropic's most capable newly released model, suitable for the most demanding reasoning and long-running agent work. Claude Fable 5.1, while keeping input and output prices the same as Claude Fable 5, reduces cache read prices to one quarter of the original and is more capable for long-running agent programming, multi-step research, and document, spreadsheet, and slide processing. (The model is extremely expensive and not recommended for casual use.)

DeepSeek-V4.1-Flash model official release. This is the smallest model in DeepSeek’s new model-architecture series, featuring native multimodal visual understanding capabilities. The new architecture is designed to deliver a higher capability ceiling, faster inference speeds, greater throughput, and scalability to larger-parameter models.

GLM-5.3-Flash is a high-efficiency multimodal model from Z.AI. It supports a context window of roughly 1 million tokens, along with text, image, and video inputs, and includes tool-calling capabilities. It is primarily designed for coding agents, complex reasoning, and long-horizon software engineering tasks. Built on the existing GLM technology stack, the model has been further post-trained and optimized to deliver strong performance while placing greater emphasis on inference efficiency, responsiveness, and cost.The model is offered at a limited-time 50% discount; users are welcome to try it.

OpenAI's latest realtime speech-to-text model, built for low-latency use — it streams incremental transcripts as audio arrives, ideal for live captions, dictation, and voice interfaces. Supports automatic transcription across 57 languages, with keyword/context hints and a tunable latency-vs-accuracy tradeoff.

xiaomi-mimo-v2.6-pro-free is the open free version of xiaomi-mimo-v2.6-pro. To maintain reliable service, each account is limited to 5 requests per minute, 100 requests per day, and 1 million tokens per day.

GPT-Realtime-2.1 is a reasoning speech-to-speech model for the Realtime API, with tool use and configurable reasoning effort, improving on GPT-Realtime-2 in alphanumeric recognition, silence and noise handling, and interruption behavior.

GLM-5.3 is Z.AI’s coding and agentic reasoning model, built for complex software engineering, long-running agent tasks, vulnerability analysis, and other demanding workloads. Building on GLM-5.2, it incorporates further post-training improvements to deliver stronger coding performance, better task execution, and greater token efficiency. We currently offer the production-ready GLM-5.3 API with unlimited concurrency, making it well suited for high-throughput workloads, coding agents, and large-scale automation. For a limited time, GLM-5.3 is available at 10% off.

Qwen3.8-27b is an Alibaba-released dense vision-language model with open-source weights. Qwen3.8 was developed based on the architecture of Qwen3.5, achieving significant improvements in encoding capability, handling professional tasks, research tasks, and tasks that require long-term completion. Qwen3.8-27B integrates these advantages into a compact, easy-to-deploy dense model — a native vision-language model capable of understanding image and video information and featuring flexible cognitive control capabilities. This model can more reliably accomplish complex, multi-step tasks.

GPT-Image-2.5 Flare is OpenAI's latest image model, the fastest and suited for everyday high-quality image generation. It accepts text and image inputs and produces image outputs. The model supports quality settings: low, medium, high, xhigh, max, and auto.

GPT Image 2.5 Sunburst is OpenAI's latest image model, capable of generating and editing images from text and image inputs. It is suitable for workflows that require extremely high editing precision. The model supports quality settings: low, medium, high, xhigh, max, and auto.

Mercury 2.5 is the latest diffusion-based large language model (dLLM) released by Inception. It is the fastest inference LLM; unlike the sequential token-by-token generation approach, Mercury 2.5 can generate and optimize multiple tokens in parallel, achieving a generation speed of 1,107 tokens per second on standard GPUs. Compared to Mercury 2, its intelligence has increased by more than 10 percentage points, and its quality rivals leading cost-optimized frontier models such as GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5.

Qwen3.8-Max-0902 (also known as qwen3.8-max-2026-09-02) is a snapshot version of Alibaba Cloud Tongyi Qianwen's qwen3.8-max. It pushes encoding depth further, enabling it to handle more complex engineering-level projects and long-term autonomous development; collaborative agent capabilities are significantly enhanced, performing more confidently in multi-tool orchestration and end-to-end delivery; visual understanding is comprehensively improved, with more sensitive and accurate chart reasoning, document parsing, and multimodal perception. Continuing the 1-million-context window, reasoning modes, and a complete tool ecosystem, it continues to evolve at a higher level of intelligence.

