Models
The models the published price list prices, on one balance. Narrow by modality or provider, search the catalogue, and switch without switching accounts
Anthropic: Claude Opus 5.5
Anthropic Claude Opus 5.5 text model. Available over the Anthropic Messages and Chat Completions protocols.
Anthropic: Claude Sonnet 5
Anthropic Claude Sonnet 5 text model. Available over the Chat Completions, Responses and Anthropic Messages protocols.
Anthropic: Claude Sonnet 5.5
Anthropic Claude Sonnet 5.5 text model. Available over the Anthropic Messages and Chat Completions protocols.
Anthropic: Claude Fable 5.1
Anthropic Claude Fable 5.1 text model. Available over the Anthropic Messages and Chat Completions protocols.
xAI: Grok 4.6
xAI Grok 4.6 text model. Available over the Chat Completions and Responses protocols.
Anthropic: Claude Opus 5
Anthropic Claude Opus 5 text model. Available over the Anthropic Messages and Chat Completions protocols.
OpenAI: GPT-5.6 Luna
OpenAI GPT-5.6 Luna text model. Available over the Chat Completions protocol.
OpenAI: GPT-5.6 Terra
OpenAI GPT-5.6 Terra text model. Available over the Chat Completions protocol.
OpenAI: GPT-5.5
OpenAI GPT-5.5 text model. Available over the Chat Completions protocol.
Anthropic: Claude Haiku 4.5
Anthropic Claude Haiku 4.5 text model. Available over the Anthropic Messages and Chat Completions protocols.
Anthropic: Claude Haiku 5.5
Anthropic Claude Haiku 5.5 text model. Available over the Anthropic Messages and Chat Completions protocols.
OpenAI: GPT-6 Sol
OpenAI GPT-6 Sol text model. Available over the Chat Completions protocol.
xAI: Grok 4.7
xAI Grok 4.7 text model. Available over the Chat Completions and Responses protocols.
OpenAI: GPT-6 Astra
OpenAI GPT-6 Astra text model. Available over the Chat Completions protocol.
Zhipu: GLM-5.3
Zhipu's flagship GLM-5.3 model for complex software engineering and agent tasks, including code review and vulnerability hunting. It has a 1M-token context, reasoning and tool calling.
OpenAI: GPT-5.6 Sol
OpenAI GPT-5.6 Sol text model. Available over the Chat Completions protocol.
xAI: Grok 4.5
xAI Grok 4.5 text model. Available over the Chat Completions and Responses protocols.
Zhipu: GLM-5.2
Zhipu's previous GLM-5 flagship for coding and agent scenarios with reasoning. It has a 1M-token context and tool calling.
Anthropic: Claude Fable 5
Anthropic Claude Fable 5 text model. Available over the Anthropic Messages and Chat Completions protocols.
Anthropic: Claude Opus 4.8
Anthropic Claude Opus 4.8 text model. Available over the Chat Completions, Responses and Anthropic Messages protocols.
Anthropic: Claude Opus 4.7
Anthropic Claude Opus 4.7 text model. Available over the Anthropic Messages and Chat Completions protocols.
Anthropic: Claude Sonnet 4.6
Anthropic Claude Sonnet 4.6 text model. Available over the Anthropic Messages and Chat Completions protocols.
Anthropic: Claude Opus 4.6
Anthropic Claude Opus 4.6 text model. Available over the Anthropic Messages and Chat Completions protocols.
Google: Gemini 3 Flash Preview
Google Gemini 3 Flash Preview text model. Available over the Chat Completions protocol.
Anthropic: Claude Opus 4.5
Anthropic Claude Opus 4.5 (2025-11-01) text model. Available over the Anthropic Messages and Chat Completions protocols.
Anthropic: Claude Sonnet 4.5
Anthropic Claude Sonnet 4.5 (2025-09-29) text model. Available over the Anthropic Messages and Chat Completions protocols.
OpenAI: Codex Auto Review
OpenAI Codex Auto Review text model. Available over the Chat Completions protocol.
Google: Gemini 3.7 Flash High
Google Gemini 3.7 Flash High text model. Available over the Chat Completions protocol.
Qwen: Qwen Image 2.1 Pro
The higher-quality tier of Alibaba's Qwen Image 2.1 series for text-to-image generation, with more quality and detail than Turbo. Up to 2K resolution with many aspect ratios.
Qwen: Qwen Image 2.1 Turbo
The fast, low-cost tier of Alibaba's Qwen Image 2.1 series for text-to-image generation, when throughput and price matter more than peak quality. Up to 2K resolution with many aspect ratios.
