Claude Opus 5.5 Tops Cognition's FrontierCode 1.1

Artificial Analysis has released initial results for Cognition's FrontierCode 1.1, a benchmark designed to assess whether coding agents produce software changes that human maintainers would actually accept. Unlike traditional evaluations focused strictly on unit tests, this 100-task private benchmark grades code against repository conventions, scope boundaries, and architectural fit established alongside more than 20 open-source maintainers. To earn credit, a model must satisfy all blocking criteria; failing any blocker results in a score of zero, while passing attempts score between 70% and 100%.

Claude Opus 5.5 earned the top aggregate score at 52%, closely trailed by a dense cluster of nine models scoring between 43% and 52%. However, performance comes with wildly divergent operational expenses across the leaderboard. Anthropic's Claude Fable 5.1 costs an average of $12.00 per task to reach a 49% score, whereas OpenAI's GPT-6.1 Sol achieves the same 49% mark for just $0.81 per task. This difference stems largely from output token volume: GPT-6.1 Sol uses 33,000 output tokens per task and GPT-6 Astra uses 27,000, compared to 157,000 for Opus 5.5 and 236,000 for Sonnet 5.5.

Other models offer specialized trade-offs in speed and deployment flexibility. DeepSeek V4.1 Flash completes tasks in roughly seven minutes—less than half the 17 minutes required by Opus 5.5—while posting a 44% score alongside Kimi K3. In open-weights options, MiMo-V2.6-Pro posted the highest result at 45%, trailing the top commercial system by seven percentage points. Meanwhile, GPT-6 Luna marked the lowest recorded expense at $0.09 per task, and nine of the benchmark tasks were solved by no evaluated model whatsoever.

For software engineering teams, FrontierCode 1.1 demonstrates that raw test completion does not guarantee production-ready code. Because the benchmark penalizes invasive refactoring and unnecessary file creation, practitioners must weigh absolute model capabilities against execution time, token budgets, and cost per task. With even the highest-performing model capping out at 52%, human code review and strict pull-request gates remain essential when integrating automated coding agents into existing development pipelines.