Claude Opus 4 (retired)
Claude Opus 4 was retired from the Claude API on 15 June 2026. Its original specifications and benchmarks, why it mattered, and exactly what to migrate to.
Claude Opus 4 (claude-opus-4-20250514) was retired from the Claude API on 15 June 2026, having been deprecated on 14 April 2026. Requests using that model ID now return an error. The current replacement is claude-opus-5 - see the Claude Opus 5 page for its specification and pricing.
Why this page still exists: Opus 4 was Anthropic's flagship for most of 2025 and is referenced in a great deal of documentation, tutorials and vendor paperwork written in that period. Knowing what it was, and what to put in its place, is more useful than a dead link.
Claude Opus 4 was released on 22 May 2025 as the top tier of the Claude 4 generation. At launch it was Anthropic's most capable model, aimed at long-running coding work and multi-step agentic tasks, and it was the first Claude model deployed under the company's ASL-3 safety standard. It ran as flagship for about eleven weeks before Opus 4.1 arrived.
#Claude Opus 4 as it shipped
- Status
- Retired 15 June 2026
- Released
- 22 May 2025
- Retired API ID
claude-opus-4-20250514- Context window
- 200,000 tokens
- Max output
- 32,000 tokens
- Price at launch
- $15 input / $75 output per million tokens
- Batch price
- $7.50 / $37.50 per million tokens
- Safety tier
- ASL-3 deployment and security standard
- Replace with
claude-opus-5
#What was Claude Opus 4 built for?
At launch in May 2025, Anthropic positioned Opus 4 as the model to reach for when a task ran long enough that giving up halfway was the real failure mode: multi-file refactors, sustained debugging sessions, research that needed dozens of tool calls. That framing has not changed in the year since - it is exactly how Claude Fable 5 is described today - but the definition of "long" has moved from tens of minutes to many hours.
Three capabilities defined it in mid-2025:
- Hybrid reasoning. Opus 4 could answer immediately or engage an explicit extended-thinking mode with a token budget you set. That manual budget has since been replaced by adaptive thinking and an
effortladder; the difference is explained on the thinking and reasoning page. - Tool use during reasoning. It could interleave tool calls with thinking steps rather than finishing its reasoning first, which is what made agent loops practical rather than merely possible.
- Local memory files. Given file access, Opus 4 could write and read its own notes across a session - the crude ancestor of the memory tool and context-management features in the current API.
It also launched alongside the general availability of Claude Code, the terminal agent covered on Claude for developers. Opus 4 and Claude Code were sold as a pair, and that pairing is most of why agentic coding became the category Anthropic is known for.
#Claude Opus 4 benchmarks at launch (May 2025)
These are Anthropic's own published figures from 22 May 2025. Treat them strictly as a historical snapshot. SWE-bench Verified and Terminal-bench in this form are no longer reported by frontier vendors, so these numbers cannot be lined up against any 2026 model score.
| Benchmark | Claude Opus 4 score | Published | For comparison |
|---|---|---|---|
| SWE-bench Verified | 72.5% | 22 May 2025 | Sonnet 4 scored 72.7% the same day |
| Terminal-bench | 43.2% | 22 May 2025 | - |
| SWE-bench Verified - successor models | - | - | Opus 4.1 74.5% (Aug 2025); Sonnet 4.5 77.2% (Sep 2025); Haiku 4.5 73.3% (Oct 2025) |
The awkward fact in that table is that Sonnet 4 matched Opus 4 on the headline coding benchmark while costing a fifth as much. Anthropic's argument at the time was that Opus 4's advantage appeared over hours of work rather than in a single scored patch - a claim benchmarks of that era were poorly designed to test, and one reason the industry has since shifted to long-horizon agentic evaluations. By October 2025 the cheapest model in the lineup had passed Opus 4's launch score outright.
