Pinned Tweet
We just benched the latest version of Descant, Soprano 7, at 82.3% on DeepSWE v1.1—ahead of every frontier model.
Good timing, too, because we’re opening Descant to teams on the waitlist tomorrow.
Descant brings multiple independent perspectives to planning, implementation, and review. It challenges assumptions, resolves disagreements through evidence, tests how components work together, and simplifies the code before changes land. It carries what it learns into the next run and keeps the remaining work aligned as the codebase evolves. Fully autonomous, with as much direction up front as you want to provide.
It built OH HAI from a single issue: a central hub connecting you and your AI agents across vendors and environments, so questions, tasks, and critical updates don’t get buried in long-running threads. It also refactored and evolved TriviaPot—a real-time multiplayer trivia app—across thousands of issues. Both are on the App Store.
Now I want to get it working on the tech debt, feature backlog, and massive refactors your team hasn’t had time to tackle. Or turning a single issue into an entirely new, production-ready product. No slop grenades.
The result below measures how it handles individual long-horizon engineering tasks. Descant is built to work far beyond that—evolving a codebase over months and years, across thousands of interconnected issues.
More on the benchmark, what we learned, and getting access: descant.run
We also published all 113 benchmark task repos and the solutions—93 passing. If you’d like to dig into the code, PRs, and task notes, you can request access here:
github.com/descant-soprano-7…
If you’re curious, I’m happy to walk you through Descant and answer questions. I’d also love to hear what you’d want it to take off your plate.
Reply here or send me a DM.
descant.run built OH HAI and the MA2H protocol behind it, fully autonomously. OH HAI is now on the App Store—free, no ads, no selling your data. And security is a priority: every change has to pass the protocol’s conformance tests to merge. LMKWYT producthunt.com/products/oh-…
I surpassed my previous most contributions in a year on GitHub (2.2K) in 3 days with our multi-model software engineering system we are building. ~13.8K contributions in the last 60 days. Fully autonomous.
Oh, and it scored a 86.46% on SWE Bench Pro.
Opus 4.8 is currently the highest scoring model available with 69.2%. Mythos 5 / Fable 5 scored 80.3%. Our architecture allows models to be hot-swapped and takes advantage of the top strengths of each. The highest scoring model currently plugged in is Opus 4.8. The average score of all models currently plugged in is 63.9%. Our system brings a +22.6% upgrade. For comparison, Mythos / Fable only scored 11.1% better than Opus 4.8 and that gap has been described as the difference between now and the stone age. We've doubled that gap. Once we plug in Fable / Mythos our new target will be 94%.
Its called Descant and its solely focused on very high quality software engineering, semi or fully autonomous. We are targeting the first release, Soprano 1, in July 2026.
Made the Descant SWE-bench Pro result artifact public:
633 / 731 resolved = 86.59%
The original post said 86.46%; one very long-running task was still finishing when I posted, and it ended up passing.
The repo includes the predictions, winning diffs, resolved list, grades, and reproduction instructions against the official harness.
This is the result artifact only — not the full Descant system — but the benchmark output is now inspectable:
github.com/autnmy/descant-sw…
A little more context on the SWE-bench Pro result:
Descant is not a single model. It is a multi-model software engineering system with separate loops for planning, implementation, review, validation, and recovery.
The 86.46% score came from the system, not from any individual model plugged into it.
That distinction is the whole thesis: long-horizon software engineering performance is increasingly a function of the engineering process around the models, not just the base model itself.
While benchmarks are cool, the thing that actually matters is the production-ready quality of the code. Below is a very early screenshot (design NOT final) of our 2nd product in development, OH HAI, a universal inbox to answer questions, complete tasks, and get updates from agentic fleets. That screenshot is from the first boot of the app, ever. 0 errors starting it up. Built entirely by Descant, per our technical direction and using the MA2H protocol ma2h.org/ (which we also designed and published).
