The shipping AI model tracker

How to read this tracker

A few conventions that apply to every table below. Skim these once, then the rows make sense at a glance.

• Params — Listed only where the lab has officially disclosed the figure. For Anthropic, OpenAI, and frontier Gemini, this column reads 'not disclosed' on purpose. We do not repeat third-party guesses.
• Context window — Maximum total tokens (input plus output) the API will accept for the model's standard tier. Some providers sell longer-context variants at higher prices; we note those inline.
• Modalities — Text in / text out is the baseline. We mark vision, audio, image generation, and tool use where the model supports them natively (not via a wrapper).
• Price per 1M tokens — Standard public API pricing in USD as of the page's June 2026 refresh. Input and output are listed separately because the gap matters for any agentic workload. Batch-API and cached-input discounts are footnoted, not in the main number.
• Release date — First public general-availability date from the provider's own announcement. Preview/beta dates are noted as such.
• Open weights — Yes means downloadable from Hugging Face (or the lab's own host) under a license that permits local inference. 'Yes, gated' means the weights ship but the license restricts commercial use or requires acceptance.
• License — The exact license name. The fine print on Llama, Gemma, and DeepSeek is non-trivial; we link to the actual license file, not a summary.

Frontier closed-weight models

These are the labs that do not ship weights. Pricing is API-only. Parameter counts are not disclosed by any of these providers and we do not estimate them. All dates and prices were verified against the provider's own documentation at the page's June 2026 refresh — re-check before quoting in a contract.

On 'GPT-5', 'Claude 4.5', 'Gemini 3', and other models that may or may not exist by the time you read this

Frontier open-weight models

These are the models you can actually download and run on your own hardware (or rent from a third-party host like Together, Fireworks, Groq, or Cerebras). Pricing is hosted-API pricing from the model's primary commercial host, where available — but the point of this tier is that you do not have to use any host at all. All have permissive-enough licenses for most builder use cases; read the fine print on the Llama and DeepSeek licenses before shipping in production.

Specialized and embedded models

These are models that are not trying to be a generalist chat assistant. They are tuned for one job — code completion, embeddings, image generation, speech, or on-device inference — and the right shopping question is 'best in class for this specific narrow task,' not 'best overall.'

Open-source heroes — the labs that ship weights

Release timeline — the last 24 months

A compressed chronology of the releases that moved the market. Dates are first-public-availability, not paper-publication.

1. 2024-04

   Llama 3 8B / 70B; Mixtral 8x22B; Command R+

   Meta ships the first Llama 3 generation. Mistral releases Mixtral 8x22B under Apache 2.0. Cohere ships Command R+ as the first open-weight RAG-tuned 100B-class model.

2. 2024-05

   GPT-4o

   OpenAI ships its first natively multimodal frontier model — text, vision, and audio in one weight set. Cuts price-per-token by roughly half vs. GPT-4 Turbo.

3. 2024-06

   Claude 3.5 Sonnet (first cut); Gemma 2; Nemotron 340B

   Anthropic ships 3.5 Sonnet, which would dominate enterprise coding benchmarks for the next nine months. Google releases Gemma 2; NVIDIA releases Nemotron-4 340B as its largest open model.

4. 2024-07

   Llama 3.1 405B; Mistral Large 2

   Meta's 405B becomes the first openly-released frontier-tier weight set. Mistral Large 2 ships at 123B with 128k context.

5. 2024-09

   OpenAI o1-preview; Llama 3.2 (vision); Qwen2.5

   o1 introduces test-time compute scaling at the frontier. Llama 3.2 adds vision to the open family. Qwen2.5 ships across seven sizes.

6. 2024-10

   Claude 3.5 Sonnet (new); Computer Use

   Anthropic ships an upgraded 3.5 Sonnet and the Computer Use beta — first frontier-lab native agentic control.

7. 2024-11

   Claude 3.5 Haiku; Qwen2.5-Coder 32B; QwQ-32B

   Lower-tier frontier models compress capability further. QwQ becomes the first open-weight reasoning model.

8. 2024-12

   OpenAI o1 GA, o3 announced; DeepSeek-V3; Phi-4; Llama 3.3 70B

   December 2024 was the densest release month of the cycle. DeepSeek-V3 in particular reset cost-per-token expectations.

9. 2025-01

   DeepSeek-R1; o3-mini; Mistral Small 3

   DeepSeek-R1 ships as the first open-weight o1-class reasoning model under MIT for the weights. o3-mini follows weeks later.

10. 2025-02

    Grok 3; Claude 3.7 Sonnet (with extended thinking)

    xAI ships Grok 3 with advertised 1M context. Anthropic adds extended thinking to the Claude line.

11. 2025-03

    Gemini 2.5 Pro (preview); Gemma 3; OLMo 2 32B

    Google's 2.5 Pro reasserts the long-context lead. Gemma 3 and OLMo 2 32B fill out the open tier.

12. 2025-04

    Llama 4 Scout / Maverick; OpenAI o3 GA, o4-mini, GPT-4.1; Gemini 2.5 Flash

    Meta moves to MoE at the frontier. OpenAI's o3 reaches general availability. GPT-4.1 reframes the API line.

13. 2025-05

    Claude Opus 4 and Claude Sonnet 4

    Anthropic's Claude 4 generation ships, with Sonnet 4 offering a 1M-token context beta.

14. 2025-06-onward

    Continued open-weight pressure; reasoning-on-everything; preview models that may or may not graduate

    The pattern from mid-2025 into 2026: every frontier model gets a reasoning variant, every open lab compresses costs, and the gap between closed and open narrows on benchmarks but widens on agentic capability. For anything released after this page's June 2026 refresh, check the provider's pricing page directly.

Cost discipline — what to actually pay attention to

Model selection in 2026 is rarely 'which model is smartest.' For 80% of production workloads it is 'which model is good enough at the lowest sustainable unit cost.' Apply the following filters in order before you commit to a SKU.

• Output tokens cost 3-5x input tokens at every closed frontier lab. If your workload generates long completions, the output column is the one that matters — input pricing is misleading marketing.
• Cached input is real. Anthropic, OpenAI, and Google all offer prompt-caching discounts (often 50-90% off cached portions). If your system prompt is long and static, build for caching from day one.
• Batch API is 50% off at every lab that offers it (OpenAI, Anthropic, Google). For non-real-time workloads — overnight evals, document processing, content backfill — this is free money.
• Mixture-of-experts pricing is per-active-parameter, not per-total-parameter. DeepSeek-V3 is 671B total but bills like a 37B model. This is why open-weight MoE has eaten the cost frontier.
• Self-hosted breakeven on a frontier-open model lands somewhere around 50-200M monthly tokens depending on the model and the GPU rental rate. Below that, hosted APIs win on TCO; above it, your own infrastructure starts to pay back.
• Reasoning models are not always worth it. o1 and R1 charge for the hidden chain-of-thought tokens too. For tasks that do not need multi-step reasoning, a cheaper non-reasoning model gets you the same answer for a tenth of the cost.
• Context window pricing is non-linear. Gemini 2.5 Pro and Claude Sonnet 4 both have higher per-token rates above 200k. If you can summarize-and-cache instead of dumping the full corpus, you usually should.

What this page does not cover

How to keep this current