Hy4 Preview is Tencent Hunyuan’s large language model for agents, coding, office automation, and complex tool use.Hy4 preview has a total of 770B parameters and 49B active parameters, and is primarily optimized for agent, coding, and productivity scenarios. It strengthens understanding, planning, tool invocation, and sustained execution capabilities for complex tasks. Compared with the previous generation, Hy4 preview further improves multi-step agents, code development, and productivity tasks, offering better task decomposition, context continuity, instruction following, and long-horizon execution. In coding scenarios, it further enhances code understanding, generation, modification, and the handling of complex engineering tasks; in productivity scenarios, it focuses on improving document processing, information analysis, office automation, game development, webpage generation, and cross-tool collaboration. Hy4 preview is suitable for coding agents, complex tool invocation, and various agent workflows that require multi-step planning and sustained execution, providing more reliable task completion for complex real-world business scenarios.

Muse Spark 1.3 is Meta’s multimodal reasoning model, designed for long-running agentic, multi-agent, and complex coding workflows. It can maintain context and task constraints across extended workflows, reconcile conflicting information, and request clarification or confirmation when needed. Compared with Muse Spark 1.2, Muse Spark 1.3 improves efficiency across long-horizon agentic and coding tasks, with fewer unnecessary turns, tool calls, and tokens, while producing more concise outputs.

Qwen3.8 Flash is Alibaba Cloud Qwen’s flagship native vision-language model for coding, office tasks, long-context reasoning, and agent workflows. It supports a 1M-token context, 128K output, web access, and tool calling. Compared with Qwen3.7-Plus, Qwen3.8-Flash significantly reduces training and inference costs—the training overhead is only about one-ninth of the former—while offering stronger capabilities on coding and office tasks.

Gemini 3.7 Flash is Google’s natively multimodal reasoning model for coding, agents, web development, and knowledge work. It supports a 1M-token context window and adjustable thinking levels. Compared with Gemini 3.6 Flash, it improves coding, tool use, multi-step planning, and instruction following.

MAI-Image-2.6 is Microsoft's latest image-generation and editing model. It can generate images from text prompts and supports image-guided editing in multiple aspect ratios.

MAI-Image-2.6 Flash is Microsoft’s low-latency version of the latest image model MAI-Image-2.6. It supports image generation in multiple aspect ratios and image-guided editing.

xiaomi-mimo-v2.6-flash-free is the open free version of xiaomi-mimo-v2.6-flash. To maintain reliable service, each account is limited to 5 requests per minute, 100 requests per day, and 1 million tokens per day.

Agnes 3.0 Flash is designed for real-world agent tasks and development workflows, covering the full execution chain from task understanding and planning to tool invocation and final delivery. The model focuses on improving stability, instruction-following, factual grounding, and output completeness in complex tasks, helping developers build more reliable agent applications.

MiniMax H3 (minimax-h3) is a general-purpose omni-modal generation model developed by the Chinese AI company MiniMax. It can jointly understand text, images, video, and audio while generating videos of up to 15 seconds in 2K resolution with native stereo sound. It is designed for advertising, branding, e-commerce, product design, UI/UX, and gaming. Compared with Hailuo 01 and Hailuo 02, H3 evolves from specialized video generation into a unified model for multimodal creation, reference-based editing, text rendering, and motion transfer.

Agents-A1 is a 35B Mixture-of-Experts (MoE) agentic model developed by InternScience for complex, long-horizon tasks across search, engineering, scientific research, instruction following, and tool use. The model is trained to scale both long-horizon agent trajectories and heterogeneous agent capabilities through a multi-stage training framework that combines full-domain supervised fine-tuning, domain-specialized teacher models, and multi-teacher on-policy distillation. Agents-A1 supports both text and image inputs and is designed for general-purpose agent workflows that require multi-step reasoning, planning, tool interaction, and sustained task execution.

GLM-5.2 is Z.ai’s flagship model for the era of long-horizon tasks. With a truly usable 1M-token context window, it can handle project-level engineering context, execute long-running tasks more reliably, follow engineering standards more consistently, and complete the full development workflow from requirements to multi-platform deployment in a single task.

Intern-S2-Preview is InternLM’s 35B scientific multimodal foundation model, designed for professional scientific reasoning, research agents, multimodal understanding, and complex scientific workflows. It scales the difficulty, diversity, and coverage of scientific tasks across the full training pipeline, while strengthening scientific agent capabilities, time-series modeling, molecular and material structure understanding, and efficient reasoning. Compared with the trillion-parameter Intern-S1-Pro, Intern-S2-Preview achieves comparable performance on multiple core scientific tasks with only 35B parameters, while retaining strong general reasoning and multimodal capabilities.

