Anthropic Uses Claude to Build Claude; New Benchmark Gives Opus 5.5 Only 45

Anthropic is using Claude to build Claude. This episode of the Latent Space podcast is exactly about this internal process. The interviewees are Katelyn Lesse, the engineering lead behind the platform, and Ang Jiang, the product lead. This platform has produced first-party tools such as Code, Tag, and Duet, and is also used by more than half of Anthropic's top customers.

The show mentions a figure: Anthropic now spends $1 in cost for every $2 in revenue. One commenter asked: does this accounting include data center take-or-pay contracts? Take-or-pay means you pay even if you don't use the capacity once it comes online. The show didn't elaborate. The question points to a common weak spot for AI companies: infrastructure bills.

Someone else in the comments asked: when Opus 5.5 launched, what changed in the harness (the toolchain outside the model)? Another person said they most wanted to hear about a change that was blocked before launch: what the model wrote, which check failed, and how it was ultimately fixed. A smooth release is the result; being blocked is the process.

Opus 5.5 was called by the podcast 'the world's number one LLM right now.' The guest's response was: number one depends on whom you ask. It is number one in the rankings, but in the new benchmark, things get subtle.

Cognition this week released FrontierCode 2.0, a coding benchmark specifically testing the ability to go 'beyond the spec.' Most coding evals look at whether a model can write code according to a specification. FrontierCode 1.0 asked: would you merge this PR (pull request)? 2.0 adds another layer: it doesn't just ask 'would you merge it,' it also asks 'should you merge it.'

The tasks are more diverse, broader in scope, and have a higher bar for mergeability. Prompts are deliberately written vaguely, leaving engineering decisions to the model. As a result, this benchmark is far from saturated.

Cognition's Walden Yan said the quality bar is very high, and while building the benchmark the team rejected 80% of PRs and sent them back to open-source maintainers for improvement. He said this is the eval that best matches his intuitive judgment of models. His colleague Cheng-Yuan Lee added: software engineering feels solved, but agents still miss details and fail to produce thorough implementations.

The score distribution says a lot. Opus 5.5 scores 45.5%, GPT-6 Astra 42.6%, and Claude Sonnet 5.5 40.8%, with only a small gap at the top. Further down, Claude Fable 5.1 scores 32.4% and Grok 4.7 28.5%. One user in the comments pointed out that Sonnet 5.5 is close to Opus 5.5 and GPT-6 Astra, but clearly ahead of Fable 5.1. Others were discussing rumors that Fable 5.5 is coming.

Opus 5.5 was released on September 22, 2026. Data from expense management company Ramp shows that Opus 5.5 accounts for 63% of Anthropic's spending on models. The company doesn't just sell Claude to customers; it also uses Claude heavily itself. Using Claude to develop Claude is a serious investment.

The new benchmark's stance is: even the strongest model only reaches 45%. There is still a long way to go before 'should merge.'

Cognition has published the evaluation methodology, scoring criteria, and sample tasks on its blog. Anyone who wants to verify a model's real-world engineering judgment can compare against them.