Anthropic's 80x miss broke the AI cost model — June 15 is the cliff

ServiceNow — a $150B enterprise software company that knows how to buy software — exhausted its full-year Anthropic budget by May. Its CDIO went on the record about it. Anthropic offered no SLA, no per-user telemetry, no comment, and no committed pricing. National Life Group's CIO summarized the buyer experience: "great for consumer usage but not great for companies."

That's the through-line for the week. Everything else — the 80x demand miss, the Colossus lease, the June 15 credit conversion, the Codex switching offer — flows out of an enterprise vendor whose financial controls are structurally behind its market position.

What actually changes June 15

Anthropic is converting every Claude subscription into a dollar-matched API credit pool. Programmatic usage through Cline, OpenCode, Zed, Conductor, T3, custom Agent SDK harnesses — all metered at list rates against a separate bucket with no rollover. The 70-90% implicit discount that power users were quietly running on subscription-tier access is gone.

The math is unkind. A team running an effective $1,400/month of API-equivalent value out of a $200 plan now pays $1,400. Same prompts, same outputs, new bill. Opus 4.7 separately tripled per-image token accounting, so any vision pipeline on the hot path is wrong by a further factor of three.

OpenAI shipped a two-month free Codex switching promo the same week. Deadline July 13. Ramp's April data put Anthropic ahead 34.4% to 32.3% — the first lead change — and OpenAI is paying to flip it back. Take the free window. Even a no-switch outcome gives you comparison data for the next negotiation.

Why the vendor broke

Dario said it on stage: planned for 10x growth, got 80x. The visible consequences were Claude Code silently nerfed for paying users, corporate accounts banned without notice, and a 220,000-GPU emergency lease from xAI's Colossus 1 — the CEO who called Anthropic "misanthropic and evil" three months earlier is now their landlord. Stitching GB200-class hardware into an existing H100/H200 fleet under one API contract is not a clean swap. Expect p95/p99 variance through the integration. Any Claude benchmark from before May 7 is stale.

The load-bearing fact for buyers isn't the capacity miss. It's that Anthropic's response to the miss — degrade silently, revoke access without disclosure, ship no observability — is now precedent. Every Claude-dependent pipeline carries availability risk with no contractual backstop, on infrastructure owned by a hostile competitor, with no native per-user attribution.

The 59% number that reframes the bill

Vercel published production telemetry across 200,000+ teams: agentic workloads now carry 59% of all token volume. Anthropic captures 61% of spend through Opus on reasoning. Google captures 38% of volume through Flash on throughput. Most production teams already route by task — multi-model is the default architecture, not the sophisticated one.

Three things break if your stack hasn't caught up. First, the eval harness. Single-turn accuracy hides 15:1 input-heavy token burn on multi-step traces; a planner burning 40,000 tokens before giving up looks identical to a clean run on pass/fail. Second, the cost model. Ratios fit on 3:1 input-output are off by ~5x on agentic traffic, and MCP without a knowledge graph layer wastes ~30% per session as system prompts and schemas re-tokenize on every hop. Third, the router. Treating each call as independent leaves session-aware routing, KV cache reuse, and tool-call reliability on the table.

The production pattern Abridge has disclosed — cheap model triages, expensive model reasons only when needed, LLM judges calibrated against human annotators — is converging. The 5-10x cost reduction at scale comes from routing, not model swaps.

What to do this week

The operator move splits into three concrete things, in order.

Run the cost audit by Monday. Pull every Claude-backed workload — Agent SDK, GitHub Actions, batch evals, third-party harnesses. Reconcile projected token burn against the new credit cap. The formula is: (current third-party token usage − plan credit equivalent) × API rates = your June 16 invoice. Teams that don't do this math discover the answer on the bill, the way ServiceNow did.

Deploy a gateway with per-user, per-feature attribution before the credit conversion lands. LiteLLM, Portkey, Helicone — pick one, tag every call, set spend alerts at the user and feature level. Anthropic explicitly offloaded observability to you. If your SOC can't tell a human from an agent in the logs and your finance team can't tell a feature from a user in the bill, you're flying instruments-out on the largest line item you'll add this year. (Note that LiteLLM itself is on CISA's KEV catalog this week — patch first if you're on 1.81.16–1.83.7, then deploy.)

Pilot Codex on one load-bearing workflow inside the 60-day free window. Not as a switching commitment — as a benchmark. Trajectory-level metrics, not pass@1: tool-call precision, steps-to-completion, cost-per-successful-task. The output is leverage in the next Anthropic conversation regardless of which model wins on your workload.

The deeper lesson is that enterprise AI revenue is reversible in ways SaaS revenue never was. Token spend cuts to zero overnight. The teams that built unit economics on a subsidy they didn't have a contract for are about to find out what their margin actually was. The ones who instrument before the cliff get to negotiate. The ones who don't get the bill.