Claude Opus 5.5 scores 45.5% on FrontierCode 2.0 against 54.6% on the earlier version. At xhigh effort it scores 43.4% for $5.18 a rollout, where max effort costs $12.82.
A 6.5 point gain in coding score comes at 12.4 times the cost. On Cognition's new FrontierCode 2.0 benchmark, Claude Opus 5.5 leads a 25 model board with 45.5% at $12.82 a run, while GPT-6.1 Sol scores 39.0% for $1.03.
The same Opus model scored 54.6% on version 1.1 of the test. The test became stricter, so every score fell. Cost now matters more than the ranking alone.
The 45.5% is a weighted rubric score. A solution that fails any blocking criterion gets zero, and Cognition lists a separate pass rate of 49.7% for the same run.
The top four sit within 6.5 points of each other but span $1.03 to $19.94. The leaderboard reports mean spend per rollout at list API prices. Sol delivers 86% of the Opus score at one twelfth of the cost.
What extra reasoning effort costs
The effort setting is where the money goes. Opus 5.5 at xhigh effort scores 43.4% for $5.18. Against Sol, that is 4.4 more points for $4.15, about $0.94 a point.
Moving to max effort is the costly choice here. The step from xhigh to max adds 2.1 points for $7.64, about $3.64 a point. Pass rates tell the same story: 47.3% at xhigh against 49.7% at max, with Sol at max passing 42.3%.
Opus at xhigh also beats GPT-6 Astra at max, 43.4% against 42.8%, for $1.50 less. Lower settings sit further down the same curve. Opus at high effort scores 34.5% for $2.42, and at medium 30.1% for $1.70.
Sol at high scores 34.8% for $0.65. That matches Sonnet at xhigh for about a fifth of the cost.
Sonnet 5.5 offers the poorest value in this comparison. At max it scores 41.0% for $19.94, which is 56% more than Opus for 4.5 fewer points. At xhigh it scores 34.8% for $3.31.
At the cheaper end, each additional point costs much less. GPT-6 Luna scores 21.9% for $0.17. Haiku 5.5 adds 5.8 points for $0.20 more, about $0.034 a point.
Sol adds 11.3 points over Haiku for $0.66, about $0.058 a point. Astra adds 3.8 points over Sol for $5.65, about $1.49 a point. Opus at max adds 2.7 points over Astra for $6.14, about $2.27 a point.
The same cost logic holds across generations. Opus 5.5 at low effort scores 14.7% for $0.75, the same score as Opus 4.8 at max for $12.13. At max effort, Opus 5.5 scores 18.9 points above Opus 5 for $0.18 less ($12.82 against $13.00).
The routing rule follows. Send first pass pull requests to Sol at $1.03 or Haiku at $0.37. Buy Opus at xhigh for changes where a failed blocker costs a human review cycle. Max effort has to earn $3.64 a point.
Scores fell from version 1.1 because the test became stricter. Cognition's blog says the 132 tasks come from 55% more repositories. Reference solutions change 67% more files and 89% more lines, in codebases 25% larger.
Tasks also carry 43% more blocking criteria each, and blockers across the suite rose from 1,345 to 1,853. Thoroughness blockers, which check that a fix reaches every code path and cleanup step it should, tripled from 168 to 501.
Cognition loosened one thing. Scope failures, where a model changes more than it should, fell from 9.0% to 3.8% across four top models. The harder test is therefore mostly a test of finishing the job. The blog calls the benchmark "far from saturated."
Every model lost ground, and the weaker ones lost most. Opus 5.5 fell 17%, from 54.6% to 45.5%. GPT-6 Astra fell 20%, from 53.3% to 42.8%.
The lowest scorers fell hardest. Grok 4.7 fell 40%, from 47.6% to 28.6%. Kimi K3 fell 60%, from 44.2% to 17.7%. Cognition says its results show "significant room for improvement for open source models."
The gap between best and worst widened. Across the eight models in Cognition's before and after table, it grew from 10.7 points to 27.8, or 2.6 times. A test that separates tiers is more useful to a buyer than one where everyone clusters in the fifties.
Thoroughness is where models fail most. It fails 26.0% of checks at low effort and 15.3% at max, pooled across Opus 5.5, Fable 5.1, GPT-6.1 Sol and GPT-6 Astra. Core behavior, where the main logic is wrong, falls from 20.9% to 12.3%.
Scope is the only category that worsens with effort, from 3.8% to 4.9%. More effort fixes sloppiness but adds changes nobody asked for. A buyer paying for max effort is paying partly for that trade.
Where models still fail
Cognition's walkthrough shows a bug in Mock Service Worker, a library for mocking network requests. Opus 5.5 fixes the problem as described but skips the user's cleanup code in a finally block. It fails in all five runs at each of five effort levels.
Cognition built the tasks with more than 20 open source maintainers, who spent over 40 hours on each one. It wrote the test and sells the Devin coding agent, so treat the cross model price gap as the safer signal. Its own model, SWE-2, ranks 13th at 23.8%.
In 80% of the 2.0 tasks the reference solution differs from the upstream pull request, up from 60% in version 1.0. Runs that consult the original pull request score zero.
Cognition's blog places Opus on the high end of the Pareto frontier, the set of models no rival beats on both score and cost. Luna, Haiku and Sol sit at the low cost end of that frontier. It singles out GPT-6.1 Sol for "particularly strong performance relative to its cost."
The leaderboard moves quickly. Version 1.0 launched on 8 June, 1.1 followed on 7 July, and the public board went live on 17 July. Cognition added nine models in the 35 days from 22 September to 27 October, including Sonnet 5.5, Haiku 5.5 and Gemini 4 Argon.
Its own conclusion is that "agentic software engineering remains far from solved." At that cadence the cost ladder will be tested again within weeks. The next entry that scores above 39.0% for under $1.03 a rollout moves Sol off the frontier and resets the price per point.