benchrTwo dates, because two models
Anthropic split this retirement across the summer. Opus 4, announced deprecated on April 14, 2026, retired first on June 15, the same day as Claude Sonnet 4. Opus 4.1 got its notice on June 5, 2026 and retires August 5, one year to the day after it launched. If you're on 4.1 you still have until August 5. If you're on 4, it's already gone — calls now fail.
There's no reason to migrate from one doomed model to the other. Both point at the same replacement, and the replacement is better than either by a margin that isn't close.
The money, before and after
| Model | Input | Output | SWE-bench Verified | Status |
|---|---|---|---|---|
| Claude Opus 4 | $15.00 | $75.00 | 72.5% | Retired Jun 15, 2026 |
| Claude Opus 4.1 | $15.00 | $75.00 | 74.5% | Retires Aug 5, 2026 |
| Claude Opus 4.8 | $5.00 | $25.00 | 88.6% | Active (floor date May 28, 2027) |
The output column is where budgets change. Agent loops and long generations on old Opus burned $75 per million tokens out. On 4.8 the same tokens cost $25, and the model finishes harder tasks in fewer turns. Teams that held Opus usage to a minimum because of the price should re-run the math; the tier got three times more affordable while most people weren't looking.
Anthropic publishes a retirement floor for the replacement, too: Opus 4.8 won't retire before May 28, 2027. You're not jumping onto another sinking ship.
The parameter that breaks
Here's the migration surprise nobody reads the docs for. Anthropic deprecated the temperature, top_p, and top_k request parameters on Opus 4.7 and later. Send a non-default value to claude-opus-4-8 and the API returns a 400 error — your request doesn't degrade gracefully, it fails.
Old Opus integrations set temperature constantly, because 2025-era prompt guides told everyone to. Before you flip the model ID, strip those parameters from your request builder and move any determinism requirements into the prompt itself. This one line of cleanup prevents the most common post-migration outage.
Migration order of operations
Swap the model string, delete the sampling parameters, re-run your eval suite, then watch the first week of production traffic. Anthropic's console exports usage by API key and model; use it to confirm nothing still calls claude-opus-4-20250514 or claude-opus-4-1-20250805 after the cutover. Bedrock and Vertex AI deployments follow separate retirement calendars; check their model tables directly. And since the per-token price just fell by two-thirds, it's a good week to re-run your monthly numbers through the cost calculator and see what you can afford to move up-tier.