Anthropic Killed the Claude Arbitrage the Same Week It Rented Musk's GPUs

On May 12, Anthropic converted Claude subscriptions into dollar-matched API credit pools. Effective June 15, every programmatic call through Cursor, Cline, Zed, OpenCode, GitHub Actions, or the Agent SDK meters at list price. The 70-90% implicit discount that heavy users have been pulling — the $200 plan that quietly bought $700-$2,000 of API-equivalent work — is gone in thirty days.

The same week, Ramp's card-spend data put Anthropic at 34.4% of US B2B AI spend against OpenAI's 32.3%. First documented lead change in the category. OpenAI answered inside hours with two months of free Codex for enterprise switchers who commit within a 30-day window. And Anthropic — planning for 10x growth, hit with 80x — signed a lease for the entire Colossus 1 cluster from xAI. Two hundred and twenty thousand GPUs, roughly 45% of xAI's capacity, from the CEO who has publicly called Anthropic "misanthropic and evil."

Read the sequence together. New CFO. October IPO. Margin recovery dressed as developer defense, timed to pre-IPO diligence, executed while renting compute from a sworn enemy because there was no other option. That is not a company with headroom. That is a company with a book to close.

Yes, but — a fair counter-read is that Anthropic is simply repricing a subsidy that never should have existed, and enterprises absorbing the increase without churning proves the product is load-bearing. That's true, and it's why the take still holds: load-bearing revenue with zero SLAs, zero per-user telemetry, and a competitor's datacenter as the relief valve is exactly the kind of revenue that gets marked down when someone with public-market discipline looks at it.

The number ServiceNow already ran

ServiceNow's CDIO said publicly that the company burned its full-year Anthropic budget by May. She could not tell which users or which workloads drove it, because Anthropic ships no granular usage attribution. National Life Group's CIO called Claude "great for consumer usage but not great for companies." PagerDuty reports the same shape. If a company that ships enterprise observability for a living cannot govern its own AI spend, the assumption that your finance partner's spreadsheet describes reality is a strong one.

Here is the mechanic teams keep missing. You price the feature with healthy margin over inference cost. Users find the workflow that saves two hours. They run it eleven times a day instead of the three the model assumed. Retention improves, usage depth climbs, gross margin goes sideways and then down. That is the base case for any AI feature good enough that people adopt it, and the new pricing model exposes it directly.

Vercel's production data — 200,000+ teams over seven months — puts agentic traffic at 59% of all token volume, up from under 20% six months ago. Anthropic captures 61% of spend via Opus on planning nodes; Google captures 38% of volume via Flash on utility calls. Cost models built on last year's 3:1 input-output ratio are off by roughly 5x on agentic traces, and the error is not symmetric across vendors. One Vercel customer burned a monthly budget in nine days. The cause was not a bad prompt. The workload mix shifted underneath the model.

Three things to do this week

First, compute the new number. Take last month's Claude usage through every third-party harness and multiply by full API rates. If you cannot produce that number by Friday, that is the finding — you have no cost attribution and you are about to discover ServiceNow's problem in your own finance review. Deploy a gateway with per-tenant, per-feature tagging (LiteLLM if you must, though patch it first — it landed on CISA's KEV catalog this week; Portkey or in-house works). No Claude call should leave your infrastructure without a tenant_id, feature_id, and prompt-family hash.

Second, run the Codex trial. Not because OpenAI is obviously better — the honest read is that on your top task classes it probably isn't, quite — but because the two-month free window closes July 13, and even a no-switch outcome produces a comparison number you wave at Anthropic during renewal. Vendor leverage peaks when both labs are fighting for the same logo. That window is open now and will not be open in Q4.

Third, ship the abstraction layer you have been deferring. A provider-switching config, not a rewrite. A few hundred lines of code buys you the ability to route Opus for reasoning and Flash for throughput — which is what Vercel's data shows mature teams already do — and it buys you the ability to fail over silently when Anthropic's capacity math stops closing. Which it will again, because 80x against a 10x plan is not a one-time miss.

The teams that model the new economics before June 15 will renegotiate from a position. The teams that discover the change in July's invoice will do what ServiceNow did — build the telemetry after the fact, staff a dedicated headcount to watch usage, and explain to the board why the AI line item quadrupled. One of those conversations goes better than the other.