At 3:19 PM UTC on June 15, Opus 4.8 started returning 529 Overloaded errors. The incident affected the API, Claude Code, Claude AI, and Claude Cowork. It lasted roughly two hours.
At 4:00 PM UTC — forty-one minutes later — Claude Opus 4 and Sonnet 4 retired permanently. Every API request to the old model IDs started returning errors. No fallback. No grace period.
The developers who had migrated from Opus 4 to Opus 4.8 — as Anthropic instructed — hit server errors on the replacement model at the moment the original died. The migration path was disrupted at the instant it became mandatory.
That same day, the subscription split took effect. Automated agent usage — the Agent SDK, claude -p, GitHub Actions, third-party tools — moved to a separate credit pool at full API rates. Pro users got $20 in credits. Max 20x got $200. The flat-rate era for programmatic access ended.
June 15 carried triple weight: credit split, model retirement, and the replacement model breaking. It was also the middle of the Fable 5 free window (June 9–22) — the most capable coding model Anthropic has ever shipped, available on subscriptions for exactly thirteen days before requiring usage credits on June 23.
None of this was a cliff. It was a step.
The Calendar
| Date | Event | Effect |
|---|---|---|
| June 1 | Copilot token billing goes live | Reports of $29→$750, $50→$3,000 first-week bills |
| June 8–11 | FinOps X, San Diego | ~25% of sessions on AI token economics. Tokenomics Foundation launches under Linux Foundation |
| June 9–22 | Fable 5 free on subscriptions | Mythos-class model ($10/$50 per M tokens) available at no extra cost. Counts 2× against limits. Thirteen days. |
| June 15 | Claude credit split + Opus 4/Sonnet 4 retirement + Opus 4.8 outage | Triple event. Automated usage repriced to API rates. Old models die. Replacement model breaks. |
| June 18 | Gemini CLI dies | Closed-source Antigravity CLI replaces it. Enterprise keeps access. Free/individual users forced to migrate. No grace period. |
| June 22 | Fable 5 free window ends | Sugar rush over |
| June 23 | Fable 5 requires usage credits | Most capable model ever → meter |
| Aug 5 | Opus 4.1 retires | Another migration deadline. The pipeline continues. |
| Sept | Copilot promotional credits expire | Enterprise bills arrive at full rates. The promotions that masked June 1 wear off. |
Nine steps. Twenty-two days for the core cluster (June 1–23). Three vendors. The press is now reporting the synchronized timing as a coordinated pattern: "Three of the most-used assistants restructured how teams pay within days of each other in June 2026."
What the Subsidy Was
The flat-rate subscription model for AI coding tools was never sustainable. It was a growth mechanism. MagnaCapax calculated the effective subsidy range: 12× at the floor (a Pro user routing $236/month in API value through a $20 subscription), 175× at the ceiling (a Max 20x heavy Sonnet user extracting ~$35,000/month in API value from $200 paid). Boris Cherny, head of Claude Code, said it directly: "Our subscriptions weren't built for the usage patterns of these third-party tools."
The subsidy worked. It built the market. Claude Code is now responsible for roughly 10% of all public GitHub commits. Copilot has 4.7 million paying users. Cursor hit $3 billion ARR. Anthropic's own revenue surged from $19 billion to $47 billion in three months. The subsidy era produced one of the fastest adoption curves in software history.
And now the subsidy is the thing being removed.
What Drove the Removal
The demand side matters as much as the supply side. Companies didn't just adopt AI coding tools — they gamified consumption.
"If a top engineer spends $500K on tokens and gets 10x, it's a no-brainer. No cap."
— Andrew Bosworth, Meta CTO
Meta built Claudeonomics — an internal dashboard ranking 85,000 employees by token usage. Sixty trillion tokens in thirty days. Gamified titles: "Token Legend," "Session Immortal," "Cache Wizard." Employees ran agents idle to inflate their rankings. The dashboard was shut down 48 hours after the data leaked externally. Jensen Huang said he'd be "deeply alarmed" if a $500K engineer wasn't burning $250K in tokens per year.
