Anthropic's June 15 pricing cliff exposes the real AI supply chain risk

On June 15, every Claude subscription converts to dollar-matched API credits. Third-party harness usage — Cursor, Cline, Zed, OpenCode, custom SDKs — loses the 70-90% effective discount it has been running as COGS for the last year. Overage bills at metered API rates. Any Claude-backed workflow still priced on subscription math is wrong by roughly an order of magnitude in 30 days.

That is the operational headline. The structural one is worse.

Dario Amodei admitted planning for 10x growth and hitting 80x. The April Claude Code degradation was a capacity miss, not a product decision. The fix is a lease of xAI's entire Colossus 1 cluster — 220,000+ GPUs owned by a CEO who has publicly called Anthropic "misanthropic and evil." ServiceNow burned its full-year Claude budget by May and couldn't tell which users caused it, because Anthropic ships no per-user telemetry and no SLAs. National Life Group's CIO said it plainly: great for consumer, not great for companies. This is the vendor Ramp now clocks at 34.4% of business AI spend versus OpenAI's 32.3%. First crossover.

OpenAI's counter is two months of free Codex for enterprise switchers within 30 days. Take the free benchmark whether or not you switch. The comparison data is the point.

The workload shape moved and your evals didn't

Vercel's AI Gateway index across 200,000+ teams and seven months puts agentic workloads at 59% of production token volume. Anthropic captures 61% of spend on Opus reasoning nodes. Google captures 38% of volume on Flash throughput. These are two different businesses inside the phrase "foundation models," and multi-provider routing is already the default in production. Application code pinned to a single vendor's SDK is out of step with what the market shipped.

Which means most eval harnesses are measuring the minority workload. Single-turn benchmarks against reference answers were correct in 2023. Today the median request is a multi-step tool loop with retries, and what breaks in prod is a planner burning 40,000 tokens arguing with itself. Cost models fit on 3:1 input-output ratios are off by roughly 5x against real 15:1 agentic traces, and the error is asymmetric across vendors.

If 59% of your tokens are agentic and 100% of your evals are single-turn, you are flying instruments-out. Update the harness before you update the model.

The counter-reading

Ramp measures credit-card billing, which skews toward developer and SMB adoption. The 2.1-point lead over OpenAI could reverse on the next GPT release, and it likely overstates Anthropic's position in $1M+ ACV accounts still on invoice contracts. Fair. The take still holds because the observability gap is real regardless of who leads the share number. ServiceNow's blown budget is not a share-of-wallet story. It is a category story — the winner of enterprise AI is selling consumer-grade plumbing at a $900B mark, and every large customer is discovering it on the same schedule.

The other thing that shipped this week

UK AISI confirmed Anthropic's Mythos cleared both simulated attack ranges — the first model to complete full network takeover autonomously. Mozilla's custom harness surfaced 271 Firefox bugs with the same model family that yielded one low-severity CVE against curl when run as a generic scanner. Same weights, 271:1 yield delta. The harness is the product, not the model. Meanwhile PraisonAI went from disclosure to weaponized exploit in four hours, LiteLLM landed on CISA KEV with active exploitation confirmed, and Traefik shipped a CVSS 10.0 auth bypass alongside an 18-year-old pre-auth RCE in NGINX's rewrite module.

The defensive stack was priced against an adversary that needed human researchers and months. That adversary is now an overnight prompt. Refusal-rate harnesses measure the wrong bottleneck. Any agent release gate not staged against a recon → lateral movement → persistence → exfil rubric is measuring what jailbreaks, not what compromises.

What to do this week

One action, deadline-driven, sized for a team that ships. Stand up a per-request cost and identity gateway in front of every LLM call by end of sprint. LiteLLM, Portkey, or a homegrown proxy — the specific choice is less important than the tags. Every call gets a team ID, feature ID, user ID, and request ID. Daily budget alerts fire per feature. Failover to a second provider on 429/5xx is behind the same abstraction.

That single piece of infrastructure solves four problems at once. It prices the June 15 change against real workload data instead of guesswork. It closes the ServiceNow-style blind spot where a compromised Claude account and legitimate use look identical. It makes multi-provider routing an operational question rather than a rewrite. And it gives you the trace-level data your eval harness needs to catch up with the fact that most of your tokens are now agentic.

If that gateway does not exist when the June 15 invoice lands, the number on it will be the least of the problems it reveals.