Frontier AI Labs

Makes ChatGPT and the GPT model family — the company most people mean when they say “AI” for the first time.

• GPT-6 Sol— Rolling out in ChatGPT from October 7 to Plus, Pro, Business and Enterprise, with responses that can include charts, forms and interactive tools.Paid only
• GPT-6 Luna— The everyday GPT-6 model rolling out to ChatGPT Free and Go from October 8; free-plan tool and usage limits still apply.Free with limits
• GPT-6 Astra— The new flagship (launched 3 September 2026): deep reasoning, agentic computer-use automation, top-tier coding and research work.Paid only
• GPT-5.6 Sol— Previous-generation model for professional and coding work; the October ChatGPT rollout introduces GPT-6 Sol.Paid only
• GPT-5.6 Terra— Previous-generation balanced model for everyday reasoning and coding on paid plans; not the current free-plan model.Paid only
• GPT-5.6 Luna— Previous-generation everyday chat model; GPT-6 Luna is rolling out to Free and Go from October 8.Free with limits
• gpt-oss-120b / gpt-oss-20b— OpenAI’s own open-weight models — download and run them on your own hardware.Free, open-weight

• Note: Specialized and restricted models (cybersecurity research, life-sciences research, image and voice generation) exist but need vetted access or a paid plan — they are not general-use free tools.
• Note: Sora (AI video) is discontinued: the app shut down 26 April 2026 and the API follows on 24 September 2026. OpenAI currently has no video-generation product.

Makes Claude and Claude Code — known for reasoning quality and for being the most safety-research-forward of the major labs.

• Claude Haiku 5.5— Released October 7 for high-volume, cost-sensitive work with adjustable effort; paid API access starts at $0.10 input / $0.50 output per million tokens for prompts up to 100K.Paid only
• Claude Fable 5.1— The newest public flagship (launched 1 September 2026), Anthropic’s top pick for demanding reasoning and long agentic tasks.Paid only
• Claude Opus 5— The recommended model for most complex agentic coding and enterprise work.Paid only
• Claude Sonnet 5— The default workhorse — best balance of speed and intelligence; included on Claude.ai’s free plan on a rolling usage window.Free with limits
• Claude Haiku 4.5— The fastest, cheapest tier, still near-frontier for everyday tasks; also included on the free plan alongside Sonnet 5.Free with limits
• Claude Mythos 5.1— Anthropic’s highest-capability model, restricted to vetted cybersecurity and bio-research organizations — not public.Paid only
• Claude Code— The agentic coding tool built on the models above; no free tier of its own, bundled from the Pro plan up.Paid only

• Note: October 26, 2026: Anthropic announced a Form S-1 filing for a proposed IPO of Class A common stock, with an intended Nasdaq Global Select Market listing under ANT. The share count and price range are not set, and the registration statement is not yet effective (source: anthropic.com/news/s1-sec).

Makes Gemini, and has the most free-tier-friendly lineup of the major labs.

• Gemini 4 Carbon— Google's newest frontier model, announced October 28, 2026, with a focus on cybersecurity and long-horizon agentic work. It is rolling out first to a limited group of cyber defenders through Google's Fairwind Program, with wider developer, business and consumer access promised later — not yet generally available.Paid only
• Gemini 4 Argon— Google's frontier model for reasoning and coding; available in the Gemini app for AI Pro and Ultra subscribers and via the API.Paid only
• Gemini 3.1 Pro— Flagship for complex reasoning, coding and creative work — free with limits in the Gemini app, paid-only via the API.Free with limits
• Gemini 3.8 Flash (and the 3.5 / 3.6 Flash-Lite tiers below it)— Fast, capable everyday and agentic-workflow models, free through both the Gemini app and Google AI Studio’s API, with rate limits.Free
• Gemini 3.1 Deep Think— An extended-reasoning mode for hard research and engineering problems, on the Google AI Ultra plan.Paid only
• Nano Banana / Nano Banana Pro— Image generation and editing — free with limits in the Gemini app and Search; full API access is mostly paid.Free with limits
• Gemma 4— Google’s fully open-weight family, sized from phone-friendly to workstation-class — download from Hugging Face, Kaggle or Ollama.Free, open-weight

