Image via latent.space
Opus 5.5 shipped the day before this conversation was recorded, and Anthropic says a large, previously unpublicized share of the launch work went into something other than capability: making the API harder to distill from.
Distillation means training a rival model on another model's outputs, and the obvious objection is lost revenue. Katelyn Lesse, Head of Platform Engineering at Anthropic, told Latent Space the commercial case actually runs the other way, and the real concern is what a copied model does without Anthropic's safeguards.
"Realistically, if you distill from us, you actually pay us, so commercially we like make the money... it's just more on like making sure that our model's like... Claude could be derived, and if it ends up negative, that would be unfortunate." — Katelyn Lesse
Katelyn runs the engineering behind the Claude API, and Angela Jiang, Head of Product for the Claude Platform, owns the product side. Between them they built the separate "brain" and "hands" design behind Claude Managed Agents and moved the entire Messages API to Rust in roughly a year.
The full interview is covered here so you can skip it. 1 hour 47 minutes of audio, 10 minutes of reading.
Here are the 9 insights that matter.
Latent Space: The AI Engineer Podcast · Oct 27, 2026 · 1h 47m · 97m saved
With Katelyn Lesse, Head of Platform Engineering at Anthropic · Angela Jiang, Head of Product, Claude Platform at Anthropic
Key Takeaways
Much of the Opus 5.5 launch went into resisting distillation, and the stated reason is safety, not revenue
Training-run incidents are staged on purpose: models are put in a box and told to solve an impossible task by any means
Claude Managed Agents keeps credentials in a vault the agent never touches, with a proxy injecting the real key only as traffic leaves the sandbox
Every model launch now carries its own policy decisions, plus a single button that pulls access back for everyone
Anthropic rewrote the Messages API in Rust with no Rust developers on the team
Claude wrote the code, and the port saved roughly 40–50% of CPU
A safety classifier has to be as capable as the model it gates, which is why Claude sometimes drops to a weaker model mid-answer
Inside Anthropic, unreviewed Claude output can be sent back to whoever produced it
1. Opus 5.5 vs. Distillation
Katelyn and Angela said much of the Opus 5.5 launch went into preserving Claude's "thinking blocks" across turns. That change stops a chain-of-thought extraction technique, published responsibly by outside researchers in a paper on stealing reasoning traces from proprietary models, from working.
Anthropic's motive for blocking distillation is safety, Angela Jiang said, because a copy would lose Claude's safeguards
"The entire point of why Anthropic exists is 'cause we want AI to create positive outcomes in the world safely, and we run a lot of safeguards classifiers on the tokens that we're producing." — Angela Jiang
A model distilled from Claude would not carry those classifiers, and could produce bad outcomes nobody intended. Before the change, the team worked directly with the maintainers of open-source agent harnesses. It also shipped a new compaction feature, for condensing long agent sessions, so customers whose approach depended on the old behavior would not break.
