How Anthropic ships Claude with Claude: Separating the Brain from the Hands

What it is

Latent Space interview with Katelyn Lesse (engineering) and Angela Jiang (product) of Anthropic, recorded the day after the Claude Opus 5.5 launch. The parts useful to an API integrator are the API-side changes behind the launch and how model access is gated.

What changed for API users

• Preserved thinking blocks. Anthropic has enforced preserving thinking blocks across turns in the last couple of models. Enforcement is now also aimed at extraction across multiple accounts. Lesse says chain-of-thought extraction is among the most powerful distillation routes.
• Compaction breakage. Harnesses that compact context manually can break under this requirement. Anthropic added a compaction feature to the API that replaces the manual approach and works with preserved thinking.
• Grace period. When preserved thinking shipped for Fable 5.1, existing customers were grandfathered for a time. Anthropic tracked them with a trust grant checked on the request hot path.
• Outreach. Anthropic contacted open source harness maintainers and large customers before the change.
• Effect. Lesse and Jiang say distillation is measurably harder in the areas Anthropic can measure. Jiang frames the motive as safety (models run without safeguard classifiers), not revenue.

Access and data policy

• Per-model policy. Data retention requirements are decided for each model launch. Opus 5.5 does not carry the retention requirements that Fable does.
• Verification programs. Cyber verification and a newer life science verification program relax online safeguards for approved users and use cases.
• Trust grants. A trust grant is an authorization record tied to access to a specific model or program. Lesse describes it as ordinary authorization, "IAM, but Anthropic's IAM". Authorization is threaded through the whole system, not handled with an API key plus a feature flag. The same primitive supports pulling back a model for everyone, the capability Anthropic wanted after the Fable de-deployment.

Other details

• Claude Code's default effort setting is medium. Jiang says defaults vary by segment (free, prosumer, Claude Code, Cowork, enterprise, API).
• Claude Managed Agents and another Anthropic product (the transcript says "Cloud Tag", likely a transcription error) share the platform team's sandbox fleet but not a harness.
• The brain/hands split and credentials held outside the sandbox are in the show notes and the existing summary. The credentials segment (about 1:05:25) is not in the text reviewed.

Caveats

• Only the first 60,000 of about 130,000 characters were available, so the Rust rewrite of the Messages API, safety classifiers, and the sandbox and credential discussion were not reviewed.
• The transcript is an unedited automatic draft with unassigned speakers. The wording above follows it and may contain errors.
• The page's intro quotes S-1 figures that differ from the spoken intro (a -126% versus -129% margin a year earlier). Treat the financial claims as unverified.
• Roadmap items such as memory and dreaming are described as slowed relative to safety work, and are not shipped commitments.