AI models people build agents on, from frontier APIs to open-weight models you can run yourself, with context, prices and licences from each provider's own pages.
14 entries. Latest source check .
What a source check covers
Documentation checked means the source was read. It does not mean the software was installed, tested or security audited.
Reading an entry does not authorise installation, credential access or spending. Tech Horizon Labs projects are labelled, and there are no affiliate links or paid placement.
Subscription, API key or your own model?
A subscription: one monthly price for a coding agent
Claude Pro costs US$20 a month and includes Claude Code; Max is US$100 and Ultra US$200. ChatGPT Plus is US$20 a month, and every ChatGPT plan includes Codex. xAI's X Subscription runs from US$8 to US$200 a month, and even its US$8 Plus plan includes Grok Build, xAI's terminal coding agent. Google AI Pro is US$19.99 a month and allows 1,500 Gemini CLI requests a day. These are US prices, read on 7 November 2026.
Check where a plan may be used before you build on it. Anthropic says its plans are for Claude Code and its own apps, and that products or services built on the Agent SDK should use an API key. Google says signing in to Gemini CLI from a third-party tool can get the account suspended. OpenAI's Sign in with ChatGPT lets open-source and locally hosted apps use your plan, and xAI supports SuperGrok and X Premium plans inside OpenCode, including on a remote server. Anthropic, OpenAI and xAI all forbid sharing an account.
Agents that run on their own, on a server or inside a workflow, are what API keys are for, and the providers say so for their own tools: Anthropic for products and services built on the Agent SDK, OpenAI for Codex in CI/CD jobs, and xAI for Grok Build where there is no browser. The table below gives each model's price per million tokens.
Anthropic, OpenAI and xAI say they do not train on API data without your permission. Google's paid tier does not either, but its free tier and AI Studio use what you send to improve Google products, and people may read it. Claude Max and Ultra include US$100 and US$200 of API credit a month, and ChatGPT Pro tiers include monthly API credit matching their price.
Your own model: no token bill, but it must fit in memory
Open-weight models such as Gemma 4, gpt-oss-120b and Qwen run on your own machine through Ollama, which is MIT licensed and runs on macOS, Windows and Linux, or LM Studio, which is free for running models but is not open source. Prompts stay on your machine when the model runs locally, and there is no per-token bill.
Memory decides what fits. Google puts Gemma 4 31B at about 17.5 GB in 4-bit for the weights alone, with more needed for long conversations. OpenAI says gpt-oss-20b runs within 16 GB and gpt-oss-120b fits a single 80 GB GPU. Qwen's own model cards give no memory figure; Ollama's download of Qwen3.8 27B is 18 GB. Qwen's largest open model, Qwen3.8-2.4T-A95B, has 2.4 trillion parameters, and Qwen also offers it hosted as Qwen3.8-Max.
Figures come from each maker's or provider's own pages on the date the entry was checked. Prices and specifications change, so open the entry and its official source before you buy or build.
A Claude model from Anthropic built for long-running agentic coding and knowledge work, such as multihour autonomous coding agents, large-scale refactoring, vision-heavy workflows and computer use.
Developers building long-running coding or knowledge-work agents who want the model Anthropic suggests starting with for most workloads.
Access and use
API model id: claude-opus-5-5 (anthropic.claude-opus-5-5 on Amazon Bedrock). It is available on the Claude API, Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry and Claude Platform on AWS. The context window is 1M tokens and max output is 128K tokens, or up to 300K on the Message Batches API with a beta header. It costs US$4 per million input tokens and US$20 per million output tokens (pricing page read 7 November 2026), with the full 1M context at standard pricing. Cache reads cost US$0.20 per million tokens, the Batch API is 50% off, fast mode costs US$8 and US$40 per million tokens, and US-only inference adds a 1.1x multiplier. It takes text and images and returns text. It supports tool use, but forced tool use (tool_choice any or tool) returns an error, and adaptive thinking is always on. Anthropic's commercial retention policy says API inputs and outputs are deleted from its backend within 30 days, except for features with longer retention under your control, zero data retention agreements, Usage Policy enforcement or legal requirements, and flagged content can be kept for up to 2 years. Anthropic's docs say retained data is never used for model training without your express permission. Proprietary; API terms apply.
A Claude model from Anthropic that Anthropic calls its best combination of speed and intelligence, built for everyday coding, agent and enterprise work such as code generation, data analysis and agentic tool use.
Teams running everyday coding and agent workloads who want lower cost and latency than Claude Opus 5.5.