#Migrating away from Claude Opus 4
If an application still carries an Opus 4 model string, it is failing today rather than merely returning dated output. Use this mapping.
| Old model string | Platform | Recommended replacement |
|---|---|---|
claude-opus-4-20250514 | Claude API | claude-opus-5 |
anthropic.claude-opus-4-20250514-v1:0 | AWS Bedrock | anthropic.claude-opus-5 |
claude-opus-4@20250514 | Google Cloud / Vertex AI | claude-opus-5 |
claude-opus-4-20250514 | Anthropic's own deprecation notice | claude-opus-4-8 (like-for-like, still callable) |
Anthropic's deprecation notice named claude-opus-4-8 as the direct successor, and that remains a defensible choice if you want the smallest possible behavioural change. For new work, claude-opus-5 is the better target: it costs $5 / $25 per million tokens against Opus 4's $15 / $75, carries a 1M-token context instead of 200k, and has a May 2026 knowledge cutoff. Current rates for every tier are on the pricing page.
Three API behaviours changed after Opus 4 and will break a copy-paste migration. temperature, top_p and top_k return a 400 error on Opus 4.7 and later. Assistant message prefill was removed on 4.6 and later. And thinking.budget_tokens has given way to adaptive thinking plus effort - on Opus 5, disabling thinking while effort is xhigh or max also returns a 400. The working details are on the Claude API page.
#One nuance: Opus 4 outlived its own retirement on one platform
Anthropic's pricing documentation still lists Claude Opus 4 with a footnote marking it retired except on Google Cloud. Partner clouds run their own deprecation calendars, so a model that is dead on the first-party API can linger on Bedrock or Vertex AI for a while. This is worth knowing if you are auditing an estate: "we still call Opus 4 in production" and "Opus 4 is retired" can both be true at once, on different platforms. It is not a reason to stay - it is a reason to find the workload before the partner cloud retires it too.
#What to use instead of Opus 4, by job
| What you used Opus 4 for | Use now | Why |
|---|---|---|
| Long agentic coding sessions | claude-opus-5 | Cheaper than Opus 4 was, 1M context, newest knowledge cutoff |
| Very long-horizon autonomous agents | claude-fable-5 | Built for multi-hour runs; costs $10 / $50 and is slower |
| Everyday production work at volume | claude-sonnet-5 | $2 / $10 and fast; the default most teams should pick |
| Minimal-change like-for-like swap | claude-opus-4-8 | Anthropic's named replacement; see the 4.x pages for the surrounding generation |
If you are re-picking a tier from scratch rather than swapping a string, the cost-per-task reasoning on the model comparison page is a better starting point than any benchmark table, and the model index lists every currently callable model with its specification. The parallel record for the other half of this generation is on the Claude Sonnet 4 archive.
#Frequently asked questions
Is Claude Opus 4 still available?
No. Opus 4 was deprecated on 14 April 2026 and retired from the Claude API on 15 June 2026. Calls to claude-opus-4-20250514 return an error. Anthropic's pricing documentation notes it remained listed on Google Cloud after that date; partner clouds retire models on their own schedules.
What should I replace Claude Opus 4 with?
Use claude-opus-5 for new work - it costs $5 / $25 per million tokens against Opus 4's $15 / $75 and has a 1M-token context. Anthropic's own deprecation notice named claude-opus-4-8 as the like-for-like successor if you want minimal behavioural change.
What did Claude Opus 4 score on SWE-bench Verified?
72.5%, published by Anthropic on 22 May 2025. Claude Sonnet 4 scored 72.7% the same day. These are historical figures for a retired model, and SWE-bench Verified is no longer reported by frontier vendors, so they cannot be compared with current scores.
How much did Claude Opus 4 cost?
$15 per million input tokens and $75 per million output tokens, or $7.50 and $37.50 through the Batch API. Every Opus-tier model released since November 2025 has cost $5 / $25, so the current flagship is a third of Opus 4's price.
Did Claude Opus 4 have a 1M-token context window?
No. Opus 4 had a 200,000-token context window and a 32,000-token maximum output. The 1M-token window first appeared in beta on Opus 4.6 in February 2026 and is now standard on Opus 5, Sonnet 5 and Fable 5 at no price premium.
Deprecation and retirement dates checked against Anthropic's model deprecations documentation at platform.claude.com/docs, and launch specifications and benchmarks against anthropic.com/news/claude-4, on 21 August 2026. This is an archive page: every capability figure on it describes the model as it stood in May 2025.