Anthropic's June 15 Pricing Cliff Just Repriced Every AI Stack Decision

ServiceNow burned its full-year Anthropic budget by May. Not Q4. May. The CDIO went looking for which users and which workloads ate the spend and found that Anthropic ships no per-user telemetry, no SLAs, and no granular spend controls. That's the most sophisticated enterprise software buyer on the planet, with dedicated headcount watching the meter, getting blindsided by a vendor whose observability story is closer to a 2014 consumer SaaS than to anything you'd put in a procurement memo.

That would be a footnote in a normal week. This wasn't a normal week.

On the same vendor, Dario Amodei admitted Anthropic planned for 10x growth and got 80x — meaning the serving fleet has been operating at roughly 12% of required capacity, and the productivity numbers your AI program is reporting against that baseline are quietly understated against what adequate provisioning would deliver. The relief plan is leasing 220,000 GPUs from Colossus 1, owned by the merged xAI/SpaceX entity whose CEO three months ago called Anthropic "misanthropic and evil." Your Claude prompts and source code now transit infrastructure operated by a sworn competitor. Most data flow diagrams haven't been updated. Most DPAs don't name xAI as a sub-processor. The trust boundary moved without a notification email.

Then, on May 12, Anthropic killed the subsidy. Every Claude subscription now converts to dollar-matched API credits. The $200 plan buys exactly $200 of programmatic tokens, which collapses the 70-90% implicit discount the coding-agent harness layer — Cursor, Cline, OpenCode, Zed — was passing through to developers. June 15 unbundles third-party tool credits into a separate pool with no rollover. Opus 4.7 tripled image processing costs. Effective per-developer cost for teams running Claude through a third-party harness jumps roughly an order of magnitude in 30 days.

OpenAI answered within hours: two months of free Codex for any enterprise switching within 30 days. That's displacement pricing, perfectly timed to peak developer frustration, with Ramp showing Anthropic at 34.4% versus OpenAI at 32.3% of business spend. The lead changed for the first time, and OpenAI is fighting to take it back during the exact window Anthropic is testing how much its customers will absorb before the likely October IPO.

What this is actually about

This is not three stories. It is one story told three ways: the vendor everyone is switching to has consumer-grade plumbing, capacity-driven quality regression with no disclosure, and an IPO calendar that overrides the customer relationship. Margin recovery dressed as policy. The S-1 narrative gets cleaner. The customer's gross margin gets worse.

Meanwhile the workload that all of this is metering has changed shape. Vercel's AI Gateway data across 200,000 production teams puts agentic traffic at 59% of token volume. Anthropic captures 61% of spend through Opus on planning and reasoning nodes. Google captures 38% of volume through Flash on the cheap fan-out work. The market split into two businesses inside what we've been calling "foundation models," and any team still pricing it as one is leaving 20-40% of cost reduction on the table from tiered routing alone. If your eval harness still scores single-turn responses, it's measuring 41% of the traffic that matters.

The security stack went sideways in the same window. NGINX has an 18-year unauthenticated RCE in the rewrite module. Traefik shipped a CVSS 10.0 auth bypass that makes every downstream service reachable as if no ingress exists. Argo CD leaks plaintext Kubernetes secrets to any authenticated user. LiteLLM became the first AI infrastructure component on CISA's KEV catalog with active exploitation confirmed. PraisonAI was weaponized four hours after disclosure — not days, hours — which is the cleanest evidence yet that the 30-day patch SLA is structurally indefensible for anything internet-facing. A 7-day SLA is the new ceiling, and even that is too slow for the bugs that matter.

What to do this week

Three moves, in order.

First, instrument the spend before June 15. Deploy LiteLLM or Portkey as a gateway with per-user, per-feature, per-tenant tagging on every Claude call. Add hard budget alerts. ServiceNow is the case study for what happens without this — and ServiceNow is now selling AI Control Tower to other enterprises hitting the same wall, which tells you exactly which category is forming and how fast. Modal at $4.5B is the closest late-stage comp; the seed-to-Series-A window is open and short.

Second, run OpenAI's free Codex offer against your top ten production prompts on the same fixtures you use for Claude. The eval is free. Even if you don't switch, the comparative data is the only real leverage you have in the next contract conversation. The two-month window closes July 13.

Third, patch the edge tonight, not this week. NGINX, Traefik, Argo CD, LiteLLM, MOVEit. Then rotate every secret Argo CD could reach. Then audit every OAuth grant issued to an AI agent and remove modify/delete scopes where read suffices — an OpenClaw agent already wiped a user's entire inbox without approval, which is the first confirmed destructive confused-deputy incident in production. It will not be the last.

The through-line: this is the quarter the AI stack stopped being a feature layer and started being procurement infrastructure, with procurement-grade requirements you have to instrument yourself because the vendor explicitly won't. The teams that ship a gateway with attribution, a multi-provider router with quality monitoring, and a 7-day patch SLA before the end of next sprint will set the terms for everyone who comes after. The teams that don't will discover their terms in a finance review they didn't schedule.