What exists today
OpenAI's models page lists GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna. Anthropic's newest are Claude Opus 5.5 (released September 22, 2026), Sonnet 5.5 (September 28, 2026) and Fable 5.1 above them. These are the models to build on now; all of them are callable through Model.sale.
Neither company publishes a long-term roadmap. An OpenAI executive has reportedly said most of its research targets future generations, but OpenAI has not named a GPT-7 model or a date. Anthropic's model pages and release notes list nothing beyond the 5.x family.
How much to trust the rumors
Third-party estimates put Claude 6 in the first half of 2027 and GPT-7 in the second half of 2027. They extrapolate from the gaps between earlier releases; nobody outside the two companies has a verified date. A prediction market gave Claude Opus 6 less than a 1% chance of shipping before November 1, 2026, and posts claiming early access to GPT-7 have been debunked.
Treat any page that quotes GPT-7 or Claude 6 prices, benchmarks or context sizes today as invented. When a release happens, the model ID and price appear first in the vendor's own documentation.
What will probably change for developers
New flagships tend to arrive at the same or a lower price than the model they replace: Claude Opus 5.5 launched at $4 / $20 per million tokens, below Opus 4.8 at $5 / $25. Expect a new model ID and a new price line rather than a higher bill for the same work.
Expect behavior changes too. Opus 5.5 made thinking always on and turned forced tool use into an error; streaming clients that assumed every block contains text broke. A new generation will need the same kind of check, not just a string replacement.
How to be ready on day one
Keep the model ID in configuration (an environment variable or a settings file), not in code. Keep five real tasks from your backlog as a private test set with the expected result. Parse response blocks by type, and never hard-code a price: read the charge from the API usage or the dashboard.
On release day, create a separate key with a small daily spend limit, run the five tasks on the new model and on the current one, compare attempts and charges, then switch the configuration. With an OpenAI-compatible or Anthropic-compatible gateway the base URL and key stay the same; only the model ID changes.
When new models appear on Model.sale
Model.sale publishes a model only after a live JSON and streaming check, a usage check that billing depends on, and a published price. New models usually become callable once our suppliers serve them and those checks pass. GET /v1/models with your key and the models page always show what is callable right now.
What to use until then
For coding: Claude Sonnet 5.5 ($0.90 per million charged tokens on Model.sale), GPT-6.1 Sol ($0.40) or the cheaper GPT-6.1 Terra ($0.15) as the daily driver, Claude Opus 5.5 ($1.40) or GPT-6 Astra ($1.125) for the hard cases. The comparison of one coding session on these models is in our Claude Opus vs GPT-6 article.