Anthropic Tells IPO Investors Claude Helps Design Its Chips
Per a report from The Information, Anthropic has a 15-person chip design team that already uses Claude for engineering and design work on custom silicon. Investors were told Claude cuts design time compared with engineers' estimates for Nvidia and AMD workflows.
Nvidia and AMD GPUs, AWS Trainium and Google TPUs stay central. The custom chips are meant to add to that pool. No plan to sell the design tools has been announced, and the speed claim is unverified. A reported amended S-1 also says Claude may “imminently” reach a recursive self-improvement threshold.
IPO Filing Reportedly Shows Nearly $895B in Compute Commitments
Coverage of Anthropic's S-1 on Nov. 3 put its future compute and infrastructure commitments at nearly $895 billion. That is the cost side of the story behind the revenue growth the filing also describes.
We could not open the full article, so check the filing itself for exact terms and timing.
Claude Code 2.1.317 and 2.1.318 Ship
v2.1.317 (Nov. 2) adds a cost report endpoint on the gateway admin API, with spend by total, developer, day or IdP group. Agent SDK hook callbacks can now set onFailure: "block". /schedule warns when a new cloud task gets connectors you did not request.
It also adds a confirmation step before a web or shell tool call sends your account email over the network, and new OTel attributes for plan and seat type. v2.1.318 (Nov. 3) warns earlier before your claude.ai sign-in expires.
Claude Code Now Refuses to Recreate Existing Games
A rule spotted in Claude Code's prompt says Claude should not build a recreation of a specific existing game, even if it is renamed or called original. Anthropic staffer Addy Osmani confirmed it is two added sentences.
Original games, genre games and games built from your own designs and assets should keep working. Critics note a prompt rule only deters low-effort copycats.
Users Push Back on Expanded Claude Watermarks
Reaction to the watermark rollout to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 is sharp. Users cannot opt out, it applies globally, and only EU-eligible organisations can detect it.
Critics say proofreading with Claude could trigger detection, while rephrasing with another AI could hide the mark. Anthropic admits short texts may not carry enough of the pattern. On Opus 4.6 and Sonnet 4.6, changing sampling settings reportedly removes it.
Receipts, Please: Chips, Commitments and Claims
Anthropic is pitching a loop: Claude helps build the chips that run Claude. It is a good story for a roadshow, especially against a reported $895B commitment bill. But a 15-person team and an unverified speed claim are not a tape-out.
Watch the next S-1/A or roadshow deck for a foundry, an EDA partner or a date. That is when the claim becomes checkable.