Claude Opus 4.1 After Retirement: Where It Still Runs, How to Leave

Model catalogues are slow to forget. Claude Opus 4.1 still appears in aggregator listings at $15 per million input tokens and $75 per million output, which makes it look like an expensive but available option. On Anthropic's own API it is neither: it was retired on 5 August 2026, and requests to it there fail.

This guide is for the two groups who still care about it: teams with Opus 4.1 calls buried in a pipeline, and anyone wondering why a retired model is still listed.

The status, platform by platform

Anthropic announced the deprecation on 5 June 2026 and retired the model on the Claude API on 5 August 2026. Its recommended replacement is Claude Opus 4.8.

Anthropic's retirement dates apply only to platforms it operates. Its pricing page describes Opus 4.1 as retired except on Amazon Bedrock and Google Cloud, which set their own schedules. Amazon's Bedrock model card lists it as a legacy model with an end-of-life date of 8 January 2027. Google Cloud's date for it is not something we have confirmed, so check the Vertex AI model page if you run it there.

That is why it can still appear in a router or aggregator catalogue: somebody is still able to serve it. If you are calling it through a third party, look at which upstream provider actually handles the request, because the answer determines how long you have left.

What Opus 4.1 was

Anthropic released Claude Opus 4.1 on 5 August 2025 as an upgrade to Claude Opus 4 on agentic tasks, real-world coding and reasoning. The headline figure was 74.5% on SWE-bench Verified. Customer reports at launch emphasised multi-file refactoring and making precise corrections in large codebases without collateral changes.

The specification was a 200K-token context window and 32K maximum output tokens, with text and image input and extended thinking via a manual token budget. It was available in Claude's apps, Claude Code, the Anthropic API, Bedrock and Vertex AI.

For its time it was the strongest coding model Anthropic sold. Within four months it had been overtaken on both capability and price.

The price gap that made migration easy to justify

Opus 4.1 kept Opus 4's pricing: $15 input and $75 output per million tokens, with batch at $7.50 and $37.50. The catalogue's :batch variant is that half-price asynchronous endpoint for the same model.

When Claude Opus 4.5 arrived in November 2025 it was priced at $5 and $25, a third of the cost. Every Opus since, including Opus 4.8, has stayed at $5 and $25 or lower. So the usual migration argument, that the new model is better but costs more, ran the other way here: the newer models were better and cheaper. Anyone still paying Opus 4.1 rates on a partner cloud is paying three times the current Opus price for an older model.

The per-token comparison slightly flatters the newer models, because Opus 4.7 and later use a tokenizer that Anthropic says produces roughly 30% more tokens for the same text. Even after that adjustment the gap is large. The cost of model upgrades covers the less obvious costs, such as re-tuning prompts and re-running evaluations.

What changes in your requests when you move up

Moving from Opus 4.1 to Opus 4.8 skips several generations of API change. The ones most likely to break working code:

• Thinking configuration. Opus 4.1 used thinking with budget_tokens. From Opus 4.6 onwards Anthropic recommends adaptive thinking steered by effort, and on Opus 4.7 and later a token budget is not accepted.
• Sampling parameters. Non-default temperature, top_p or top_k values return a 400 on Claude 4.7 and later models. Pipelines that set temperature 0 for repeatability need another approach.
• Assistant prefill. Prefilling the final assistant turn returns an error on the 4.6 generation and later. Use structured outputs or explicit instructions.
• Token budgets and limits. The newer tokenizer changes every count you have measured, so recheck max_tokens values and cost estimates.
• Context and output. Opus 4.8 has a 1M-token context window, so code that chunked documents to fit 200K may no longer need to, and it is worth testing whether it should.

If you must keep it running until January

Some teams cannot move quickly: a regulated workflow with signed-off outputs, or a vendor product that hardcodes the model ID. If that is you, run it on a partner cloud that still serves it, record the provider's end-of-life date in your runbook, and set an alert well before it. Do not rely on an aggregator listing as evidence of availability; listings can lag the underlying provider.

Meanwhile, start the replacement evaluation now, because a forced migration in the last week before end of life is the expensive kind.

Migration checklist

1. Search your code and configuration for claude-opus-4-1 and the dated ID claude-opus-4-1-20250805, including environment variables and vendor settings.
2. Pick a target. Anthropic recommends Opus 4.8; newer Opus models are also candidates. Test at least two.
3. Replace thinking budgets with adaptive thinking and an explicit effort level, remove sampling parameters, and replace prefill.
4. Recount tokens on the new model and revise max_tokens and budgets.
5. Run your own task set through old and new models side by side. How to benchmark LLMs on your own work describes a 20-task version that takes an afternoon.
6. Cut over, then delete the old model ID from configuration so it cannot creep back.