Short answer: On July 24, 2026 Anthropic removes fast mode for Claude Opus 4.7 — a request to claude-opus-4-7 with speed: "fast" returns an error, it does not quietly drop to standard speed. The fix is one line: change the model id to claude-opus-4-8, keep speed: "fast", and re-run your evals. While you're in the console, four other recent platform changes are already moving your bill.

If any part of your stack still calls Claude Opus 4.7 in fast mode, it stops working tomorrow. This isn't a deprecation warning or a soft downgrade — after the July 24 removal, claude-opus-4-7 with speed: "fast" returns an error (release notes). The model stays available at standard speed; only the fast-mode path breaks.

Why this one is stricter than the last#

Anthropic removed fast mode for Opus 4.6 three weeks earlier, on June 29 — but that change was gentle. Requests to claude-opus-4-6 with speed: "fast" kept running: standard speed, standard price, no error, with usage.speed reporting what actually happened. You could miss it and nothing would break.

Opus 4.7 gets the opposite treatment. There is no fallback, so a request in fast mode errors out (fast mode docs). That's arguably the safer design — a hard failure surfaces in your logs immediately instead of silently changing latency and cost under you — but it means you can't sleep through this one.

The one-line fix#

Change the model id and keep the rest of the request identical:

- "model": "claude-opus-4-7",
+ "model": "claude-opus-4-8",
  "speed": "fast",
  "messages": [ ... ]

Fast mode for Opus 4.8 ships as a research preview on the Claude API, including Claude Managed Agents. Two footnotes worth knowing before you flip the switch:

While you're in there: four changes already on your invoice#

The fast-mode removal is the one with a deadline, but it's not the only July change touching cost and reliability. Four quieter ones landed in the same release-notes stream:

  1. Refused requests are no longer billed (June 2). When Claude returns stop_reason: "refusal" without generating output, you aren't charged for it. If you run heavy safety guardrails, this is a small automatic discount you didn't have to ask for.
  2. Rate limits went up, tiers got simpler (June 26). Sonnet and Haiku limits now match Opus at every usage tier, and the tiers consolidated to three — Start, Build, Scale. Most organizations moved up; none moved down; no action required. Check your tier under console → limits before you architect around a limit that no longer binds.
  3. API keys can expire now (July 8). You can set an expiration when you create an API or Admin API key — a preset, a custom duration, or Never — and Anthropic emails the creator before a key with a 7-day-plus lifetime expires. Existing keys are untouched, but this is the moment to decide whether your long-lived production keys should be immortal.
  4. The legacy Workbench sunsets August 17 (announced July 17). The old Workbench and the experimental prompt-tools endpoints (generate_prompt, improve_prompt, templatize_prompt) are being retired; after removal those endpoints return an error. If you have saved prompts, variables, or evals in there, export them now.

The founder read#

Two of these have hard dates on the calendar — July 24 (fast mode) and August 17 (Workbench). The rest are automatic and in your favor. The pattern underneath is worth noticing: the platform is repricing downward while adding the guardrails — key expiration, no-charge refusals, consolidated tiers — that make an always-on agent cheaper and safer to run. The only thing that bites is the model you forgot you were still pinning. Grep your codebase for claude-opus-4-7 today; ship the one-line change before tomorrow.