Perplexity: Decider Flash V1.2
A 4B decision model from Perplexity, the smaller, lower-cost sibling of Decider V1.2 with the same API contract. It reads a state of text or images and returns calibrated probabilities for typed questions.
Perplexity: Decider V1.2
A 27B checkpoint of Perplexity's decision model, successor to Decider V1.1. Instead of generating text, it reads content passed as state — text or images — and returns calibrated probabilities for typed questions.
Respan: Span-02
An open-weight decision model from Respan. It reads a state of text and images and answers the questions you define about it — picking one of several options, a yes/no probability or a score — with calibrated probabilities.
Liquid AI: D1 Ultra
Liquid AI's largest decision model, served as a System One endpoint. It reads a state of text and images with typed questions and returns a choice, a score or a yes/no probability, with more accuracy than d1 on harder judgements.
TypeSafe: Jev 1.14
The latest version of Jev, TypeSafe's structured decision model. It reads a state of text or images with typed questions and returns calibrated probabilities for a choice, a score or a yes/no answer, built for fast decisions inside software.
ScaleDown: Classify-1
A structured decision model from ScaleDown. It scores text, images or a document against caller-defined choice, yes/no and ordered-scale questions and returns calibrated probabilities; it holds a 1M-token state.
Google: Gemini 4 Argon
Google's flagship reasoning model for complex software engineering, deep math and science, and long-horizon autonomous agents. It has a 1M-token context, reads text, images, video, audio and PDF, and supports reasoning at low, medium and high effort, tool calling and structured output.
Nace.AI: Drex v2.0
A 35B mixture-of-experts decision model (3B active) from Nace.AI, the larger successor to Drex v1.5. Given a state (text or JSON) and typed yes/no, multiple-choice or ordinal questions, it returns calibrated probabilities.
Cloudflare: Clef Omni
The mixture-of-experts member of Cloudflare's open-source Clef decision family, a fine-tune of Qwen3-Omni-30B-A3B (30B total, 3B active) served on Workers AI. It decides over states of text, images, audio and video and returns calibrated probabilities for typed questions.
Nace.AI: Drex v1.5
A 9B decision model from Nace.AI. Given a state (text or JSON) and typed questions — yes/no, multiple choice or ordinal score — it returns a calibrated probability for each answer.
Microsoft: Microsoft-Decision-1
A small Microsoft model built for fast decision-making. Instead of generating text, it reads the provided content and returns a calibrated probability for each fixed answer option.
Upstage: Solar Decide Flash
Upstage's low-latency structured decision model, a faster variant of Solar Decide built on Solar Mini 4 and served through System One. It returns a choice, a score or a yes/no probability for typed questions over a 512K-token state.
Qwen: Qwen3.8 27B Antislop
Qwen3.8 27B fine-tuned by RawModels with Antislop: it writes in natural language, without the usual AI clichés and stock phrases. It reasons before answering and supports tool calling; 262K-token context, text only. Models run on GPUs that are started on demand, so the first request after a quiet spell waits for a worker to boot — usually one to five minutes. Everything after that answers in seconds.
Cloudflare: Clef
Cloudflare's open-source 27B multimodal decision model, a fine-tune of Qwen3.8-27B served on Workers AI. Instead of writing text, it reads a state (text, structured JSON or images) and answers typed questions with calibrated probabilities: yes/no, a choice between options, or a score.
Cloudflare: Clef Flash
The fast 9B member of Cloudflare's open-source Clef decision family, a fine-tune of Qwen3.5-9B served on Workers AI. It reads a state of text, JSON or images and returns calibrated probabilities for yes/no, choice and score questions at low latency.
Liquid AI: D1
Liquid AI's structured decision model, served as a System One endpoint. Send a state of text or images with typed questions and it returns a choice, a score or a yes/no probability instead of generated text.
Inception: Mercury Decide
Inception's structured decision model, served as a System One endpoint. Send a state of text or images with typed questions and it returns a choice, a score or a yes/no probability instead of generated text.
Together AI: Tev1 4B Experimental
An experimental decision model from Together AI, a supervised fine-tune of Qwen3.5-4B trained to choose one option from a structured state, a question and a list of 2 to 24 options, returning a probability for each.
OpenAI: GPT-6.1 Sol
OpenAI GPT-6.1 Sol text model. Available over the Chat Completions and Responses protocols.