• Note: Veo (video) and Lyria (music) generation exist but are paid-plan features, not free.
• Note: October 28, 2026: Google announced Gemini 4 Carbon as its strongest model, reporting state-of-the-art cybersecurity scores (SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0% — Google's own figures) and first place on Artificial Analysis's Intelligence Index v5 at 59. Introductory API pricing matches Argon at $2 in / $10 out per million tokens, rising to $4 / $20 after the introductory period (source: deepmind.google, Gemini 4 Carbon: our next frontier model).
• Note: October 27, 2026: Google announced research on Genie 4, a world model for real-time interactive simulation with native audio. It is rolling out to a limited group of trusted testers first, then to Google AI Ultra subscribers in Project Genie and early access partners in the Gemini API and AI Studio; it is not free (source: blog.google, Introducing Genie 4).

Meta now runs two separate model lines worth knowing apart.

• Muse Spark 1.3— Meta’s new proprietary flagship (launched 2 September 2026), powering the Meta AI app, Facebook, Instagram, WhatsApp and Ray-Ban/Oakley Meta glasses; free to chat with inside Meta’s own apps, API paid-only.Free with limits
• Llama 4 Scout / Maverick / 4.5— Meta’s open-weight line — no longer the frontier priority but still actively maintained; download and self-host.Free, open-weight

• Note: Meta’s own hosted API has no free tier for Llama either. If you want Llama without your own hardware, a third-party host with a free tier (OpenRouter, Groq) is the practical free option, not Meta directly.

Formerly xAI — merged into SpaceX in February 2026 and rebranded as SpaceXAI in July 2026. The Grok product name didn’t change; the company behind it did.

• Grok 4.6— Current flagship: agentic tool use and configurable reasoning depth. The free web tier does not include it — only an older, lesser model with unspecified limits.Paid only
• Grok Build 0.1— A dedicated coding-agent model, bundled into all plans including free as of August 2026.Free with limits

• Note: Grok Imagine (image and video generation) and the full Voice API are paid-only; basic voice chat is included free in the app.
• Note: Of the five major labs this is the least free-tier-friendly: there is no free API tier at all, and you must load paid credits before the API answers any request.

A French lab known for open-weight releases alongside its paid flagship line.

• Mistral Large 4— 1-trillion-parameter natively multimodal MoE (52B active), in public preview on Mistral Studio with weights scheduled for end of October.Paid only
• Mistral Large 3— Flagship for complex reasoning and agentic tasks — paid via API, usable free with limits in Le Chat, which also includes $10 a month of free API credit.Free with limits
• Mistral Small 4, Devstral 2 (coding agent) and Pixtral (vision)— Downloadable and self-hostable under permissive licenses.Free, open-weight

A Chinese lab that moved past its well-known V3/R1 naming to a new V4 generation in 2026, generally released with open weights.

• DeepSeek-V4-Pro— Current flagship for general and agentic tasks with adjustable reasoning effort — open weights on Hugging Face, also available paid via API.Free, open-weight
• DeepSeek-V4.1-Flash— Smaller, faster sibling with native image understanding.Free, open-weight

One of the most prolific open-weight labs — dozens of model sizes, most free to self-host.

• Qwen3.8-Max— Flagship, best for autonomous coding on complex projects; API-only.Paid only
• Qwen3-Coder-480B and Qwen3-235B-A22B— Top open-weight coding and general models, Apache 2.0 licensed.Free, open-weight
• Qwen3 4B / 1.7B / 0.6B— Smaller models that run on a laptop.Free, open-weight

• Note: There is also a free consumer chat app at chat.qwen.ai.

A Chinese lab whose Kimi K-series is frequently cited alongside DeepSeek and Qwen as top-tier open-weight competition.

• Kimi K3— Current flagship (July 2026) with strong coding and agentic performance; public weights under Moonshot’s own license. The free tier of Moonshot’s own Kimi app does not include K3 — only the older K2.6.Free, open-weight
• Kimi K2.6— Prior flagship — free with limits in Moonshot’s free “Adagio” app tier, and also free to download and self-host.Free with limits

Known for competitive benchmark performance and a free small-model tier.

• GLM-5.3— Current flagship with strong general and coding performance; API-only.Paid only
• GLM-4.5-Flash (and related Flash-tier models)— Zhipu’s own pricing lists these at $0 per token.Free

• Note: Older GLM-4.x generations are open-weight; GLM-5.3’s open-weight status is unconfirmed as of this writing.