Access and use
API model id: claude-sonnet-5-5 (anthropic.claude-sonnet-5-5 on Amazon Bedrock). It is available on the Claude API, Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry and Claude Platform on AWS. The context window is 1M tokens and max output is 128K tokens, or up to 300K on the Message Batches API with a beta header. It costs US$2 per million input tokens and US$10 per million output tokens (pricing page read 7 November 2026), with the full 1M context at standard pricing. Cache reads cost US$0.10 per million tokens, the Batch API is 50% off, and US-only inference adds a 1.1x multiplier. It takes text and images and returns text. It supports tool use, but forced tool use (tool_choice any or tool) returns an error. Adaptive thinking is on by default; setting temperature, top_p or top_k returns an error. Anthropic's commercial retention policy says API inputs and outputs are deleted from its backend within 30 days, except for features with longer retention under your control, zero data retention agreements, Usage Policy enforcement or legal requirements, and flagged content can be kept for up to 2 years. Anthropic's docs say retained data is never used for model training without your express permission. Proprietary; API terms apply.
Teams running high-volume classification, routing, extraction or subagent steps where cost and latency matter most.
Access and use
API model id: claude-haiku-5-5 (anthropic.claude-haiku-5-5 on Amazon Bedrock). It is available on the Claude API, Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry and Claude Platform on AWS. The context window is 1M tokens and max output is 128K tokens, or up to 300K on the Message Batches API with a beta header. It is priced by prompt length: US$0.10 per million input tokens and US$0.50 per million output tokens for prompts up to 100,000 tokens, and US$0.50 and US$2.50 for prompts over 100,000 tokens (pricing page read 7 November 2026). Cache reads cost US$0.01 per million tokens (US$0.05 over 100,000 tokens), the Batch API is 50% off, and US-only inference adds a 1.1x multiplier. Its tokenizer counts the same text as about 30 percent more tokens than Claude Haiku 4.5. It takes text and images, returns text, and supports tool use. Anthropic's commercial retention policy says API inputs and outputs are deleted from its backend within 30 days, except for features with longer retention under your control, zero data retention agreements, Usage Policy enforcement or legal requirements, and flagged content can be kept for up to 2 years. Anthropic's docs say retained data is never used for model training without your express permission. Proprietary; API terms apply.
A Claude model from Anthropic for demanding reasoning and long-horizon agentic work, which Anthropic suggests when Claude Opus 5.5 at higher effort still does not meet a team's requirements.
Teams whose hardest reasoning, research or long-running agent tasks still fall short on Claude Opus 5.5 and who can accept 30-day data retention.
Access and use
API model id: claude-fable-5-1 (anthropic.claude-fable-5-1 on Amazon Bedrock). It is available on the Claude API, Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry and Claude Platform on AWS. The context window is 1M tokens and max output is 128K tokens. It costs US$10 per million input tokens and US$50 per million output tokens (pricing page read 7 November 2026), with the full 1M context at standard pricing. Cache reads cost US$0.25 per million tokens, the Batch API halves prices to US$5 and US$25, and US-only inference adds a 1.1x multiplier. It takes text and images and returns text. It supports tool use, but forced tool use (tool_choice any or tool) returns an error. Anthropic treats it as a Covered Model: it requires 30-day data retention, so zero data retention is not available unless Anthropic expressly authorises it. Anthropic's commercial retention policy says API inputs and outputs are deleted from its backend within 30 days, except for features with longer retention under your control, zero data retention agreements, Usage Policy enforcement or legal requirements, and flagged content can be kept for up to 2 years. Anthropic's docs say retained data is never used for model training without your express permission. Proprietary; API terms apply.
A proprietary OpenAI model that OpenAI calls its most capable for the most demanding work, for complex reasoning, coding, computer use, research and document creation.
Developers building coding, computer use or research agents on the OpenAI API who want OpenAI's flagship model rather than the cheaper GPT-6.1 Sol or GPT-6 Luna.
