Anthropic ($16.96B), Anthropic (−35% TTFT) and Trajectory ($300M) — autonomy paid in compute

Anthropic reports $16.96B in third-quarter revenue and $895.4B in compute commitments while warning it may be about to cross its AI-accelerated progress threshold. Anthropic launches message threads in public beta and cuts time-to-first-token by 35% on long Claude Code sessions. Trajectory raises $45M at a $300M valuation so non-engineers can post-train models on real feedback. Agent autonomy is being paid for in three places at once: infrastructure bills that no longer fit in a quarter, latency that drops without touching the per-token price, and capital looking for whoever can adapt the model without a research team.

Anthropic ($16.96B) — Q3 at $16.96B, $895.4B in compute and RSI threshold warning

On November 2 Anthropic filed its S-1/A. Revenue for the quarter ended September 30: $16.96B, versus $1.38B a year earlier (+1,127%) and $11.56B in Q2. Sequential growth fell to 47% from 144% the previous quarter. Operating loss of $567M after the only prior quarter of operating profit, largely attributed to roughly $5B of more expensive short-term compute and a $2B non-cash charge for donation matching. Net loss $23.29B, mostly fair-value of convertibles.

The filing totals purchase and lease commitments for the first time: about $895.4B, “substantially all” in compute and infrastructure, with a large portion due by 2030 and beyond. In the same document it warns that its internal estimates on future models (not current ones) indicate it may “imminently” meet its Responsible Scaling Policy definition of “dramatically accelerated progress attributable to AI.” No current model is designated above the threshold. More than 11,900 customers with ≥$100k TTM.

Anthropic (threads) — public beta: −35% TTFT on 500k-1M contexts

On November 2 Anthropic opened message threads in public beta on the Messages API. The agent sends the full conversation once and afterwards only the new messages plus the previous response ID. The server holds the thread in memory, encrypted and not written to disk. It expires after 5 minutes of inactivity; then the client must resend everything. Claude still reads the full context on every turn, so token usage does not change. Prices and rate limits stay the same.

In Claude Code (on by default for most users) median time-to-first-token drops by about 35% on requests of 500k to 1M tokens. Reports of long sessions mention payloads falling from ~800 KB to ~1.8 KB. Useful for coding agents, computer-use and heavy document or image workflows. It does not save money; it saves waiting. If the session pauses for five minutes the benefit disappears.

Trajectory ($45M) — Series A at $300M for continual learning without engineers

Trajectory, a one-year-old startup founded by three Stanford classmates (ex-DeepMind/Windsurf and Apple), announced a $45M Series A on November 2 led by Sequoia, with Nvidia, Bessemer and Bain Capital Ventures. Valuation $300M. Post-training / continual-learning platform: connect an open-source model to your applications, collect production feedback and the system retrains. Aimed at lawyers, analysts or doctors defining “what good looks like” without building a research team.

Named customers: Harvey, Rogo, Decagon, Mercor and Nvidia itself. 18 people. They are opening the platform to the public. The pitch is classic: a foundation model is a brilliant PhD who starts a new job every day; what the economy values is the person who has been in the role for 30 years. The real friction remains the cost of generating and filtering high-quality feedback at scale, and that the base model has to be good enough for the post-training to matter.

The connecting thread The three pieces point to the same bottleneck. Anthropic books $17B in a quarter and already has nearly $900B in commitments because agent demand outruns long-term contracted capacity and forces expensive short-term compute purchases. Message threads cuts wait time on long sessions without touching the per-token price: the saving is latency, not the bill. Trajectory raises money so post-training stops being a lab project and moves into the hands of the domain expert. Autonomy is not getting cheaper; it is being reconfigured around who pays the wait, who pays the retraining and who absorbs the infrastructure bill.