Developed by Stealth, Union Alpha is a multimodal model built for research, coding, and agentic workflows. It delivers frontier-level performance across a broad range of general-purpose tasks. Featuring a massive 262,144 token context window, this model is highly capable of handling extensive and complex tasks.

DeepSeek-V4-Flash-0731(deepseek-v4-flash-0731) is an open-source MoE large language model developed by the Chinese AI company DeepSeek, with support for a million-token context window. It is designed for coding, complex reasoning, tool use, agentic workflows, and long-document processing. Its advantages include strong performance with fewer active parameters and improved efficiency through DSpark speculative decoding. Compared with DeepSeek V4-Flash Preview, it offers significantly stronger coding and agent capabilities, while outperforming DeepSeek V4-Pro Preview on several benchmarks with fewer active parameters.

DeepSeek’s officially released new multimodal visual-understanding model, DeepSeek‑V4‑Flash‑Vision‑Exp, is experimental in nature and supports multimodal inputs. In pure-text capabilities (agents, reasoning, world knowledge, etc.), DeepSeek‑V4‑Flash‑Vision‑Exp is on par with the official DeepSeek‑V4‑Flash release.(The model name deepseek-v4-flash-vision-exp remains callable, but the official corresponding model has been taken offline; requests will be served by the DeepSeek-V4.1-Flash model and billed at Flash pricing.)

DeepSeek V4 Pro 0813 is DeepSeek’s high-performance general-purpose reasoning and agent model, designed for complex reasoning, coding, long-document analysis, and agentic workflows. It supports thinking and non-thinking modes, a 1M-token context window, up to 384K output, tool calling, and the Responses API. Compared with V4 Flash 0731, Pro prioritizes capability on complex tasks, while Flash focuses on speed, cost efficiency, and high concurrency.

Dots3-Note Preview is an open-weight mixture-of-experts model developed by Dots Studio, featuring 16B active parameters out of 280B total. As the lightest model in the Dots 3 family, it is designed for efficient performance while supporting an expansive context length of 512,000 tokens. This preview version provides an accessible way to experience the capabilities of the Dots 3 architecture.

Grok 4.6 is xAI’s (SpaceXAI) flagship multimodal reasoning model for coding, long-running agents, knowledge work, and interactive application development. It supports image understanding, a 500K context window, tool calling, and structured outputs. Compared with Grok 4.5, it offers stronger multi-step execution, self-verification, coding, and visual project generation.

The Hy3 official version is honed for real-world business scenarios, using a Mixture-of-Experts (MoE) architecture with 295B total parameters and 21B activated parameters. It natively supports a 256K context window and offers multiple thinking modes: no_think (ultra-fast response), think_low (quick thinking), and think_high (deep reasoning), balancing ultra-fast responses, complex reasoning, and invocation cost. Compared with the Preview version, Hy3—based on real business feedback from Tencent Yuanbao, WorkBuddy, ima, Marvis, and others—focuses on improving the Coding Agent, long-form understanding, multi-turn context continuity, search QA, and complex task execution, performing more stably in reducing hallucinations, improving task completion, and engineering usability. It is better suited to practical scenarios such as frontend tasks, cross-file code development, long-document analysis, office automation, and multi-step Agent workflows.

Developed by Minimax, MiniMax-M2.7 is a next-generation large language model designed for autonomous, real-world productivity and continuous improvement. Featuring a generous 196,608-token context window, the model integrates advanced agentic capabilities through multi-agent workflows. It is uniquely built to actively participate in its own evolution, delivering highly adaptable and intelligent performance.

Only supports OpenAI-compatible formats.

DeepSeek V4 Flash 0731 Fast is a high-speed deployment of DeepSeek’s agentic model provided by Wafer, designed for coding, tool use, and high-volume agent workloads. It preserves the capabilities of V4 Flash 0731 while delivering much faster inference. Compared with V4 Pro, it prioritizes latency and execution efficiency.

MAI-Thinking-1 is Microsoft’s first inference model in the MAI series, built for enterprise-scale workloads. With excellent reasoning, mathematical, and general intelligence capabilities, combined with superior cost-effectiveness, it makes high-throughput, 24/7 AI workloads economically viable.