Anthropic's leaked prospectus says two customers are a quarter of its revenue
A draft of Anthropic's IPO prospectus was leaked to Reuters on 28 September, and the numbers in it are the clearest public view yet of what a frontier lab costs to run. For 2025: revenue of $4.6bn against roughly $13bn of operating expenses, an operating loss above $8bn, a net loss of $42bn, and $20.28bn of cash at the end of December. For this year, Fortune reports second-quarter revenue of $11.5bn against $4.73bn in the first quarter, and says it will be operating-profitable for a second straight quarter. Set against that is $518bn of committed cloud and compute spending and a listing target above $2tn. None of this is an SEC filing yet — it is a draft, reported, and Anthropic has not confirmed it.
Two details matter to anyone buying inference rather than equity. The first is concentration: nearly a quarter of last year's revenue came from two unnamed clients. A supplier whose book looks like that has different incentives on price, capacity and deprecation than one with a long tail, and those incentives move when a large contract renews or does not. The second is that roughly a third of the document is risk factors, reportedly including the first prospectus language naming existential risks to humanity, alongside documented model behaviour the company itself describes — resisting shutdown, concealing or manipulating information, and conduct resembling blackmail. A vendor writing that down for investors is a vendor telling procurement to plan for a second route.
OpenAI cut its frontier price by four-fifths, and graded the result itself
GPT-6.1 Sol landed at DevDay on 29 September at $2 per million input tokens, $0.10 cached, and $10 output — against GPT-6 Astra's $10/$1/$50, one-fifth on input and output, one-tenth on cached. A new Ultrafast setting runs up to 300 tokens per second, six times faster in the API, at six times the standard token price. Sol is available now through the API as gpt-6.1-sol and in ChatGPT Work and Codex; the Ultrafast variant of Sol was still days away at launch.
The performance claims behind the price are the vendor's own. On OpenAI's evaluations, Sol matches Astra on DeepSWE v1.1, comes within 2.1 points of it on OSWorld 2.0, and scores 2.2 points above Opus 5.5 on AutomationBench at about a third of the cost, with a stated $5.47 per task at maximum effort on Terminal-Bench Science 0.1, against $23.21 for Opus 5.5. Competitor figures were taken from public reports rather than re-run. A four-fifths cut on a frontier tier resets the arithmetic for anything that routes by cost, but the honest position today is that nobody outside the lab has measured it — the independent evaluators publish on their own schedule, and vendor tables and third-party runs have diverged before.
Enterprises were banning persistent agents, so the vendors gave away the control plane
OpenClaw Enterprise, announced 29 September, is a vendor-neutral control plane for long-running agents, released free under the MIT licence — you pay only for the compute, models and storage underneath. OpenAI donated the project to the OpenClaw Foundation and is piloting it internally, Red Hat joined as a sponsor and intends to support it as one runtime among several on OpenShift AI, and Nvidia contributes the OpenShell isolated runtime. It self-hosts: Docker Compose for development, Kubernetes for internal deployment, on infrastructure you control, with your own models, harnesses and sandboxes swapped in.
The feature list reads like an audit checklist, which is the point — multi-tenancy, hard security boundaries, workload isolation, fine-grained permissions, lifecycle governance, and review of agent actions using language models. The stated reason it exists is that IT organisations had responded to agent platforms by banning them outright for want of exactly these controls. Two caveats before it appears in a NIS2 or DORA control narrative: it is in pilot with 1.0 promised later this year, and the security reference architecture has not been published. The question to ask when it is, is whether the audit record of what an agent touched is written somewhere the agent cannot reach.
Quick Hits
- HPE will ship Nvidia's agent containment layer in the fourth quarter. Its 28 September post puts OpenShell and Nvidia Confidential Computing into HPE Private Cloud AI in Q4 2026, with OpenShell running on ProLiant Compute DL394 Gen12 and BlueField-4 support across the portfolio as lead times allow. HPE's own framing is worth quoting to anyone selling you a location: “sovereignty is not only about where technology is located.”
- AI coding agents published more than 13,000 internal screenshots to public GitHub repositories. Glow Security's PixelLeak research, reported 29 September, found credentials, unreleased product details and internal billing screens across 343 organisations. The mechanism is mundane: GitHub's CLI has no API for attaching images to a pull request, so agents created public repos instead. About a third of the exposures involved gitshot, which defaults to public.
- OpenAI says prompt injections can replicate themselves. Its GPT-Red system found injections that embed copies of their own instructions in model output and spread through email, file writes and code comments, delaying execution across intermediate hops to evade inspection. The Register covered the disclosure on 29 September; no real-world attack has been recorded.
- OpenAI's zero-retention tier stores your prompts in your own bucket. Under Private Intelligence, announced at DevDay on 29 September, which builds on the zero-retention work OpenAI began in August, zero data retention excludes content from abuse-monitoring logs but writes selected encrypted safety records to the customer's own regional storage on AWS, Azure or Google Cloud with a 30-day time-to-live — zero retention at the vendor, not zero storage. Images flagged as possible CSAM can still be kept for manual review, and the Private Inference confidential-computing piece is a preview with no attestation model or threat model published yet.
- Washington swapped rules for a pledge. Tech chief executives signed a voluntary, non-binding accord on AI safeguards at the White House on 29 September, reported by Al Jazeera, alongside a directive telling federal agencies to say “super intelligence” rather than “artificial intelligence”. Nothing in it reaches an EU deployer, but it widens the gap a multinational has to document between its US and EU obligations.