AI Model Releases 2026: Every Launch, Newest First

This page is a dated list of the AI models released in 2026, newest first. I update it whenever a new model ships.

September 2026

GPT-6 Astra (3 September, OpenAI). The first widely available model that can drive a computer on its own. A 1.05 million token context window, 10 dollars in and 50 dollars out per million tokens. Access is rolling out in stages.

Claude Fable 5.1 (1 September, Anthropic). Built for the hardest reasoning and long running autonomous agent tasks. The price stayed the same and cache reads dropped from 1 dollar to 0.25 dollars. On agent setups that work with long context, that cut shows up clearly on the bill.

August 2026

No new model shipped in August among the ones I track. The notable move was on pricing. Anthropic dropped the increase it had planned for Claude Sonnet 5, so 2 dollars in and 10 dollars out per million tokens became permanent. Good news for teams that had built a budget around that model.

July 2026

The densest month of the year. Four separate releases in three weeks.

• Claude Opus 5 (24 July, Anthropic). Code, multi step agent work and long document analysis. A 1 million token context, 5 and 25 dollars. The effort setting lets you scale cost to the weight of the task.
• Gemini 3.6 Flash (21 July, Google). Unusually strong for the price. It scores above Google's own Pro tier on the independent intelligence index and spends fewer output tokens than 3.5 Flash.
• Kimi K3 (16 July, Moonshot AI). The first open model at 2.8 trillion parameters. It took first place on an independent frontend code arena. The weights are downloadable, so you can run it on your own hardware.
• GPT-5.6 family (9 July, OpenAI). Three tiers at once. Sol at 5 and 30 dollars, Terra at 2.50 and 15 dollars, Luna at 1 and 6 dollars. While Astra rolls out in stages, most corporate work still runs on this family.

Earlier in 2026

Claude Sonnet 5 (9 June, Anthropic). The balanced model for daily work at 2 and 10 dollars. Good value for writing, analysis and moderate coding.

Gemini 3.1 Pro (Google). At 2 million tokens, the largest context window on this list. Built for huge document sets and multimodal input such as video, audio and PDF.

What is next?

Gemini 3.5 Pro has been delayed. The expected features include a 2 million token context and autonomous workflow abilities. The stated reason is unresolved inconsistency in the rebuilt model.

Reading the prices correctly

Model pricing is per million tokens, with input and output billed separately. A token is roughly three quarters of a word. Output costing far more than input is not an accident: generating text is a much more expensive operation than reading it.

The practical consequence: work that reads long documents and answers briefly is cheap, work that takes a short question and produces long text is expensive. The same model can produce completely different bills across two workflows.

Note: the Claude Sonnet 5 price is introductory and runs through 31 August 2026. Gemini 3.1 Pro moves to 4 and 18 dollars for requests above 200 thousand tokens. These details change often, so confirm on the provider's own page before deciding.

Do these numbers even apply to you?

Probably not. Those are API prices, what you pay when you wire a model into your own software. If you open the ChatGPT, Claude or Gemini app and use it there, you pay a flat monthly subscription and token math never touches you.

For app users the real question is which model sits behind which plan. Roughly: free plans get lighter models, mid tier paid plans get strong models, top plans get the flagships. For most everyday work the free tiers are still more than enough.

How much should you trust benchmarks?

Every model launch comes with a chart showing it beating everything else. Treat those charts carefully:

• Most measurements come from the vendor. Independent tests arrive weeks later and usually land on more modest results.
• The tested task is not your task. Performance on olympiad math questions says little about how well a model writes your customer emails.
• The gaps are usually small. A few percent looks dramatic on a chart and is imperceptible in daily use.
• Specialization is real. Independent testing of Kimi K3 found it leading in some areas and trailing in others. There is no single ranking.

A better method: take a real task from your own work, give the same prompt to three models and compare. The answer you get in ten minutes beats any comparison table.

How much does switching models actually gain you?

The honest answer: less than you think. For most users the quality gap between models is small next to the gap created by how you ask. A badly written prompt produces bad output on the strongest model too.

A concrete example: someone who types "write me an Instagram post" gets mediocre text no matter which model they use. Someone who defines the role, the audience, the length and the constraints gets usable output even from a free model. That is why prompt writing habits deserve attention before model comparisons do. The free prompt library has ready made templates and the five part prompt formula.

Which one for which job

Writing, email, everyday questions

Claude Sonnet 5 or GPT-5.6 Terra. Both are more than enough here without charging flagship prices. In the chat apps even the free tiers usually do the job.

Long documents, contracts, archive analysis

Gemini 3.1 Pro's 2 million token window is the largest capacity available. But if you need deep reasoning over what it reads rather than just reading, Claude Opus 5 produces better results.

Code and app development

Claude Opus 5. Its coding benchmark results and self verification behavior mean fewer correction rounds on long tasks. For anyone building apps through vibe coding, this is where the difference shows most concretely.

High volume automation

GPT-5.6 Luna or Gemini 3.6 Flash. If you run thousands of small jobs a day, the bill is set by token price, not benchmark scores. These two exist precisely for that.

Organizations whose data cannot leave

Kimi K3. The only model here with downloadable weights. It needs a technical team and servers, so it is not practical for small businesses. But for organizations handling sensitive data it is the only real option.

Conclusion

The list changes every month but the logic of choosing does not. Define the job first, then pick the cheapest model that does that job. I keep the current price and context comparison side by side on the model table.

If you want to work out which model your team should use for which task, see corporate AI training.