Cognition launches FrontierCode 2.0, where its top model scores 45%

• FrontierCode 2.0 grades underspecified coding changes on mergeability. Claude Opus 5.5 leads with a 45.5% weighted score in Cognition's launch results.
• GPT-6.1 Sol scores 39.0% at $1.03 per rollout, versus Opus 5.5's $12.82, according to the leaderboard screenshot.
• Independent runs produce a similar ceiling: Opus 5.5 reaches 46.4% in Artificial Analysis's results, averaged over three attempts per task.

The best trap is a Mock Service Worker fix that restarts generators but strands the user's cleanup code. In a PostHog SDK task, turning up reasoning effort led GPT-6 Astra to delete a public class while fixing a race condition.

Cognition calculates rollout costs at list API prices, including public discounts, according to its scoring notes.

• Cost gap: At max effort, GPT-6.1 Sol trails Opus 5.5 by 6.5 percentage points at about 8% of its rollout cost, using the leaderboard's figures.
• Ranking reversals: Muse Wonder 1.0 and Gemini 4 Argon lead DeepSWE v1.1, but fall behind Opus, Astra and Grok on FrontierCode 2.0 in Cognition's comparison.

An attempt that fails any blocking criterion receives zero. Surviving attempts receive a weighted rubric score, according to Cognition's methodology, so the headline score is distinct from the leaderboard's pass rate.

The release post reports three expansions:

• Repository coverage: 55% more source repositories.
• Grading: 43% more blocking criteria covering correctness and design requirements.
• Task size: Against FrontierCode 1.1, reference solutions modify 67% more files and 89% more lines, in codebases that are 25% larger on average.

The evaluation also grades design choices, interpretation of ambiguous prompts and integration with existing code.

Thoroughness blockers increased from 168 in FrontierCode 1.1 to 501 in 2.0, according to Cognition's rubric breakdown. These check obligations the agent can infer without an exhaustive prompt:

• Fix sibling paths: Correct the same bug elsewhere in the codebase, even when the prompt names only one occurrence.
• Match existing support: Handle additional input types when neighboring implementations establish that expectation.

In Mock Service Worker, Opus 5.5 fixed resetHandlers() by discarding the old generator reference. The next request starts a fresh generator, but the abandoned generator never closes, so user cleanup inside finally never executes, as Cognition's walkthrough explains.

Across four models at high effort, the failure rate for thoroughness checks rose from 11.7% to 19.2% between benchmark versions in the published measurements.

GPT-6 Astra's PostHog results show how additional reasoning can hurt. The task walkthrough compares fixes for a startup race between cached and live analytics events:

• Low effort, 5/5 passing runs: Move cache loading into PostHogQueue.kt, remove sending logic from the legacy integration and preserve the public API.
• Max effort, 1/5 passing runs: Often delete PostHogSendCachedEventsIntegration.kt and redistribute its responsibilities. The race disappears, but removing the public class breaks compatibility.

Cognition also acknowledged that earlier scope blockers rejected some reasonable solutions. It loosened those criteria in 2.0; the reported scope-check failure rate fell from 9.0% to 3.8% across effort levels.

The boundary between thoroughness and over-editing drew discussion about subjectivity in automated review.

Artificial Analysis evaluated all 132 private tasks three times using its Stirrup harness. Its benchmark documentation adds several implementation details:

• Each run is capped at 500 turns, with tasks weighted equally.
• Shell commands can continue in the background while the agent works.
• GPT-6 Luna at medium effort serves as the LLM judge.
• Attempts flagged for looking up the solution receive zero, as do attempts failing a blocking check.

Its scores differ from Cognition's launch figures: Opus 5.5 at max effort with fallback reaches 46.4%, Haiku 5.5 reaches 34.9%, and DeepSeek V4.1 Flash 1017 reaches 28.5%.

FrontierCode 2.0 also receives 15% of the weight in Intelligence Index v5, alongside Terminal-Bench 4.0's 10% within the coding category.

Open-source maintainers spent more than 40 hours building each task, according to Cognition's account. Human and automated audits changed reference solutions relative to the original upstream PR in 80% of tasks, up from 60% in FrontierCode 1.0.

One audit agent also discovered a security vulnerability in an upstream implementation. Cognition says it reported the issue to repository maintainers after human review in its quality-control account.