Access and use
Call it as gpt-6-astra through the OpenAI Responses API or Chat Completions, and through the Batch API. OpenAI says it is also available to ChatGPT Plus, Pro, Business and Enterprise users and through Microsoft Azure and Amazon Bedrock (model id openai.gpt-6-astra), with Google Cloud in private preview. The context window is 1,050,000 tokens, with up to 922,000 input tokens and 128,000 output tokens. Standard pricing is US$10 per million input tokens and US$50 per million output tokens (pricing page read 7 November 2026); prompts over 272K input tokens are charged at twice the input and cache rates and 1.5 times the output rate for the whole request. API credits can only be bought in US dollars. It takes text and image input and returns text. OpenAI says data sent to the API is not used to train its models unless you opt in. Abuse monitoring logs, which can include prompts and responses, are kept for up to 30 days by default, and the Responses API keeps response data for at least 30 days by default or when store is true. Zero Data Retention and Modified Abuse Monitoring need OpenAI's prior approval. OpenAI monitors Astra for misalignment, and its extra safety checks can stop legitimate work: in the API, a flagged task stops. Proprietary; API terms apply.
An open-weight OpenAI reasoning model for production, general purpose and high reasoning use, with native function calling, web browsing, Python code execution and structured outputs for agentic tasks.
Teams that want an OpenAI reasoning model they can download, fine-tune and run on their own GPU or cloud instead of calling a hosted API.
Access and use
Download the weights from Hugging Face as openai/gpt-oss-120b, for example with huggingface-cli download openai/gpt-oss-120b --include "original/*" --local-dir gpt-oss-120b/, or pull it with Ollama (ollama pull gpt-oss:120b) or LM Studio (lms get openai/gpt-oss-120b). OpenAI's docs also list it on Amazon Bedrock as openai.gpt-oss-120b. The context window and maximum output are both 131,072 tokens. It takes text in and returns text, with low, medium or high reasoning effort set in the system prompt. It supports function calling with defined schemas, web browsing and Python code execution through built-in tools, and structured outputs. It must be used with OpenAI's harmony response format or it will not work correctly; the Transformers chat template applies it automatically. Its full chain of thought is visible to developers but is not intended to be shown to end users. It has 117B parameters with 5.1B active. With MXFP4 quantisation of its MoE weights it runs on a single 80GB GPU such as an NVIDIA H100 or AMD MI300X, and it can be fine-tuned on a single H100 node; the smaller gpt-oss-20b runs within 16GB of memory. The usage policy asks you to comply with all applicable law and adds no other use restrictions. Apache-2.0.
A Google reasoning model built for complex software engineering, deep mathematical and scientific problem solving, autonomous multi-step agents and very long outputs of up to 1 million tokens.
Developers building coding, research or multi-step agents on the Gemini API who need a 1 million token context and long outputs.
Access and use
API model id: gemini-4-argon. Google offers it through the Gemini API and Google AI Studio; it is not on the free tier, so API use needs a Cloud project with an active billing account. The context window is 1,048,576 tokens shared between input and output, with a default maximum output of 65,536 tokens, or up to 1,048,576 with long decode. Paid tier pricing is an introductory US$2.00 per million input tokens and US$10.00 per million output tokens through 30 November 2026, rising to US$4.00 and US$20.00 from 1 December 2026 (pricing page read 7 November 2026); output includes thinking tokens. Inputs are text, image, video, audio and PDF; output is text. Google says paid-tier prompts and responses are not used to improve its products, but it logs them for a limited period to detect Prohibited Use Policy violations and for legal disclosures, and that data may be stored or cached in any country where Google has facilities. Separately, your project's own Logs and Datasets logging is on by default for Interactions API calls and for Generate Content calls in projects created on or after 26 October 2026, with logs kept for up to 55 days by default; a dataset you choose to share with Google may be used to train its models. Proprietary; API terms apply.
A Google model, described as its most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents and complex enterprise workflows with the speed and cost efficiency of the Flash line.
Developers running high-volume or cost-sensitive agents on the Gemini API who still need a 1 million token context and tool use.
Access and use
API model id: gemini-3.8-flash. Google offers it through the Gemini API, Google AI Studio and Gemini Enterprise Agent Platform. The input limit is 1,048,576 tokens and the output limit is 65,536 tokens. Paid tier pricing is an introductory US$0.75 per million input tokens and US$3.75 per million output tokens through 31 December 2026, then US$1.50 and US$7.50 from 1 January 2027 (pricing page read 7 November 2026); output includes thinking tokens. On the free tier, Google uses prompts and responses to improve its products and machine learning technologies, human reviewers may read them, and Google asks users not to submit sensitive, confidential or personal information. Google says paid-tier prompts and responses are not used to improve its products, but it logs them for a limited period to detect Prohibited Use Policy violations and for legal disclosures, and that data may be stored or cached in any country where Google has facilities. Separately, your project's own Logs and Datasets logging is on by default for Interactions API calls and for Generate Content calls in projects created on or after 26 October 2026, with logs kept for up to 55 days by default; a dataset you choose to share with Google may be used to train its models. Proprietary; API terms apply.