Uber built its own leaderboard. Adoption hit 95%. The entire 2026 AI budget was burned in four months. The CTO spent $1,200 in a two-hour demo. The COO coined "tokenmaxxing." The response was a $1,500/month cap — a smaller subsidy with a shorter horizon.
Microsoft's EVP Rajesh Jha issued a memo canceling Claude Code licenses by June 30. Walmart capped Code Puppy tokens. An unnamed company reported a $500 million Claude bill in a single month.
The pattern: encourage → overflow → ration → split. The subsidy didn't fail organically. It was overwhelmed by demand amplification — companies incentivizing exactly the consumption patterns that made flat-rate pricing unsustainable.
The Vendor's Dilemma
Anthropic tried four times to fix the economics. Each attempt was more structured than the last.
January: Blocked subscription OAuth for third-party tools. Reversed within days after backlash.
April (pricing): A/B tested removing Claude Code from Pro for 2% of new signups. Reversed within 24 hours.
June 9 (safety): Shipped invisible distillation classifiers in Fable 5 that silently degraded responses for AI development work. Called "secret sabotage" by researchers. Walked back to visible fallbacks.
June 15: Subscription split with separate credit pool. Adds something ($20–200 credits) rather than taking something away. The most carefully designed attempt because each previous failure taught the design.
Cursor's CTO at SaaStr said it plainly: "Our users love per-seat pricing — it's just the cost side makes it harder." The market wants flat-rate. The economics can't support it. The gap is the staircase.
The Numbers Under the Step
Token prices fell 98% — from roughly $20 to $0.40 per million tokens over two years. Enterprise AI bills rose 320%. Average enterprise AI budget went from $1.2 million (2024) to $7 million (2026). Per-developer token consumption increased 18.6× in nine months. "Companies started calling in April saying they were already 3× over their entire 2026 token budget" — TechCrunch.
The paradox is the point. Jevons' Paradox for AI tokens: efficiency gains increase total consumption. Per-token cost drops make agentic workflows viable, and agentic workflows consume 10–15× more tokens than chat. A single Fable 5 session can cost $2–$8. Daily heavy usage reaches $40–160 before subscription costs. One developer reported a five-hour usage window burned through in a refactor.
The cost crisis became institutional enough to warrant its own standards body. The Tokenomics Foundation — backed by Accenture, Google Cloud, IBM, JPMorgan Chase, Microsoft, Oracle, Salesforce, and SAP — launched at FinOps X with new metrics: cost-per-intelligence, tokens-per-watt. A dedicated spinoff conference (Tokenomicon) is scheduled for Amsterdam in September and London in February 2027. The cost of AI coding now has its own conference circuit.
The Escape Valve
As commercial tools raise prices, open-source absorbs price-sensitive developers at exactly the moment it matters.
OpenCode went from zero to 8 million monthly active users in one year, without marketing. 172,000 GitHub stars. $25 million ARR. Provider-agnostic — supports 75+ models. It topped LogRocket's power rankings in June, displacing Cursor. The open-source escape valve accelerates precisely as commercial tools add meters.
Zed Industries created the Agent Client Protocol (ACP) — the LSP for AI agents — specifically so users can switch providers when pricing shifts. Vendor lock-in protection as a protocol. Twenty-five editors and agents have adopted it. The infrastructure to leave is being built in parallel with the reasons to leave.
What the Staircase Means
The subsidy era produced the market. Anthropic's revenue IS the enterprise cost — same number, opposite meaning. A $47 billion run-rate for Anthropic is $47 billion flowing out of enterprise budgets. The growth story and the cost crisis are the same event viewed from different ledgers.
Each step on the staircase creates a smaller platform with a shorter horizon. Uber's $1,500 cap. Copilot's promotional credits masking through August. Anthropic's $20–200 credit pool. Fable 5's thirteen-day free window. Each is a subsidy — just smaller, and with an expiration date visible from the moment it's offered.
Gemini CLI dies in two days. Fable 5's free window closes in six. Copilot's promotions expire in August. Enterprise bills arrive in September. Opus 4.1 retires in August, adding another migration on top of another repricing.
The staircase doesn't stop. The next step is always visible. And each step is shorter than the last.