Upstage: Solar Decide
Upstage's structured decision model on Solar Mini 4, served as a System One endpoint. Send a state with typed questions and it returns a choice, a score or a yes/no probability; it holds a 512K-token state.
Respan: Span-01
Respan's behaviour scoring model. It reads a conversation span and returns, for each plain-language behaviour you define, the probability that the behaviour is present.
Jared Palmer: Kev 4B
A small open-weight decision model from Jared Palmer, built as a LoRA adapter and pointer head on Qwen3.5-4B-Base and served over the same System One contract as TypeSafe's Jev. It returns calibrated probabilities for typed questions over a short text state.
OpenAI: GPT-6 Luna
OpenAI GPT-6 Luna text model. Available over the Chat Completions protocol.
TypeSafe: Jev 1.13
TypeSafe's structured decision model and the first of its System One models. It makes fast, structured decisions for software: given a state and typed questions it returns calibrated probabilities for a choice, a score or a yes/no answer rather than free text.
Zhipu: GLM-5.3 Raw
Uncensored GLM-5.3 from Z.ai: RawModels removed its built-in refusals with abliteration, so it answers requests ordinary models decline, including cybersecurity and penetration-testing tasks. Reasoning and tool calling are kept; 1M-token context. Models run on GPUs that are started on demand, so the first request after a quiet spell waits for a worker to boot — usually one to five minutes. Everything after that answers in seconds.
DeepSeek: DeepSeek V4.1 Flash
The economy tier of DeepSeek's V4.1 family: a sparse mixture-of-experts model on a 552B-parameter backbone that activates about 8B parameters for input and 16B for output. Image understanding is trained in from the start, and a compressed KV cache lowers the cost of long agent sessions. It fits agents that write code, drive a terminal or operate a computer, as well as long-document analysis, with a 1M-token context and tool calling.
Inception: Mercury 2.5
A diffusion language model from Inception: instead of writing tokens one after another, it drafts and refines many tokens in parallel, reaching over 1,000 tokens per second on standard GPUs. It offers adjustable reasoning, parallel tool calls and schema-conforming JSON, and suits latency-sensitive uses such as search agents, voice pipelines and coding sub-agents. Text only, with a 260K-token context.
Qwen: Qwen3.8 Max 0902
The September snapshot (0902) of Qwen3.8 Max, the largest model of Alibaba's Qwen team: a mixture-of-experts with about 2.4 trillion parameters. It takes text, images and video, reasons by default and was post-trained for multi-step software projects, orchestration of many tools and analysis of charts, long documents and long video. Context is 1M tokens; tool calling, structured output and adjustable reasoning effort are available.
Google: Gemini 3.8 Flash
Google Gemini 3.8 Flash text model. Available over the Chat Completions and Responses protocols.
Google: Gemini 3.8 Flash High
Google Gemini 3.8 Flash High text model. Available over the Chat Completions protocol.
Qwen: Qwen3.8 27B
An open-weight dense 27B-parameter vision-language model from Qwen. It covers coding, research, everyday professional work and long agent runs, and its thinking mode can be switched on or off per request. Accepts text, images and video, with a 1M-token context and tool calling.
Qwen: Qwen Image 3
Alibaba's Qwen Image 3 image-generation model: it renders text and details as small as 10px precisely and draws on broad world knowledge. Up to 2K resolution.
Qwen: Qwen Image 3 Pro
The higher-quality tier of Alibaba's Qwen Image 3: precise rendering of text and small details, richer world knowledge and higher quality than Qwen Image 3. Up to 2K resolution.
Google: Gemini 3.6 Flash
Google Gemini 3.6 Flash text model. Available over the Chat Completions and Responses protocols.
Google: Gemini 3.6 Flash High
Google Gemini 3.6 Flash High text model. Available over the Chat Completions protocol.
Moonshot AI: Kimi K3
Moonshot AI's flagship mixture-of-experts model with a 1M-token context. It reads text, images and video and is built for long coding tasks, end-to-end agent work and deep reasoning, with tool calling and structured output.
Google: Gemini 3.1 Flash Lite
Google Gemini 3.1 Flash Lite text model. Available over the Chat Completions protocol.
DeepSeek: DeepSeek V4 Flash
A throughput-oriented DeepSeek model (April snapshot 0423): mixture-of-experts with 284B total and 13B active parameters, plus hybrid attention for cheaper long inputs. It fits coding assistants, chat products and agents where latency and price matter, offers reasoning levels up to xhigh, a 1M-token context and tool calling.
