Devin’s own benchmark puts Claude Fable 5.2 on top at 49.1%

Cognition, the company behind Devin and Windsurf, said Claude Fable 5.2 is now live in Devin. On its own FrontierCode 2.0 benchmark, the model scores 49.1%, the highest on the board and well above Fable 5.1’s 32.4%.

A 17-point jump over Fable 5.1

According to Cognition, FrontierCode 2.0 judges models on real-world engineering tasks, looking at quality and whether the result is mergeable. Fable 5.2 takes first place with 49.1%. Fable 5.1 reached 32.4%, and Claude Opus 5.5 reached 45.5%.

Cognition adds that the gain is not limited to the top score. At low reasoning effort, it says, Fable 5.2 already beats the best result Fable 5.1 achieved, while costing less than a third as much per task. The company did not publish the exact dollar figures in the announcement text; they sit in a score-versus-cost chart on its leaderboard.

On scoring, the page says the score is a weighted aggregate of rubric items, and that solutions failing blocking criteria receive 0. Cost is the mean spend per rollout at list API pricing, public discounts included.

Users can reach the model through Devin Desktop and the Devin CLI, both linked from the post. The announcement does not mention a price for Devin plans, any usage limits, or whether the model is the default.

How the rest of the field scores

Cognition’s separate FrontierCode 2.0 write-up lists the other models it tested, each at its best reasoning effort. GPT-6 Astra scores 42.8%, Grok 4.7 28.6%, Gemini 4 Argon 26.3% and Muse Wonder 1.0 24.3%. Further down are SWE-2 at 23.9%, DeepSeek V4.1 Flash at 19.3% and Kimi K3 at 17.7%.

That puts Fable 5.2 about 3.6 points ahead of Opus 5.5 and 6.3 points ahead of GPT-6 Astra. Anthropic models hold the top three places among those listed except for GPT-6 Astra in third; the gap to the non-Anthropic field is wide, with the next rival after GPT-6 Astra more than 20 points behind.

The cost side matters as much as the score. In our earlier report on Fable 5.2 pricing and availability, the model lists at $8 per million input tokens and $40 per million output tokens, against $10 and $50 for Fable 5.1. A model that is 20% cheaper per token and, by Cognition’s account, needs fewer tokens to reach a given score would lower the bill twice over. Anthropic also sells a fast mode at double the standard rate, which Cognition does not discuss.

Our coverage of the Fable 5.2 and Mythos 5.2 launch covers the model itself; Cognition’s post concerns only how it performs inside Devin.

A benchmark run by a competitor in the market

The caveats are plain. FrontierCode 2.0 is built and run by Cognition, a company that sells a coding agent and has a commercial interest in showing that its tooling gets the most from new models. The announcement does not say how many tasks the benchmark contains, and the 49.1% figure has not been independently reproduced. Cognition says the tasks were written by more than 20 open-source developers from repositories they maintain, which is a reasonable design, but outsiders cannot yet check it.

Also unclear is whether the scores reflect the bare model or Devin’s own scaffolding around it. Results from other agents, or from the same models in other tools, would show how much of the lead belongs to Fable 5.2 itself.

The signals to watch are independent runs on public coding benchmarks, and whether other agent makers such as Cursor report similar gains once they add the model. If Fable 5.2 keeps its edge outside Cognition’s own test, the claim of a large step over Fable 5.1 will hold up.