The Week in One Paragraph
Anthropic's S-1 is the event of the week. It gives buyers the first audited look at frontier-lab economics: 11,900+ customers spending over $100K a year, 250 spending over $10M, 900% net revenue retention, and a +19% operating margin (up from -221% in Q1 2025). Revenue grew 1,300% while compute cost grew 300%, which shows that cheaper tasks bring more usage, not lower spend. The same filing says agentic workloads use disproportionately more compute per revenue dollar, which puts pressure on future pricing and capacity. OpenAI's GPT-6.1 Terra fits the same pattern. It is a $1/$5 mid-tier model aimed at computer use and document work, and it halves the price of its predecessor tier. The week also brought two lessons about evaluation. Leash Bio showed that grading agents against a known answer key exposes decision-quality failures that outcome-only evals miss. The 404 Media episode showed a public assistant sending users to a third party that could not help them, a failure accuracy evals don't catch.
The Three Things That Mattered
1. Anthropic is now an audited, profitable vendor with founder control. The filing lowers counterparty risk on multi-year commitments. Governance cuts the other way. Seven co-founders hold 50.1% of the vote through a single Class F share, Class A holders have no practical say, and the Long-Term Benefit Trust can replace a board majority. Strategy and safety posture are decided by a small group that outside shareholders cannot influence. The filing also calls "closing the loop" (each Claude helping build the next) its core advantage and admits Anthropic may only detect acceleration after it has happened. Model behavior can change faster than the vendor's own forecasts.
2. The ROI benchmark for AI-built software now exists. More than 90% of Anthropic's September code was written by Claude. Since April 2026, internal Claude spend has exceeded payroll excluding stock comp. Revenue per employee went from $1.4M to $6.6M. This is the first credible public number for what an agent-heavy engineering org spends on tokens relative to people. Expect boards to cite it, and expect CFOs to be surprised that the result is a large token bill rather than a smaller one.
3. The commercial target has moved to computer use, which makes routing and security core concerns. Terra matches Sol and Astra on vision and computer use, runs much faster, and is weaker on most other tasks. OpenAI now sells four overlapping tiers, so "default to the flagship" no longer works as a policy. Model choice has become a per-task routing decision that only your own evals can make. Screen-driven agents on cheap, fast models also increase prompt-injection exposure, and the release offers no evidence of mitigation.
Direction of Travel
- Unit price falls, total spend rises. Per-tier prices drop about 2x within a generation, and Opus 5.5 cut per-task cost by about 40%. Usage grows faster than prices fall. Budgets should model volume growth, not savings.
- Agentic compute is the cost variable that can blow up a budget. Anthropic says agents use more compute per revenue dollar and that it has underestimated capacity needs before. Rationing, priority tiers, or repricing for agentic use are the likely next moves from vendors.
- The reasoning setting is now a cost line of its own. One max-effort Terra run cost about 11 cents and took over 4 minutes. Spend is now shaped by model choice multiplied by reasoning level.
- Rankings change every ~11 days. Leash could not keep its public leaderboard current. Open-weight models (DeepSeek V4 Flash, Qwen, GLM, Kimi) are competitive on hard domain tasks, which supports private deployment where data sensitivity requires it.
- Evals are shifting from scoring outcomes to scoring decision quality. Environments with known ground truth expose early quitting, redundant retesting, and choices that contradict the agent's own hypotheses. Proprietary data that can serve as an answer key is becoming a strategic asset.
What To Do This Week
1. Re-price live business cases. Any model built on GPT-5.6 Terra/Sol or Opus 5 pricing is stale. Rerun the numbers with the new tiers, separate cost by reasoning level, and assume volume growth. 2. Build a routing eval harness for one client workload. Run Terra, Sol, Opus 5.5, and one open-weight model on the client's actual tasks. Report cost, latency, and accuracy per task class. That deliverable is worth more than any vendor's tier names. 3. Draft agentic capacity terms. Use the S-1's own risk disclosures as leverage to ask for committed capacity, price protection on agentic SKUs, and notice periods for model deprecation or behavior changes. 4. Add regression tests for vendor model updates. "Closing the loop" means behavior can shift between versions. Pin versions where possible and gate upgrades on your own eval suite. 5. Prototype an answer-key eval. Find a client process with historical ground truth (claims adjudication, ticket resolution, procurement decisions) and score agent decision paths against it, not just final outputs.
Customer Conversations to Have
- CFO / procurement: "Your AI bill will go up as unit prices fall. Here's the Anthropic benchmark, and here's how to budget by workload and reasoning level."
- CIO / architecture: "Single-vendor exposure now includes founder-controlled governance and agentic capacity risk. A routing layer and a portable eval suite are the hedge."
- Engineering leadership: "Anthropic has 90% of its code written by AI and spends more on tokens than on payroll. What's your target ratio, and what does the eval and review pipeline look like at that scale?"
- CISO: "Computer-use agents on fast, cheap models multiply your prompt-injection exposure. No vendor has shown a fix. We test this in your environment before any RPA replacement."
- Customer experience owners: "Public chatbots are already answering questions about your support paths, and some answers are wrong. We should monitor what they say, and constrain your own assistant's claims about escalation to verified content."
- Regulated / data-sensitive clients: "Open-weight models are competitive on hard domain tasks. Private deployment is now a performance-viable option, not only a compliance concession."
Risks and Watch-Items
- Agentic repricing. Watch for capacity-based SKUs, priority tiers, or rate-limit tightening from Anthropic and OpenAI over the next two quarters.
- Capacity constraints. Anthropic has under-provisioned before and is exposed to chip supply and Taiwan risk. Single-region, single-vendor deployments carry availability risk.
- Governance drift. Post-IPO, founder control means policy changes on safety, usage terms, or acceptable use can happen without shareholder check. Track the LTBT's board actions.
- Undetected capability jumps. Leash saw chemistry competence appear abruptly in February 2026, and Anthropic admits acceleration may go undetected. Workloads that depend on stable behavior need continuous evals.
- Sycophantic misrouting. The 404 Media pattern (assistants inventing referrals and remedies) is anecdotal and its cause unconfirmed. It is a plausible failure for any customer-facing deployment and is invisible to accuracy evals. Add tests for invented referrals.
- Prompt injection through screens. Terra's computer-use push comes with no published mitigation. Treat every screen and web surface as untrusted input.
- Revenue concentration. Anthropic's revenue is concentrated in large accounts. A downturn in any one of them could change its pricing behavior.