OpenAI: GPT Image 2
OpenAI GPT Image 2 image-generation model. Available over the Images protocol.
OpenAI: GPT Image 1
OpenAI GPT Image 1 image-generation model. Available over the Images protocol.
OpenAI: GPT Image 1.5
OpenAI GPT Image 1.5 image-generation model. Available over the Images protocol.
OpenAI: GPT Image 2.5
OpenAI GPT Image 2.5 image-generation model. Available over the Images protocol.
OpenAI: GPT Image 2.5 Flare
OpenAI GPT Image 2.5 Flare image-generation model. Available over the Images protocol.
OpenAI: GPT Image 2.5 Sunburst
OpenAI GPT Image 2.5 Sunburst image-generation model. Available over the Images protocol.
RawModels: OpenJev
RawModels' open decision model on Qwen3.5 4B (a cross-encoder), trained as an open counterpart of Jev. It generates nothing: it reads a state — text, JSON, an image or a video — and answers typed questions with probabilities: yes/no, a choice between options or a score on a scale. The same request always returns the same answer; 32K-token context. Models run on GPUs that are started on demand, so the first request after a quiet spell waits for a worker to boot — usually one to five minutes. Everything after that answers in seconds.
RawModels: OpenJev Small
A lighter OpenJev on Qwen3.5 0.8B: faster and half the price, weaker on multi-step policies and on reading values off images. It reads a state — text, JSON, an image or a video — and answers typed questions with probabilities: yes/no, a choice between options or a score. 32K-token context. Models run on GPUs that are started on demand, so the first request after a quiet spell waits for a worker to boot — usually one to five minutes. Everything after that answers in seconds.
xAI: Grok Imagine Lite
xAI Grok Imagine Lite image-generation model. Available over the Images protocol.
StepFun: Step 5 Preview
A preview of StepFun's top agent model: a sparse mixture-of-experts with 600B total and 27B active parameters. It is aimed at software engineering and professional knowledge work, finance in particular, where it iterates with tools across large codebases and document sets. Reads text, images and video, with a 1M-token context and tool calling.
Mistral AI: Mistral Large 4
Mistral AI's large general model for reasoning, programming and agent pipelines. It reads text and images, holds a 1M-token context and can produce up to 256K tokens in one reply. Tool calling and structured JSON output are supported.
AionLabs: Aion 3.5
A role-play and fiction system from AionLabs built on GLM models, in which several specialised models jointly compose each reply to give stories tighter structure and more dramatic tension. Text only, with a 256K-token context and tool calling.
Cohere: Command A+
Cohere's top model for enterprise agents. It reads text and images, offers a 192K-token context, native tool calling with strict schemas, JSON object and JSON schema output, optional reasoning and adjustable safety modes.
Qwen: Qwen3.8 Omni Flash
A fast omni-modal reasoning model from Alibaba's Qwen3.8 line, the first Qwen model built around agent capabilities with native audio and video understanding. It suits audio and video analysis and summarisation, image work and low-cost agent tasks. Context is 1M tokens, with tool calling and structured output.
Xiaomi: MiMo V2.6 Pro
Xiaomi's largest model, with more than 1T parameters, intended for the heaviest agent workloads. It natively handles text, images, video and audio and is tuned for coding, visual, research and general tasks that run over many steps in different agent frameworks. Context is about 1M tokens; tool calling is supported.
Xiaomi: MiMo V2.6 Flash
An open-source Xiaomi model built as a mixture-of-experts with 309B total and 15B active parameters and hybrid attention for efficiency. Like the Pro version it natively takes text, images, video and audio and targets multi-step agent work in coding, vision and research, at a lower price. Context is about 1M tokens; tool calling is supported.
Zhipu: GLM-5.3 FlashX
A speed-tuned edition of Zhipu's GLM-5.3 Flash that generates up to about 200 tokens per second. It shares the base architecture of hybrid sparse and linear attention with 320B total and 18B active parameters, and reads text, images and video. Suited to quick coding, visual tasks and long agent runs, with a 1M-token context and tool calling.
Moonshot AI: Kimi K2.8
Moonshot AI's Kimi K2.8 Preview, a coding model that Moonshot positions close to Kimi K3. It has a context of about 1M tokens, reads text, images and video, and supports reasoning at low, high and max effort (max by default) and tool calling.
