Anthropic's June 15 Meter Reset Is the Real Story This Week

On May 12, Anthropic converted Claude subscriptions into dollar-matched API credits. On June 15, third-party tool usage — Cline, Zed, OpenCode, Cursor via Claude — moves to a separate credit bucket that bills at API rates on overflow. Opus 4.7's image cost tripled the same week. ServiceNow, one of the most disciplined enterprise software buyers on earth, already burned its full-year Claude budget by May. Its CDIO is now selling the workaround as a product.

That is the story. Everything else this week is context.

The pricing change is a margin recovery timed to an IPO

Anthropic hired a CFO. The likely IPO target is October. The 70-90% implicit subsidy that made Claude-via-harness the cheap option was always a growth-phase feature, and growth-phase features don't survive S-1 diligence. If your COGS assumed subscription-token arbitrage, your gross margin changed on Friday and most of your team hasn't noticed yet.

OpenAI answered within hours: two months of free Codex for enterprise switchers, offer closing July 13. Ramp puts business share at Anthropic 34.4%, OpenAI 32.3% — the first lead change, and it's contested enough that OpenAI will pay for your evaluation. Take the offer. Run it against your actual eval harness on matched prompts. Even if you don't switch, you get a real number for the next renewal conversation.

Yes, but — a fair counter: switching costs are non-zero for teams with Claude-specific prompt engineering or fine-tunes, and Vercel's production data shows heavy multi-model routing is already the norm anyway, which means the pricing change is a rounding error against portfolio consumption. The counter holds for teams that already treat models as commodities behind a gateway. It doesn't hold for anyone running Claude Code, Cline, or a single-vendor agent stack, and that's most of the shops I talk to.

The enterprise-plumbing gap is the second-order problem

Anthropic ships no per-user telemetry. No SLAs on latency or availability. No admin dashboard worth the name. When ServiceNow blew its budget, they couldn't diagnose which users or workloads drove the burn — because the data doesn't exist on the vendor side. National Life Group's CIO said it plainly: great for consumers, not great for companies.

Dario admitted the shape of the problem when he said they planned for 10x growth and got 80x. An 8x forecast miss is why Claude Code degraded through April, why rate limits tightened without changelogs, and why Anthropic is now leasing xAI's entire Colossus 1 cluster — 220,000+ GPUs — from a CEO who called them misanthropic and evil three months ago. When a competitor becomes your landlord, the trust boundary moved and your DPA probably didn't. GDPR Article 28 wants a word.

The operator move is the same regardless of which vendor you pick: gateway everything. LiteLLM or Portkey, tenant tagging, per-user token attribution, hard budget alerts. If you can't answer which team, which feature, which prompt family is burning credits, you are ServiceNow in April.

One caveat: LiteLLM landed on CISA's KEV catalog on May 8. First AI infrastructure component on the list. Versions 1.81.16–1.83.7 are actively exploited. If it's already your gateway, patch this week and rotate every provider key it touched. The gateway thesis stands; the specific implementation needs supervision.

The agentic mix inverted while the tooling didn't

Vercel's AI Gateway index, across 200,000+ production teams, puts agentic workloads at 59% of token volume. Anthropic takes 61% of spend via Opus. Google takes 38% of volume via Flash. Input-to-output ratios shifted from roughly 3:1 to 15:1. If your cost model still uses last year's ratio it's off by about 5x in the direction that shows up as an invoice you didn't budget for.

Most eval harnesses still score single-turn responses against reference answers. That measures 41% of your traffic. What matters now is tool-call precision, steps-to-completion, cost-per-successful-task, and recovery-from-error rate. A planner that burns 40,000 tokens arguing with itself before succeeding looks identical to a clean three-step resolution under pass/fail — and costs an order of magnitude more.

Glean's benchmark says off-the-shelf MCP burns 30% more tokens than retrieval-tuned knowledge graphs on comparable tasks. It's vendor-published, so discount it, but the direction matches what verbose tool outputs do to context windows. Measure your own traces before you argue with the number.

The security floor moved under the same feet

Three perimeter auth bypasses landed together: an 18-year-old unauthenticated RCE in NGINX's rewrite module, Traefik CVE-2026-35051 at CVSS 10.0, and MOVEit CVE-2026-4670 at 9.8. Argo CD leaks plaintext Kubernetes secrets at 9.6. The chain writes itself: Traefik bypass → internal service → Argo CD → cluster.

PraisonAI was weaponized four hours after disclosure. Mythos cleared both UK AISI attack ranges — first model to do it. Microsoft's MDASH shipped 16 real Windows fixes in one Patch Tuesday cycle. A patch SLA measured in months is now measured against exploit windows measured in hours. That gap is where breaches live.

What to do this week

One action, not five. Deploy a gateway with per-user, per-feature token attribution before June 15. That single move gives you cost visibility Anthropic won't ship, model portability the OpenAI promo lets you test for free, agentic-workload instrumentation your eval harness is missing, and a place to enforce budget caps before the invoice teaches you what your team was actually doing. If your gateway is LiteLLM, patch it first. Then instrument. Then decide.

The teams that own the meter set the terms. The teams that don't get their terms set for them, on a schedule Anthropic's CFO is picking.