A family of Google DeepMind open-weight models in five sizes that take text and image input and suit reasoning, agentic workflows, coding and multimodal understanding on devices from phones to workstations.
Developers who want an Apache-2.0 model with function calling that they can run on their own phone, laptop, GPU or server.
Access and use
Download the weights from Hugging Face (for example google/gemma-4-31B-it, not gated) or Kaggle; Google publishes official 4-bit GGUF builds for local use on CPU, Apple Silicon or consumer GPUs. The Gemini API also hosts the two largest sizes as gemma-4-31b-it and gemma-4-26b-a4b-it, free of charge on the free tier only, where Google uses prompts and responses to improve its products. The context window is 128K tokens on E2B and E4B and 256K tokens on 12B, 26B A4B and 31B. All sizes take text and image input and produce text. Google lists native function calling, a system role and configurable thinking. Parameter counts are 2.3B effective (5.1B with embeddings) for E2B, 4.5B effective (8B) for E4B, 11.95B for 12B, 30.7B for 31B, and 25.2B total with 3.8B active for the 26B A4B mixture-of-experts model. Google's own table puts the GPU or TPU memory to load the weights at 69.9 GB in 16-bit, 34.9 GB in 8-bit and 17.5 GB in 4-bit for 31B; 57.7, 28.8 and 14.4 GB for 26B A4B (all of its parameters must be loaded); 26.7, 13.4 and 6.7 GB for 12B; and 2.9 GB in 4-bit for E2B; the context window and supporting software need memory on top. Google says the models permit responsible commercial use, and its Gemma 4 licence page is the Apache License 2.0. Apache-2.0.
A Grok model from xAI (published under the SpaceXSI name in its current docs) that the company calls its flagship for code, agentic tool calling and knowledge work, with configurable reasoning.
Developers building coding and tool-calling agents who want a hosted model with a 500,000-token context and xAI's server-side search and code execution tools.
Access and use
Call it as grok-4.7 through the Responses API at https://api.x.ai/v1 with an API key from the xAI Console; the OpenAI SDK works with that base URL. The context window is 500,000 tokens, xAI lists no text output limit, and the knowledge cutoff is May 2026. Pricing is US$2.00 per million input tokens and US$6.00 per million output tokens for prompts under 200,000 tokens, rising to US$4.00 and US$12.00 once a prompt reaches 200,000 tokens, and the higher rate then applies to every token in that request (pricing page read 7 November 2026). It takes text and image input and returns text, with reasoning effort set to low, medium, high (the default) or xhigh. It supports function calling and structured outputs, plus xAI's server-side web search, X search and code execution tools. xAI says it never trains on API inputs or outputs without explicit permission; by default it stores requests and responses, encrypted, for 30 days for abuse auditing and then deletes them. Team admins can turn on Zero Data Retention, which also switches off the stateful Responses API, Files, Collections and the Batch API. The default endpoint does not guarantee a processing region; https://us.api.x.ai/v1 keeps inference and retained request data in the United States. Proprietary; API terms apply.
An open-weight DeepSeek mixture-of-experts model, the smallest in its V4.1 family, that takes text and images and that DeepSeek says improves on reasoning, coding and agentic tasks.
Developers who want a low-cost hosted model for coding and agent harnesses, and teams with multi-GPU servers who want MIT-licensed weights to run themselves.
Access and use
Call it as deepseek-flash on the DeepSeek API at https://api.deepseek.com (OpenAI format) or https://api.deepseek.com/anthropic (Anthropic format). The weights are published on Hugging Face and ModelScope. Context length is 1M tokens with a maximum output of 384K tokens. Peak-hour pricing is US$0.30 per million input tokens on a cache miss and US$1.20 per million output tokens, and off-peak rates are half that (pricing page read 7 November 2026). DeepSeek's Open Platform Terms handle a developer's own personal information under the DeepSeek Privacy Policy, which says DeepSeek collects, processes and stores personal data in the People's Republic of China, uses personal data to train and improve its models, gives a right to opt out of that training, and keeps personal data for as long as necessary to provide its services. DeepSeek's training disclosure says a small portion of its fine-tuning data may be based on user input, de-identified first. The open weights (deepseek-ai/DeepSeek-V4.1-Flash-1017) are a 552B-parameter mixture-of-experts model with 8B active parameters for input and 16B for output. The repository and the model weights are under the MIT License, which grants use without restriction provided the copyright and permission notice is kept. MIT.