IBM: Granite 4.2 8B
A small dense 8B-parameter reasoning model from IBM for mathematics, code, multilingual dialogue and agent tasks with several steps. It can reason fully, with low effort or not at all, and supports 12 languages including English, German, Japanese, Arabic and Chinese. Text only, with a 128K-token context and tool calling.
Tencent: Hy4 Preview
A preview of Tencent's Hy4, a mixture-of-experts text model with 770B total and 49B active parameters. It is aimed at coding agents, workflows with many tool calls and productivity tasks that need planning and a consistent context over many steps. Context is 1M tokens; tool calling is supported.
Zhipu: GLM 5.3 Flash
A fast, low-cost Zhipu GLM-5.3 model and the first natively multimodal one in the line: it reads text, images and video. Suited to coding, visual interface work and documents, with a 1M-token context, reasoning and tool calling.
Qwen: Qwen3.8 Flash
A fast, low-cost reasoning model from Alibaba's Qwen3.8 line that reads text, images and video. It is used for coding help, agents, document and repository analysis, screen interaction and long-video understanding. Context is 1M tokens and tool calling is supported.
DeepSeek: DeepSeek V4 Pro 0813
The generally available August 2026 snapshot (0813) of DeepSeek V4 Pro, the larger mixture-of-experts model of the V4 line. It is a general text model for reasoning, coding and agent tasks, with a 1M-token context and support for tool calling.
ByteDance Seed: Seed 2.1 Turbo
ByteDance Seed's multimodal model for programming and extended agent workflows. It can carry software work from plan to delivery, debug and correct itself along the way, and understand images and video. Context is 256K tokens; tool calling is supported.
ByteDance Seed: Seed 2.0 Code
A ByteDance Seed model tuned for agentic programming, with frontend work and code in many languages as its focus. It is intended to run inside coding agents like Claude Code, Kilo or OpenCode. Reads text, images and video, with a 256K-token context and tool calling.
Upstage: Solar Pro 4
An affordable large language model from Upstage with a 512K-token context. It is built for long tasks and agents, with an emphasis on office productivity, document-heavy work and coding. Text only; tool calling is supported.
Meta: Muse Glimmer 30B
Meta Superintelligence Labs' dense open-weight multimodal model with 30B parameters; it is a distilled version of Muse Spark, made so that agents can run on consumer hardware. It handles multi-step reasoning, tool use, recovery after errors and image input, and works in over 100 languages. Context is 128K tokens; tool calling is supported.
Poolside: Laguna S 2.1
Poolside's coding-agent model with 118B total and 8B active parameters, intended for software engineering and agentic programming. The weights are open under the OpenMDW-1.1 license. Text only, with a 1M-token context and tool calling.
Meituan: LongCat 2.0
A sparse mixture-of-experts text model from Meituan with 1.6T total and 48B active parameters. It is used for coding, changes that span a whole repository, long problem-solving sessions and agent workflows. Context is 1M tokens; tool calling is supported.
Qwen: Qwen3.8 Max Preview
The preview of Qwen3.8 Max that Alibaba showed at WAIC in July 2026: a mixture-of-experts with about 2.4 trillion parameters and Qwen's first multimodal model above a trillion parameters. It reads text, images and video, reasons, and targets coding and complex productivity work. Context is about 1M tokens and tool calling is supported.
Thinking Machines: Inkling
Thinking Machines Lab publishes the weights of Inkling, a multimodal MoE network with 975B total and 41B active parameters. It is a general model for reasoning, coding, agents, retrieval-augmented generation and multilingual chat, and it natively understands images and audio. Context is 512K tokens; tool calling is supported.
Moonshot AI: Kimi K2.7 Code
The coding variant of Moonshot AI's Kimi K2 family, meant to finish programming tasks end to end over long contexts. It is a multimodal mixture-of-experts with about 1T parameters, 32B of them active, accepts text and images and always works in thinking mode, keeping its reasoning across turns. Context is 256K tokens; tool calling is supported.
NVIDIA: Nemotron 3 Ultra
The largest open model of NVIDIA's Nemotron 3 line: its mixture-of-experts design combines Transformer and Mamba layers, with 550B total and 55B active parameters. It is built for planning-heavy agent pipelines such as orchestrating other agents, coding agents and deep research, with throughput suited to high request volumes. Text only; here it is served with a 256K-token context and tool calling.
Qwen: Qwen3.7 Plus
A mid-priced model of Alibaba's Qwen3.7 series with upgraded vision. Besides coding and tool use, it can read screenshots and operate graphical interfaces, turn visual mockups into code and navigate mobile apps step by step. Accepts text and images, with a 1M-token context and tool calling.
