For a new production API deployment, neither is a sound default today—but their lifecycle status is different. As of November 12, 2026, Anthropic lists claude-opus-4-1-20250805 as retired on Anthropic-operated platforms; its scheduled retirement was August 5, 2026. Requests to retired models fail. OpenAI labels gpt-5-2025-08-07 deprecated, but deprecation alone does not mean shutdown. OpenAI’s public GPT-5 model page does not establish whether this model is callable for your account or give its shutdown date. Check the model’s shutdown_date through OpenAI’s model API before deciding what to do with an existing deployment. (platform.claude.com)
For a new deployment, avoid building against these older model versions: Opus 4.1 is retired on Anthropic-operated services, and GPT-5 is deprecated. For an existing GPT-5 deployment, don’t assume either that calls still work or that they have stopped; check the model metadata and your account’s access. For an existing Opus 4.1 deployment on the Anthropic API, plan for calls to fail after retirement. Partner-operated platforms such as Amazon Bedrock and Google-managed Claude can have different lifecycle schedules. (developers.openai.com)
If you’re comparing the models for historical research, an existing benchmark, or a migration decision, the available material does not establish a universal performance winner. It offers task ideas and vendor-reported claims, but no fair, reproducible head-to-head result. For a current purchase, compare models that are active and available through your intended API channel instead.
At a glance
| Question | Claude Opus 4.1 | GPT-5 |
|---|---|---|
| Production API status | Anthropic lists claude-opus-4-1-20250805 as retired; the scheduled Anthropic-operated API retirement date was August 5, 2026. Partner platforms set their own schedules. |
OpenAI labels gpt-5-2025-08-07 deprecated. The public model page does not establish account-specific callability or show a shutdown date; check the model API’s shutdown_date. |
| Historical API model ID | claude-opus-4-1-20250805 |
gpt-5-2025-08-07 |
| What the sources say about strengths | Anthropic described it as an upgrade for agentic tasks, real-world coding, and reasoning. A third-party article also characterized it as suited to long-context, multi-step work, but provided no comparative test evidence. | OpenAI describes it as a previous reasoning model for coding and agentic tasks. The same third-party article suggested strengths in front-end polish and developer experience, without naming tests or measurements. |
| Is there enough evidence to declare a winner? | No. | No. |
The exact IDs are useful for recognizing old deployments, not a recommendation to build against them. Anthropic’s retirement status applies to its operated platforms; Amazon Bedrock and Google’s provider-operated Claude service may follow different schedules. OpenAI says to check the optional shutdown_date field on the model object to monitor shutdowns; a null value means no shutdown date has been announced. (platform.claude.com)
What the comparison evidence actually shows
Anthropic announced Opus 4.1 on August 5, 2025, describing it as an upgrade to Opus 4 for agentic tasks, real-world coding, and reasoning. The announcement reported 74.5% on SWE-bench Verified for Opus 4.1. That is a vendor-reported result, not a head-to-head comparison with GPT-5. (slagbot.ai)
The supplied third-party comparison article offers a possible way to frame coding work: Opus 4.1 for long-context, multi-step agentic workflows, and GPT-5 for front-end polish and developer-facing work. Treat those as hypotheses to test, not established advantages: the captured article gives no benchmark names, paired scores, methodology, or model outputs. Its example tasks—single-file generation and multi-file refactoring—and measures—tests passed, runtime correctness, and review effort—are useful evaluation prompts, not evidence that one model wins them.
In short, separate vendor benchmark claims or broad descriptions cannot answer “which is better?” unless the models were tested on comparable tasks, under disclosed conditions, with the same scoring method. The available evidence here does not supply that comparison.
API and historical pricing: don’t use old rates as a current quote
The historical API identifiers were claude-opus-4-1-20250805 and gpt-5-2025-08-07. OpenAI’s GPT-5 model page labels the model deprecated and displays API rates of $1.25 per million input tokens and $10 per million output tokens. Anthropic said Opus 4.1 pricing matched Opus 4; Anthropic’s Opus 4 launch announcement listed $15 per million input tokens and $75 per million output tokens. Treat these as listed rates for the named model versions, not a guarantee of current access or a reliable estimate of what a production deployment can use or cost. Don’t compare them as live offers or confuse API token charges with consumer subscriptions. (developers.openai.com)
For a current buying decision, verify the exact model ID, lifecycle status, platform, billing unit, and applicable rates with the provider or cloud platform you intend to use. That distinction matters because Anthropic says its retirement dates apply to Anthropic-operated platforms; Amazon Bedrock and Google’s provider-operated Claude service set their own schedules. (platform.claude.com)
How to choose a production replacement
Rather than infer a winner from the old model names, run an evaluation on the workload you actually need to ship:
- Select representative tasks. Include examples such as a single-file change, a multi-file refactor, or a tool-using workflow only if those resemble your production jobs.
- Set shared conditions. Use the same task inputs, repository or context, tool access, stopping rules, and acceptance criteria for each candidate. Record the exact model IDs and API platform.
- Measure outcomes that matter. Track tests passed, runtime behavior, defects introduced, human review or correction time, and latency. For cost, record actual input and output token use and apply the current rates for that API channel.
- Repeat and inspect failures. A small set of runs can reveal issues worth investigating, but don’t generalize beyond the tasks and conditions tested.
- Check lifecycle and migration impact. Confirm that the chosen model is active and supported on your deployment platform, then regression-test it before switching production traffic.
Bottom line: Opus 4.1 and GPT-5 can still be compared as historical model versions, but they do not have the same API status. Anthropic lists Opus 4.1 as retired on its operated platforms; OpenAI labels GPT-5 deprecated, while account-specific callability and its shutdown date need to be checked through the model API. There is no supported basis here for declaring one the better model overall—or for treating historical pricing as a current quote. Choose an active, available model only after testing it against your own workload. (platform.claude.com)
Sources
Model deprecations | Claude Platform Docs
モデルの非推奨化 | Claude Platform Docs