A 2.4 trillion parameter open-weight mixture-of-experts model from Alibaba's Qwen team, with 95B parameters active per token, built for coding, professional work, research and long-horizon agentic tasks.
Teams who want to self-host Qwen's largest open-weight model for coding and agent work, or call its hosted version, Qwen3.8-Max, through Qwen Cloud.
Access and use
The open weights are Qwen/Qwen3.8-2.4T-A95B on Hugging Face; the hosted version on Qwen Cloud takes the API model id qwen3.8-max. Self-host with SGLang, vLLM or TokenSpeed, or reach Qwen3.8-Max through the Qwen Cloud API (OpenAI-compatible). The open weights have a native context of 262,144 tokens, extensible to 1,010,000 tokens; Qwen Cloud lists a 1M context, 991K maximum input and 131K maximum output for Qwen3.8-Max. Qwen3.8-Max costs US$2 per million input tokens and US$6 per million output tokens on Qwen Cloud (Singapore region; pricing page read 7 November 2026). Qwen Cloud's data security page says request data is stored in the region whose endpoint you call; that page does not address training on API data. The model has 2.4T total parameters with 95B activated, and the card gives no hardware requirement. The Qwen3.8-Max License permits use, modification, hosting and resale, but products with more than 100 million monthly active users or US$20 million monthly revenue must display the model name in their interface, and a Model as a Service or AI Work Assistant business with more than US$50 million revenue in 12 months needs a separate licence from Qwen for commercial use; internal use that exposes nothing to third parties is exempt. Qwen3.8-Max License (custom, not OSI).
An open-weight, general-purpose multimodal mixture-of-experts model from Mistral that combines instruction following, reasoning, coding and agentic workflows in one set of weights, with a 1M token context window.
Developers building agents who want one open-weight model for chat, reasoning, coding and tool use, called through the Mistral API or hosted on their own GPUs.
Access and use
The API model id is mistral-large-4 (also mistral-large-4-0), and the open weights are mistralai/Mistral-Large-4-1T-A52B on Hugging Face, with a quantised NVFP4 checkpoint that targets Blackwell GPUs. The context window is 1M tokens. It costs US$0.70 per million input tokens and US$2.10 per million output tokens on the Mistral API (pricing page read 7 November 2026). It takes text and image input and returns text. On Mistral Studio's free mode, Mistral may use API inputs and outputs to train its models unless you opt out; pay-as-you-go customers can opt out at any time; thumbs up or down feedback on any plan authorises Mistral to use that input and output. Zero data retention is available on paid plans for stateless endpoints such as /v1/chat/completions, but not for agents, conversations or batch processing files. The model has about 1 trillion total parameters with 49B active per token. The card's serving examples assume a node of 8 GPUs. The Mistral Ecosystem Open-Weight License is a modified MIT licence: a business that operates a Model as a Service must sign a separate commercial licence with Mistral before any commercial use, while internal use and use through Mistral's official products are exempt. Mistral Ecosystem Open-Weight License (custom, not OSI).
An open-weight mixture-of-experts model from Meta with 816B total and 42B active parameters, built for coding and agentic work, natively multimodal and with a 1M token context window.
Organisations with very large GPU memory, or two Mac Studios with M5 Ultra, that want to self-host a multimodal open-weight model for coding agents and long tool-using workflows.
Access and use
The model card says the model is available by API through third party inference providers; the open weights are meta-models/Muse-Spark-Open-816B-A42B on Hugging Face, in BF16 and NVFP4 versions. Download is gated: you log in and accept the licence conditions, and the information you provide is handled under the Meta Privacy Policy. The context length is 1M tokens at inference. It takes text, image, video and audio input and returns text. The model has 816B total and 42B active parameters. Meta says it needs a GPU with at least 700GB of VRAM for NVFP4 and 2TB for BF16, and that it is optimised to run on two Mac Studios with M5 Ultra using MLX. Meta recommends deploying it inside a broader AI system with guardrails suited to the use case, and says it is not intended for individuals under 18. The Muse Model License allows use, modification, distillation and resale, but you must keep the notice and the Usage Policy with copies, follow the Usage Policy's list of prohibited uses, display "Built with Muse Spark Open" in each product's interface or its documentation, website or settings, and notify Meta and sign a separate agreement before offering it as a Model as a Service once revenue exceeds US$20 million over 12 months. Muse Model License (custom, not OSI).