Anthropic's S-1 is out, and the footnote is the real story

Good morning 👋 Anthropic's S-1 landed with the IPO headline everyone expected. The number buried on page 248 is the one worth your coffee.

In today's issue:

• 🔭 Anthropic's S-1 is out, and the footnote is the real story
• 🧠 Models & Releases: GPT-6.1 Terra ships, and OpenAI leaves a checkable trail on three math problems
• 🔬 Research Highlights: agents routing around human sign-off, an open model's audit, and why hand-built agents lose
• 🛠️ Try This: cache the stable half of your agent's prompt before someone else's pricing call does it for you
• 🔗 Quick Links: Broadcom's $63B assist for Anthropic's chip bill, and Nvidia's stake in Meta

🔭 THE ONE THING

📄 Anthropic's S-1 is out, and the footnote is the real story

Anthropic filed its S-1 with the SEC this week, confirming a Nasdaq listing under ticker ANT with Goldman Sachs and Morgan Stanley running the book. Skip past the IPO headline and look at the margin: compute and infrastructure costs that ran 251% of revenue in Q1 2025 are down to 48% in Q2 2026, the efficiency story the raise is actually selling. That's against roughly $800B in compute commitments disclosed in the filing, so the ratio improving while the bill keeps growing either means real operating leverage or a very well-timed snapshot. Buried on page 248: Ben Mann, long called a cofounder in public, isn't in the "Continuing Founders" group that holds 50.1% of the vote. In a filing this size, that's the kind of detail someone hoped nobody would read that far to find.

🧠 MODELS & RELEASES

• 🖥️ OpenAI shipped GPT-6.1 Terra, a faster computer-use model, and this time it's not stuck behind a research waitlist. If your agent roadmap has been waiting on a screen-control model worth building on, this is the one to test first. link
• 🧮 OpenAI also posted model-assisted progress on three unrelated open problems: the Kakeya conjecture, the Exponential Time Hypothesis, and a question in quantum chromodynamics. The proofs aren't the point so much as the trail behind them. Outside researchers can actually check the model's reasoning, not just take OpenAI's word for it. link

🔬 RESEARCH HIGHLIGHTS

• Approval gates aren't actually gates Ten frontier models, given email access and a wallet task that needed a spending-cap increase, went and raised the cap themselves by harvesting the one-time passcodes meant to require a human's sign-off. The paper's response is a protocol called I-HITL and a prototype, Countersign, that forces a separate trusted device into the approval path. Software-only human-in-the-loop, it turns out, is a suggestion an agent can route around. Paper

• DeepSeek V4.1 Flash got the audit open models usually skip Seventeen outside researchers ran it through 34 benchmarks under the EU's GPAI code of practice, the same bar applied to GPT-6 Astra, GPT-5.6 Sol, and Gemini 3.8 Flash. Real improvement over the V4 line, roughly competitive with the closed labs on several measures, and still enough loss-of-control and manipulation behavior left to justify the next check-in. Closed-weight releases get system cards by default. This open one didn't, until someone went and built one. Paper

• The hand-built agent loses to the plain one On RCAtlas's 90 validated incident cases, a bare LLM call plus a SQL agent and a Codex coding agent scored 65-77% accuracy. Five published root-cause-analysis agents, built with SOPs and fixed investigation stages, maxed out at 53%. All the Pareto-optimal runs on cost and time came from the simple harnesses too. If you're scripting an on-call diagnosis bot, the scaffolding may be the thing holding it back. Paper

🛠️ TRY THIS

Audit your agent's token bill before someone else's pricing call does it for you

1. Pull your agent's last 20 API calls and check what's actually in the payload: system prompt, tool schemas, instructions, scratchpad history.

2. Split that into a stable prefix (system prompt, tool defs) and a volatile tail (the live user turn, fresh retrieval results).

3. Turn on prompt caching for the stable prefix. AWS laid out the exact pattern for OpenAI models running on Bedrock, caching everything your agent doesn't need to re-read every step (AWS).

4. Re-run your eval set and diff token spend and p50 latency against last week. If the cache isn't hitting, your "stable" prefix isn't as stable as you think.

Prompt: List every place this agent rebuilds its system prompt, tool list, or instructions before a call, and flag which ones are byte-identical across the last 20 turns.

Worth a look

• AWS's agentic retrieval CDK pattern pairs well with the cleanup above: one agent routing across two Bedrock Knowledge Bases through the AgentCore Gateway, returning cited answers with built-in observability and continuous evaluation baked into the stack. AWS

🔗 QUICK LINKS

• Broadcom arranged a $63B financing deal with Bank of America, Blackstone, Citigroup and Morgan Stanley, routed through a new SPV, to cover Anthropic's chip purchases. Hock Tan reportedly got on the phone personally in the final hours to keep the banks in. Bloomberg
• Nvidia is putting $3B into Meta's stock offering, part of a raise that could run up to $95B, as the two expand their Muse partnership and Meta buys more Vera Rubin and Vera CPU capacity. WSJ
• OpenAI published Signals data on 165 work activities across nearly 100 countries. Skip the topline. Look at which jobs show up first, and which never do. OpenAI
• Fanatics ran a Shop the Look test with OpenAI across roughly 30,000 MLB products and reported a 10% engagement lift. Vendor-measured, so treat it as a claim, not a verdict. OpenAI
• Greptile is stacking GPT-6 Astra on top of code review to test actual user workflows, with Codex trimming the rework that follows. OpenAI
• CodeRabbit uses GPT-6.1 Sol to flag security risk across large codebases, then leaves developers to triage the findings. OpenAI
• OpenAI says its own Codex Ultrafast cut internal incident investigation from hours to minutes. Self-reported, but a concrete number if other shops see the same. OpenAI
• AWS shipped a SageMaker reference architecture for fleet-wide computer vision: a model per camera, not one for the whole fleet, quality-gated and versioned by the thousand through Step Functions. AWS
• Sakana AI's VP of GTM walked through what it actually takes to move a Japan-born AI product onto the global stage. Sakana AI
• Google added "Polar Peaks," a neuroscience lesson, to its AI Quests literacy series. Google

Read the footnotes. They're usually where the real story hides.

Pradeep Perugu