Anthropic broke the Claude economy in a week nobody was ready for

On May 12, Anthropic converted every Claude subscription into dollar-matched API credits and set June 15 as the day third-party harness usage (Cursor, Cline, Zed, OpenCode) starts drawing from a separate, capped credit pool. After that pool drains, overflow bills at full API rates. The 70-90% implicit discount that quietly underwrote most Claude-wrapper economics and every heavy developer's workflow — gone.

On the same news cycle, OpenAI offered two months of free Codex to any enterprise that switches within 30 days. Ramp's April data has Anthropic at 34.4% of business AI spend versus OpenAI's 32.3% — the first documented lead change. The counter-offer was priced against the exact developer cohort Anthropic just alienated. Both windows close before most planning cycles complete.

The capacity story underneath explains the pricing. Dario Amodei disclosed at Code with Claude that Anthropic planned for 10x growth and got 80x. ARR moved from $9B to $30B+ in roughly four months. The relief valve is a lease on xAI's entire Colossus 1 cluster — 220,000+ GPUs — from the CEO who called Anthropic "misanthropic and evil" three months ago. Rivals don't rent from enemies unless the alternative is worse.

The silent Claude Code degradation users have been logging since April was never a product change. It was quantization, smaller-model routing under load, and scheduler unfairness dressed up as a roadmap. Anthropic ships no per-user telemetry, no meaningful SLAs, and no enterprise dashboard that would embarrass a mid-tier SaaS vendor. ServiceNow — one of the most sophisticated enterprise buyers on the planet — burned its full-year Anthropic budget by May and had to build AI Control Tower internally to figure out why. It now sells that tool to its own customers. That is the market routing around a vendor deficiency.

Yes, but — the counter-reading is that this is IPO hygiene. Anthropic hired a CFO, is targeting an October filing, and needed the arbitrage gone from the S-1. Prices stabilize after October, developers absorb the change, the wrapper economy adapts. The problem with that read is Anthropic's ARR tripled in four months while customers absorbed prior price hikes without churning. This isn't cleanup. It's pricing power, and there will be more of it before October.

What actually changed for anyone shipping on Claude

If your team uses Claude through a third-party harness, your per-developer cost assumption is wrong by close to an order of magnitude starting in 30 days. Model ten engineers on Pro plans running Claude through Zed eight hours a day. That is the case to price against API rates, not against the plan number on the invoice. Opus 4.7 separately tripled per-image token accounting, so anything that fans out across image batches now pays 3x for identical bytes.

The production data confirms this isn't a single-vendor story anymore. Vercel's AI Gateway, covering 200,000+ teams, reports 59% of production tokens are now agentic multi-turn traces. Anthropic captures 61% of spend via Opus for heavy reasoning. Google captures 38% of volume via Flash for cheap throughput. Teams are already routing across providers organically. Any harness hardcoded to one provider SDK is out of step with what production actually looks like.

Which means eval harnesses built on single-turn accuracy are measuring the minority workload. Cost-per-successful-task, tool-call precision, steps-to-completion, recovery-from-error — these are the trajectory-level metrics that map to what the bill actually reflects. A cost model built on last year's 3:1 input-to-output ratio is off by roughly 5x on spend forecasting for agentic workloads. The forecasts in most finance decks right now are wrong in a specific, quantifiable way.

The operator move this week

Before Friday, one calculation: (current third-party token usage − plan credit equivalent) × API rates = your July invoice. If that number surprises anyone on the leadership team when it lands, do it now instead of finding out in July.

Before end of sprint, one build: an LLM gateway (LiteLLM, Portkey, or a thin custom wrapper) in front of all Claude traffic with per-user, per-feature tagging and daily token budget alerts. Anthropic has explicitly offloaded observability to the customer. ServiceNow's failure is the cautionary tale — you either build this now or discover you needed it after the fact.

Before July 13, one experiment: run OpenAI's free Codex trial against your top workloads with matched prompts and tool schemas. Even if you don't switch, the comparison is the contract leverage. A free evaluation window against your incumbent is an asymmetric-payoff bet, and the payoff decays with every day you don't take it.

The pricing didn't change quality. It changed who pays for the arbitrage. That was always going to be the customer eventually. It's the customer now.