Anthropic's June 15 pricing reset just broke your AI unit economics

ServiceNow — a $150B enterprise software company that ships governance tooling for a living — publicly conceded this week that it burned through its full-year Anthropic budget by May. National Life Group's CIO called Claude "great for consumer usage but not great for companies." Anthropic responded with silence, no SLAs, and no per-user telemetry. Three weeks from now, on June 15, the same vendor is closing the arbitrage that has been quietly subsidizing most Claude-dependent stacks: every subscription converts to dollar-matched API credits, and third-party tools (Zed, Cline, OpenCode, Conductor, the GitHub Actions harness you didn't tell finance about) get a separate credit pool that overflows at list rates.

The effective cost delta is 3-10x on the same workload. Same prompts. Same outputs. New bill.

OpenAI shipped a two-month free Codex offer for enterprise switchers the same day, with a July 13 deadline. Ramp's April data has Anthropic at 34.4% of business AI spend versus OpenAI at 32.3% — the first lead change, being contested with displacement pricing. Read it as a subsidy war both sides need to be seen losing money on for a few quarters, timed to Anthropic's October IPO window and OpenAI's need to arrest a defection cohort.

Yes, but — a fair reading is that Anthropic is just charging what the tokens cost, and the 70-90% "discount" was always a fiction subsidized by capacity Anthropic didn't have. That's technically right and operationally irrelevant. The fiction was priced into hundreds of production stacks. Removing it in three weeks is a repricing event whether or not it was ever a sustainable price.

Why the number under the number is worse

The token bill is the visible part. The invisible part is the forward-deployed engineer tax that every serious vendor now quietly acknowledges is required to actually ship. Google is hiring hundreds of FDEs. OpenAI stood up DeployCo with Bain. Anthropic, Salesforce, and ServiceNow are staffing the same function under different names. At $300-500K loaded per FDE and five to ten required per deployment, true program cost runs 3-5x model fees. Budgets built on token pricing are approving roughly a third of the actual spend.

ServiceNow's failure was not overspend. It was instrumentation. The company that couldn't tell which teams drove the burn is now building AI Control Tower and selling it to everyone else in the same position. That's the market routing around a vendor deficiency in real time, and it's the clearest tell for where the next Datadog-scale category is forming.

Meanwhile, Vercel's AI Gateway index — 200,000+ teams, seven months of production traffic — puts agentic workloads at 59% of all token volume. Anthropic takes 61% of spend on Opus reasoning. Google takes 38% of volume on Flash throughput. If your eval harness is single-turn and your cost model runs on 3:1 input-output ratios, you are measuring the minority of your traffic and miscosting the majority by roughly 5x. Agentic traces run closer to 15:1, input-heavy, with variable cache-hit rates across providers. A five-hop MCP session without prompt deduplication wastes 30% of its tokens re-shipping context the previous hop already paid for.

What to do this week

Three moves, in order.

First, audit every Claude-backed workload against the new credit math before Monday. The formula is: current third-party token usage minus plan credit equivalent, times list API rates. If the answer is uncomfortable, you have two weeks to either renegotiate directly with Anthropic (their pre-IPO posture is the most leverage you'll get for a year), pilot the free Codex window on one load-bearing workflow, or accept the new bill with your eyes open. Do the arithmetic before the pricing changes; "we'll figure it out in June" is how ServiceNow's budget went.

Second, ship per-user, per-feature token attribution into your gateway this sprint. LiteLLM, Portkey, or an internal shim — the tool doesn't matter, the meter does. Anthropic explicitly offloaded observability onto customers. Absent your own instrumentation, you are running a black-box P&L, and the failure mode is discovering it in a monthly invoice rather than a daily dashboard. Add trajectory-level metrics while you're there: tool-call precision, steps-to-completion, cost-per-successful-task. Pass/fail accuracy hides the 40,000-token planner spiraling into itself before giving up.

Third, decide whether your product's core workflows can be invoked by an agent that isn't yours. SAP put €100M behind Knowledge Graph. ServiceNow shipped Action Fabric on MCP. Notion, Airtable, and Salesforce are running the same play. Procurement managers are already asking the question in RFPs: can our agents call this directly, or does a human have to click through your UI? Two to four weeks of scoping, one quarter of build, assuming your API isn't a mess. The window before this becomes a renewal blocker is 2-3 quarters. After that, the agent living in the buyer's stack routes around you to reach the system of record, and your product becomes a reporting surface on execution someone else owns.

The compounding effect of the three moves is that you know what you spend, you can prove what you get, and you're callable by the agents that will be doing the calling. Miss any one of them and you'll spend Q3 explaining an invoice to a CFO who has read the ServiceNow story.