Amazon’s own model family, accessed through AWS Bedrock rather than a consumer app.

• Nova 2 Lite and Nova 2 Sonic— Amazon’s current fast multimodal and speech models, paid via Bedrock; new AWS customers get a one-time $200 credit but there is no ongoing free tier.Trial credits only

• Note: Nova Premier (the previous flagship) and Nova Canvas/Reel (image and video) were both discontinued in September 2026.
• Note: Of all twelve labs in this directory, Nova has the weakest free-tier story — no standing free access at all.

Ships the Copilot assistant across Windows, Edge and Microsoft 365 — built on a mix of OpenAI models and Microsoft’s own in-house “MAI” model family.

• MAI-Transcribe-2-Streaming— In-house streaming speech-to-text model with 2.5% WER on Artificial Analysis; available in public preview on Azure.Paid only
• MAI-Voice-2.1— In-house text-to-speech model for natural, expressive generation and voice matching; available in public preview on Azure.Paid only
• Copilot (consumer)— Microsoft’s assistant app and web chat. The September 2026 relaunch added a Home hub, a Code mode for programming tasks, and an Autopilot mode that carries out multi-step tasks; the base tier is free, with paid Pro and Microsoft 365 bundles on top.Free with limits
• MAI-1-preview— Microsoft’s first fully in-house general-purpose model — testable free on LMArena, with limited free use inside some Copilot features.Free with limits
• MAI-Code-1-Flash— A lightweight coding model, now the default model in VS Code and bundled into GitHub Copilot.Free

• Note: Copilot still routes many requests to OpenAI models; Microsoft does not label which model answers a given prompt.
• Note: MAI-Transcribe-2-Streaming and MAI-Voice-2.1 (October 2026) provide low-latency transcription and speech synthesis via Azure; other MAI models remain bundled inside Microsoft 365 apps.

Founded in 2024 by Ilya Sutskever (OpenAI's former chief scientist), Daniel Gross and Daniel Levy with a single stated goal: building safe superintelligence, with no near-term product line.

• No public models— SSI ships no consumer or API products — it is a pure research lab, so there is nothing to sign up for or try.Paid only

• Note: There is no free tier, no API and no waitlist product — the lab's site is a mission statement, not a storefront.

Founded in 2025 by Mira Murati (OpenAI's former CTO) with a team of well-known researchers, focused on making AI systems more understandable, customizable and generally capable.

• Tinker— A fine-tuning API for open-weight models — researchers and developers bring their own data and training loops; access has been via a waitlist.Free with limits

• Note: The lab also funds external work: its Safety Research Grants program supports independent research on AI safety and evaluation.

Founded in 2024 by Fei-Fei Li (creator of ImageNet) to build “world models” — AI that understands and generates 3D environments rather than just text.

• Marble— A generative world model that creates persistent, navigable 3D environments from text or images — free tier available with limits, paid plans for heavier use.Free with limits
• Atlas (research preview)— A larger world model announced on the lab's blog, aimed at more coherent, physically consistent generated worlds.Free with limits

A US lab founded in 2024 by former DeepMind researchers, working on open-weight frontier models and autonomous coding systems.

• Asimov (research preview)— A code-understanding agent designed to read and reason over entire codebases; access is via a waitlist rather than open sign-up.Free with limits

• Note: The lab has said it plans to release open-weight frontier models; as of this writing no weights are publicly downloadable.

A Canadian lab focused on enterprise AI — its models are built for business deployment, retrieval and private-cloud setups rather than a consumer chat app.

• Command A— Cohere's flagship enterprise model for agents, retrieval-augmented generation and multilingual business use; open weights for research, API is paid with a free trial tier.Free with limits
• Embed and Rerank— Search and retrieval models that power document search inside other products; free trial-rate API keys available.Free with limits

• Note: Cohere's free API tier is rate-limited and intended for evaluation and prototyping, not production traffic.

A non-profit research institute in Edmonton, founded by reinforcement-learning pioneer Richard Sutton and colleagues, doing basic research on how minds can form from computation.

• Note: Openmind ships no products or models — it is a basic-research institute. Everything it produces is papers and open-source code, shared publicly with no intellectual-property restrictions.
• Note: Its research follows The Alberta Plan for AI Research, a 12-step roadmap co-authored by Sutton toward continually learning agents.