Claude Fable 5.2 in Devin scores 49.1% on FrontierCode 2.0, versus 32.4% for 5.1

A familiar story: the agent reports that the task is done and the tests are green, but code review reveals the code can't be merged. It ignores the repository's conventions, breaks a neighboring module, or drags in unnecessary changes. This gap between "it works" and "it can go into main" is exactly what the FrontierCode 2.0 benchmark from the Windsurf team tries to measure. It now has a new leader: Claude Fable 5.2, which has become available in Devin.

What the numbers show

According to Windsurf, Fable 5.2 scored 49.1% on FrontierCode 2.0. For comparison:

• Fable 5.1 — 32.4%.
• Opus 5.5 — 45.5%.

The gain over the previous version is almost 17 percentage points. This is not a cosmetic update but a noticeable jump in code quality. The gap to Opus 5.5 is more modest, about 3.6 points, but the model still came out on top.

The second part of the claim is more interesting for those who count the money. According to the authors, even at a low reasoning effort level Fable 5.2 outperforms the best result of Fable 5.1, while costing less than a third of the price per task. If this holds up on your data, the upgrade delivers two effects at once: the code gets better and the API bill gets smaller. Usually you have to trade one of these for the other.

How the evaluation works and why it matters

FrontierCode 2.0 evaluates models on real engineering tasks against criteria of quality and mergeability. The final score is a weighted sum of rubric items. There is also a hard rule: a solution that fails the blocking criteria receives zero. This is fundamentally different from familiar tests such as "did the unit tests pass or not." Code that formally works but violates project standards is not counted here.

On its chart, Windsurf calculates cost as the average dollar amount per run at public API prices, taking open discounts into account. In other words, this is not internal cost of goods but what you would pay. Such figures can at least roughly be mapped onto your own budget.

Limitations and what to do in practice

The main caveat is obvious: the benchmark was created and is run by Windsurf itself, which also sells Devin. That doesn't mean the numbers are unreliable. But it does mean there is no independent verification. Nor are there details worth knowing before drawing conclusions: how many tasks are in the set, which languages and stacks are covered, and how stable the result is between runs. The 3.6-point difference between Fable 5.2 and Opus 5.5 may be within the noise, or it may not; the publication doesn't let you tell.

So the practical takeaway is as follows.

• If you already use Devin, try Fable 5.2 on several of your typical tasks, including the unpleasant ones: refactoring, bug fixes in someone else's code, edits that must respect conventions.
• Start with low reasoning. The claimed savings, a third of the cost of the previous version's best result, are potentially the most valuable part of the release. Check whether this mode is enough for you before paying for high.
• Measure the metric that matters to you. The share of PRs accepted without changes and review time will tell you more than any public leaderboard.

Bottom line: Fable 5.2 looks like a strong update for agentic development, especially thanks to the "better and cheaper" combination. But for now this is a vendor's claim on its own benchmark, and the decision whether to switch should be based on results on your codebase.