MiniMax: MiniMax M3
MiniMax's multimodal base model for long agent sessions, coding and tool use; it reads text, images and video. MiniMax Sparse Attention picks only relevant KV blocks instead of attending to everything, which makes 1M-token inputs far cheaper and faster to process than in the previous generation. Supports tool calling.
StepFun: Step 3.7 Flash
An efficient multimodal mixture-of-experts from StepFun: 196B parameters in the language part plus an image and video encoder, with roughly 11B active per token. Callers choose high, medium or low reasoning to balance speed, price and depth. Fits coding, agents, structured output and long-context office work; reads text, images and video, with a 256K-token context and tool calling.
Qwen: Qwen3.7 Max
The flagship model of Alibaba's Qwen3.7 line, working with text. It is built for agent workloads and is strongest in coding, office and productivity tasks. Context is 1M tokens, with reasoning, tool calling and structured output.
Mistral AI: Mistral Medium 3.5
A dense 128B-parameter open-weight model from Mistral AI aimed at agents, coding and multi-step reasoning. It accepts text and images, has a 256K-token context and lets the caller set reasoning effort per request. Its main strength is dependable calling of several tools within one long task.
Moonshot AI: Kimi K2.6
A multimodal Moonshot AI model for long coding tasks, building user interfaces from prompts or images, and coordinating many sub-agents in parallel. It writes Python, Rust and Go and is able to produce documents, websites or spreadsheets in a single run. Accepts text and images, with a 256K-token context and tool calling.
Arcee AI: Trinity Large Thinking
An open-source reasoning model from Arcee AI designed for agent workloads. It works best when its earlier reasoning is passed back in later turns (interleaved thinking). Text only, with a 256K-token context and tool calling.
MiniMax: MiniMax M2.7
A MiniMax text model for practical agent work: planning and carrying out multi-step tasks, often with several cooperating agents. Typical uses are live debugging, root-cause investigation, financial models and producing Word, Excel and PowerPoint documents. Context is 200K tokens and tool calling is supported.
Mistral AI: Mistral Small 4
A compact Mistral AI model (snapshot 2603) that merges reasoning, image understanding and agentic coding into a single system, so one model can cover analysis, software work and visual input. It has a 256K-token context and supports tool calling.
NVIDIA: Nemotron 3 Super
An open NVIDIA model with 120B total and 12B active parameters, whose mixture-of-experts design mixes Mamba and Transformer layers and predicts several tokens at once for faster generation. It targets multi-agent systems, cross-document reasoning and multi-step planning; weights, data and training recipes are published under the NVIDIA Open License. Text only; here it is served with a 256K-token context and tool calling.
Google: Gemini 3.1 Pro Low
Google Gemini 3.1 Pro Low text model. Available over the Chat Completions protocol.
OpenAI: GPT-5.3 Codex Spark
OpenAI GPT-5.3 Codex Spark text model. Available over the Chat Completions protocol.
Qwen: Qwen3 Coder Next
An open-weight Qwen coding model for agents and local developer tooling. The sparse mixture-of-experts design has 80B total parameters but only 3B active per token, which keeps always-on agents cheap. It answers without a separate thinking block, handles long coding sessions and recovery after failed commands, and offers a 256K-token context with tool calling.
Mistral AI: Devstral 2
Mistral AI's open coding model for agents: a dense 123B-parameter transformer with a 256K-token context, released in December 2025 (snapshot 2512). It is built to navigate a codebase, make coordinated edits across many files, notice failing runs and retry with fixes. Supports tool calling; the weights ship under a modified MIT license.
Amazon: Nova 2 Lite
An inexpensive reasoning model from Amazon's second Nova generation for routine workloads. It processes text, images, video and documents, and is used for extracting data from files and video, writing code, answering with grounded sources and automating multi-step agent tasks. Context is 1M tokens; tool calling is supported.
Meta: Llama 4 Maverick
Meta's open multimodal Llama 4 model, released in April 2025: a mixture-of-experts with 128 experts, 400B total and 17B active parameters. It is instruction-tuned for assistant use, image reasoning and general multimodal chat in 12 languages, with a knowledge cutoff of August 2024. Accepts text and images, with a 1M-token context and tool calling.
Rates come from the current published price list. Token rates are shown per million tokens; image rates are shown per image. Context, maximum output and release date appear only when published in the model catalogue. A dash means no value is published. Performance in each model card comes from completed requests over the last 24 hours, not from estimated or reference values