Anthropic AI Updates: November 3, 2026

1. Messages API Adds Message Threads in Beta, So Long-Running Agents Send Only New Turns

Anthropic launched message threads in public beta on November 2 behind the message-threads-2026-08-12 beta header. The first request carries the full conversation, system prompt and tools plus "thread": {"type": "create"}; each later request sends only the new messages with "thread": {"type": "continue", "previous_message_id": "<id>"} and leaves out system and tools, which the thread keeps. Threads are held encrypted in memory and not written to disk, expire 5 minutes after the last request and are deleted no more than 1 hour after last use; a continue request against an expired thread returns a 404, so clients must keep their own copy of the conversation and replay it in a new create request, which the SDK helper client.beta.messages.thread() does automatically. Token usage, pricing and rate limits are unchanged because Claude still reads the whole conversation each turn; the gain is latency, with Anthropic reporting about 35% lower median time to first token for 500K to 1M token requests in Claude Code, where threads are now on by default for most users. Threads work on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry, but per the documentation they are not ZDR or HIPAA eligible, and because a thread keeps one model, a turn answered by a server-side fallback model is not stored and the conversation has to be replayed. Source

2. Anthropic’s Amended S-1 Reports $17.0 Billion in Q3 Revenue and Warns It May Soon Meet Its RSP Threshold for AI-Accelerated R&D

Anthropic filed Amendment No. 1 to its Form S-1 on November 2, adding results for the quarter ended September 30, 2026: revenue of $17.0 billion, up from $1.4 billion a year earlier, and an operating loss of $567 million (a 3% negative operating margin) after $974 million of operating income in the June quarter. The filing attributes the swing mainly to $5.0 billion of compute bought under short-term commitments, which costs significantly more than its multi-year capacity, and $2.0 billion of non-cash charges from its employee donation matching program; adjusted operating income rose to $2.5 billion, a 15% adjusted margin, and customers generating $100,000 or more in trailing revenue grew from more than 950 to more than 11,900 in a year. Nine-month 2026 revenue was $33.2 billion and net loss $104.0 billion, of which $94.3 billion came from fair-value changes on convertible notes, derivative liabilities and SAFEs, and commitments for compute, infrastructure and leases now total about $895.4 billion, roughly $447.5 billion of it due before the end of 2030. A new risk factor, first highlighted by TechCrunch, says internal estimates for future models, not current ones, “indicate that we may imminently meet” the Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI”, that Anthropic has in the past not identified such accelerations until after they occurred, and that it is considering RSP updates to keep humans in control of frontier model development because it does not expect to be able to safely let Claude autonomously build successive frontier models. Source

3. Anthropic’s Frontier Red Team Open-Sources FlashProve, a Zero-Knowledge System That Proves Which Weights Produced a Response

Anthropic published “Progress on provable inference for large language models”, describing FlashProve, a prototype in which a provider publishes a commitment to a model’s weights and attaches a zkSNARK proof that a response came from exactly those weights, without revealing them; Anthropic says Claude produced the performance improvements behind the system. On Llama 4 Maverick (a 400B-parameter mixture of experts with public 4-bit weights), one H100 with 24 CPU cores proves about 110 tokens per second, each proof for a 1,024-token request is about 19 MB, and verification takes about 17 seconds on a 24-thread CPU with no GPU, at roughly 70 to 95 times the GPU time vLLM needs to serve the same requests; the only earlier proof of the same model, VerInf, ran at about 0.02 tokens per second. The release includes a technical report, a protocol description, the prover, a fast C++ verifier, a slower verifier written in Lean that is proven to accept only valid proofs, and Lean proofs of completeness, soundness and zero knowledge, all under Apache 2.0. Anthropic positions it for detecting model substitution by routing services, tampered inference servers, and regulators checking that only approved models are served, and lists the limits: proofs cover an integer version of the model under greedy decoding, so sampled or floating-point outputs can diverge, contexts are capped at 1,024 tokens, the architecture is not hidden, and the fast code carries no formal guarantee. Source

4. Claude Code 2.1.317 Adds a Gateway Cost Report, Blocking Hook Failures in the Agent SDK, and an Email-Leak Check

Anthropic released Claude Code v2.1.317. The Claude apps gateway admin API gains a cost report, GET /v1/organizations/analytics/cost, that returns past spend over a date range in total, per developer, per day or per identity provider group, backed by a new daily record of each developer’s IdP groups on gateways with admin: configured, retained for admin.identity_retention_days. Agent SDK hook callbacks can now set onFailure: "block" so a callback that errors blocks the action instead of letting it run, and Claude Code asks before a web or shell tool call would send the user’s account email address over the network. OpenTelemetry metrics, events and spans gain anthropic.plan, anthropic.seat_type, anthropic.provider and anthropic.custom_base_url attributes; /schedule warns when a new cloud scheduled task receives connectors the user did not ask for; cloud sessions show a link card for links an MCP server asks the user to open; and setting CLAUDE_CODE_DISABLE_NESTED_SESSION_ID=1 on a self-hosted runner restores the earlier behavior in which claude -p run inside a session shares that session’s ID and transcript. Source

5. Anthropic Confirms Claude Code on Opus 5.5 Now Steers Away From Recreating Specific Existing Games

Anthropic added two sentences to Claude Code’s system prompt on Opus 5.5 telling the model to build original games, genre games and games from the user’s own designs and assets, but not “a recreation of a specific existing game, even if it is renamed or called original.” A user first spotted the rule in late October, XDA found it in the Claude Code 2.1.314 binary released October 30 under the flag name tengu_heron_brook, and Anthropic’s Addy Osmani confirmed the change, saying “using Claude to infringe IP is against our usage policy,” that original and genre games “are meant to keep working,” and that Anthropic will “continue to iterate” and wants to hear about refusals of users’ own projects. Ars Technica reports GitHub issues where the prompt is getting in the way of some legitimate game development, and notes the rule lands amid a wave of AI-assisted game decompilation and porting, but that developers can switch to other frontier or open-weight models without such restrictions. Ars Technica The Verge