Models 28 August 2026 5 min read
Summit, and the longer think
Our most capable model reasons for as long as the problem needs. Here is what changed, and when to reach for it.
Summit is the third model in the family and the one we reach for when the question is genuinely hard: a legal review across two hundred pages, a refactor that touches forty files, a plan with more assumptions than facts. It is available today on Pro and Max, and through the API as dune-summit.
What is different
The obvious change is the context window. Summit reads a million tokens in one pass, which in practice means a whole repository or a whole deal room without splitting anything up. The less obvious change is that Summit decides for itself how long to think. On an easy question it answers as quickly as Ridge does. On a hard one it can spend several minutes checking its own reasoning before it says a word, and it tells you it is doing so.
We built the longer think because the failure we heard about most was confident nonsense — an answer that arrived fast and read well and was wrong in the third paragraph. Summit's extra time goes on exactly that paragraph. In our own evaluations it halves the rate of unsupported claims on long documents, and it flags the places where the source material disagrees with itself instead of quietly picking a side.
When to use it
- Anything with real stakes and a lot of material: due diligence, an audit, a grant application.
- Code changes that cross module boundaries, where the cost of a subtle mistake is a week.
- Research where you want the argument laid out, not just the conclusion.
For most everyday work Ridge remains the right default. It is faster, it costs a fifth as much, and it is very good. Summit is for the days when you would have booked a meeting with the smartest person you know.
Pricing and limits
Summit is priced at $15 per million input tokens and $75 per million output tokens. On Pro it counts against your usage at five times the rate of Ridge; on Max the multiplier is the same but the ceiling is far higher. Extended thinking is on by default and can be capped per request with the max_thinking parameter.
More updates
- 14 July 2026 Dune in your editor A native extension for VS Code and a command-line client. It reads the repository, proposes the change, and runs the tests.
- 2 June 2026 Memory you can see Dune now remembers across conversations — and shows you every single thing it has kept, with a switch beside each one.
- 19 April 2026 Why Dune shows its sources A note on the design decision behind the little citation line, and why we think confidence without evidence is the wrong product.