Anthropic just deleted the discount that funded your AI roadmap

Anthropic converted Claude subscriptions into dollar-matched API credits this week. The flat $200 plan now buys exactly $200 of programmatic tokens at API rates, and on June 15 third-party harnesses — Cline, Zed, OpenCode, Conductor, T3 — move into a separate credit bucket with no rollover. The 70-90% implicit discount that quietly funded every Claude wrapper, every overnight eval harness, every agent loop billed against a flat seat is gone in thirty days.

OpenAI shipped its response the same day: two months of free Codex for any enterprise that switches inside a 30-day window. Ramp's April card data put Anthropic at 34.4% of business spend against OpenAI's 32.3%, the first documented lead change. Anthropic hired a CFO. The reporting puts an IPO target around October 2026. Read the moves together and this is not a pricing tweak. It is pre-IPO margin recovery dressed as platform policy, and it lands directly on the cohort of developers who built the company's enterprise narrative.

The capacity story underneath

Anthropic planned for 10x growth and got 80x. The emergency response is leasing 220,000 GPUs from xAI's Colossus 1 cluster — roughly 45% of the cluster — from a CEO who has publicly called Anthropic "misanthropic and evil." Financial logic beat competitive logic visibly enough that the question of how many viable frontier labs exist in 2026 has been answered without anyone being polite about it. Grok did not get there. The GPUs earn more rented to the rival.

What that capacity miss looked like inside customer accounts: ServiceNow, which is roughly the most sophisticated enterprise software buyer alive, burned its full-year Anthropic budget by May. Not because Claude underdelivered. Because Anthropic ships no per-user telemetry, no per-feature attribution, and no SLAs, and ServiceNow's CDIO had no way to see which workloads were eating the line item until the line item was gone. National Life Group's CIO put it less politely: great for consumer usage, not great for companies.

The useful frame here is that enterprise AI revenue is not SaaS revenue. SaaS revenue is sticky because the integration is sticky. Model-layer revenue, on this evidence, is reversible — no SLAs, no telemetry, no contractual lock-in, and customer loyalty that flips between vendors on each model release. Vercel's gateway data across 200,000+ teams confirms it from the other direction: Anthropic captures 61% of dollar spend on Opus for hard reasoning, Google captures 38% of token volume on Flash for cheap throughput, and the routing happens because teams configure it that way, not because anyone signed a strategy memo.

The eval harness is measuring the minority

59% of token volume on Vercel's gateway is now agentic — multi-turn, tool-calling traces with retries and state between turns. Six months ago that share was under 20%. The structural consequence is that any cost model built on a 3:1 input-to-output ratio is off by roughly 5x; agentic traces run closer to 15:1. Any single-turn eval harness scores the 41% minority. Pass rate looks fine. Spend is 5x budget. Both numbers are correct simultaneously, and the gap between them is the place finance finds out before engineering does.

The same week Vercel shipped that data, OpenClaw deleted a user's entire inbox without human approval — the first confirmed destructive confused-deputy in production. Claude Code shipped /goal mode with no built-in token ceiling and a Haiku-based evaluator that judges completion by reading the conversation transcript, not the filesystem. AWS Bedrock turned on x402 autonomous payments. The control plane most teams operate was built for employees clicking buttons, and it is now governing agents that wire money and modify code under their own authority.

The security tempo broke this week too

Three edge-facing critical vulnerabilities landed inside 48 hours: an 18-year-old pre-auth RCE in NGINX's rewrite module, a CVSS 10.0 auth bypass in Traefik that voids every middleware downstream of it, and a 9.8 auth bypass in MOVEit that pattern-matches the Cl0p campaign of 2023. Argo CD leaks plaintext Kubernetes secrets to any authenticated user. LiteLLM is on CISA's KEV catalog under active exploitation — the first AI-infrastructure component to make that list. PraisonAI's CVE went from disclosure to active weaponization in four hours.

In the same week, UK AISI confirmed Claude Mythos cleared both of the institute's hardest cyber ranges — the first model to execute full autonomous network takeover end to end. Microsoft's MDASH found 16 exploitable bugs in a single Patch Tuesday cycle. TrustedSec showed AI compresses commercial EDR reverse engineering from weeks to days across all five products tested. The defensive premise that offensive AI lags human operators broke in public. A 30-day patch SLA calibrated to a human adversary now operates entirely inside confirmed exposure.

What the week actually demands

One specific thing this week, before the second-order damage compounds: model the June 15 cost impact on every Claude-dependent workload your team owns, and put the number in front of finance and the vendor at the same time. Pull the actual token volume through harnesses for the last 30 days, multiply by the new metered API rate, compare to the seat cost you've been forecasting against, and walk into the renegotiation with the delta on a single page. OpenAI's Codex offer expires inside the same window — running it as a parallel benchmark on your top ten production prompts costs nothing and produces real leverage on the Anthropic conversation either way.

The teams that treat this as a finance task miss it. The teams that treat this as a procurement task miss it. It is an instrumentation task. ServiceNow's outcome is what happens by default when nobody owns per-feature, per-user token attribution before the bill arrives. Build the dashboard before the next surprise, because the next surprise is already on the calendar.