Anthropic just killed the subsidy — and reset the whole stack in one week

On Monday an engineer running Claude through Cursor was paying somewhere between a tenth and a third of API rates. On June 15 she pays list. Anthropic converted every subscription plan into a dollar-matched API credit pool and unbundled third-party tool usage into a separate bucket that bills at full rate once the plan's dollar equivalent is spent. The 70-90% implicit discount that funded the harness ecosystem — Cursor, Cline, OpenCode, Zed, T3 — is gone in 30 days.

That's the news. The reason it's the news is the timing. Anthropic hired a CFO. It just overtook OpenAI in Ramp's B2B panel, 34.4% to 32.3%. It's raising at $900B ahead of an October window. ServiceNow's CDIO said publicly they burned the full-year Anthropic budget by May because the vendor ships no per-user telemetry. Dario conceded the company planned for 10x growth and got 80x, which is why Claude Code quietly degraded for months and why Anthropic is now leasing 220,000 GPUs from Colossus 1 — xAI's cluster, Musk's cluster, the competitor's cluster. When financial logic overwhelms competitive logic between two people who publicly dislike each other, the market has decided something.

What it's decided is that the subsidy era is over and the margin cleanup for the S-1 starts now. Every product built on Claude subscription arbitrage lost 20-40% of effective runway on Friday. Most founders haven't put it on a board slide yet.

Yes, but — the counter-reading is that this is ordinary pricing discipline that would have happened whether an IPO was scheduled or not, and the harness discount was always a billing accident rather than a strategy. Fine. The CFO hire and the October window still rhyme, and the arithmetic on your COGS doesn't care about intent.

OpenAI answered inside hours with two months of free Codex for enterprise switchers on a 30-day enrollment window that closes July 13. Read that as displacement pricing timed to a moment of developer frustration, not as a durable price signal. Both vendors are watching the same dashboard.

The production surface already moved

While the pricing shock plays out, Vercel published the number worth pinning to the wall: 59% of production tokens across 200,000 teams are now agentic — multi-turn, tool-calling traces, not single-shot completions. Six months ago that figure was under 20%. Anthropic captures 61% of spend via Opus on reasoning. Google captures 38% of volume via Flash on cheap fan-out. Those are two different businesses inside one category, and any eval harness still measuring single-turn accuracy is measuring the minority of production traffic.

The operational consequence is that trajectory-level metrics — tool-call precision, steps-to-completion, cost-per-successful-task, recovery-from-error — are the new bar. If you can't tell me your p95 tokens-per-successful-task by next week, you don't have a cost model. You have a monthly invoice and a hope.

The perimeter broke at the same time

Same week: an 18-year-old pre-auth RCE in NGINX's rewrite module, which ships in something like 90% of production deployments. Traefik shipped two CVSS 10.0 auth bypasses — the rubric ran out of knobs — which makes every auth middleware sitting behind Traefik decorative until the patch lands. Argo CD's authenticated-read-to-plaintext-Kubernetes-Secrets bug (9.6) chains cleanly behind either. MOVEit shipped a 9.8 that pattern-matches the 2023 Cl0p campaign. LiteLLM hit CISA's KEV. PraisonAI went from disclosure to active exploitation in four hours.

Four hours is the number to hold against your patch SLA. If yours is measured in weeks, it's measuring the wrong clock. This isn't a coincidence of bad luck landing on top of bad pricing. It's what happens when AI-assisted vulnerability discovery compresses the disclosure-to-working-exploit window and defensive tooling gets reverse-engineered faster than it gets updated. TrustedSec pointed frontier models at all five major EDR products this week and found them structurally similar and readable in days rather than weeks. UK AISI confirmed Anthropic's Mythos cleared both simulated attack ranges — full network takeover, not persistence — which is a discrete step above the prior generation, not a smooth interpolation.

Mozilla's team wrapped Mythos in a custom harness and reported 271 Firefox bugs. Daniel Stenberg pointed the same model at curl with a generic scanner and got one low-severity CVE. Same model. 270x yield gap. The harness is the product now, not the model.

What to do this week

Three things, in order, before Friday.

First, patch NGINX and Traefik tonight, not this weekend. If patching means downtime you can't take, put a WAF in front. The PoC for the rewrite module lands inside a week; the Traefik diff is already public. If you run Argo CD, rotate every secret it could read during the exposure window — the patch alone doesn't close it.

Second, before Monday, pull every Claude-backed workload — Agent SDK, claude-p, GitHub Actions, harness usage — and calculate the June 15 bill at full API rates. Not the estimate. The number. If you don't have per-user, per-feature attribution today, stand up an LLM gateway with tagging and daily budget alerts this sprint. ServiceNow couldn't manage this passively at $9B revenue. You won't either. In parallel, enroll the Codex evaluation on the free window — even if you don't switch, the number gives you leverage in the renegotiation you're about to have.

Third, if your product is enterprise-facing, ship an agent-callable surface — an MCP server against your existing API is a two-to-four week build — before the next renewal cycle. SAP's €100M partner fund, ServiceNow's Action Fabric, and Notion's External Agents API all landed in the same week for a reason. The procurement question in the demo has already changed from show me the dashboard to can our agents call this directly. Two vendors couldn't answer that last month. The one that could moved to the next round.

The subsidy, the perimeter, and the agent surface all reset in the same seven days. Model the cost by Wednesday, patch the ingress by Thursday, and put the MCP scope on the Q3 roadmap by Friday.