Release guide · July 24, 2026
Claude Opus 5 is here: what to test before you switch
Claude Opus 5 is still active. Check its status, compare Opus 5.5 migration requirements, and run a ten-task test before changing your agent.

Opus 5 remains active, but Anthropic now recommends starting most workloads with Opus 5.5. Keep a working Opus 5 baseline while checking migration compatibility, then compare accepted results, review time, and total cost before switching.
Claude Opus 5 launched on July 24, 2026, according to Anthropic’s release notes. The original Thursday rumor is no longer the basis for an adoption decision: there is an official model, price, context window, and API name to test.
Sources checked October 26, 2026. This is a documentation refresh. The exercises below are proposed evaluations; Superbash has not run this ten-task comparison or measured a migration success rate.
October update: keep Opus 5 as a baseline
Anthropic’s current model overview recommends starting most workloads with Opus 5.5. Opus 5 remains active, with retirement not sooner than July 24, 2027, according to the model status table. Its legacy label does not mean requests have stopped working, and that retirement commitment is not a scheduled shutdown date.
For an existing agent, the useful question is whether Opus 5.5 improves an already measured workflow without breaking the integration.
| Standard Claude API detail | Opus 5 | Opus 5.5 |
|---|---|---|
| Model ID | claude-opus-5 | claude-opus-5-5 |
| Released | July 24, 2026 | September 22, 2026 |
| Input / output per million tokens, USD | $5 / $25 | $4 / $20 |
| Context / standard maximum output | 1M / 128K tokens | 1M / 128K tokens |
| Default effort | high | medium |
| Thinking | On by default; can be disabled at supported effort levels | Always on |
Sources: Opus 5 specifications, Opus 5.5 specifications, and the July 24 release notes. These are standard token rates, not a quote for a complete agent job. Record cache usage, paid tools, retries, and the actual provider bill separately. Lower rates do not establish a lower total task cost.
Check compatibility before measuring quality
Anthropic’s Opus 5.5 migration guide identifies changes that a model-name swap can miss:
| Existing behavior | Check when evaluating Opus 5.5 |
|---|---|
| Sends disabled or manual-budget thinking | These settings return 400 errors. Omit thinking or use adaptive. |
Forces a tool with any or tool | These choices return 400 errors. Use auto or none; verify that required work actually happens. |
| Relies on default effort | Set effort explicitly: Opus 5 defaults to high, Opus 5.5 to medium. |
| Streams narration between tool calls | Progress moves into thinking blocks, empty by default. Review display settings and rendering. |
| Reuses a conversation for fallback | Check thinking-block compatibility before sending Opus 5.5 history to another model. |
Computer-use integrations also need the provider-specific section: the Claude API and Google-managed Google Cloud route reject computer_20251124 on Opus 5.5; Amazon Bedrock still accepts it. On the affected routes, review computer_toolset_20260801 before migrating.
Start with a disposable project and one known file-edit task. Save the request settings, returned model, tool trace, diff, and test exit code. Fix request-validation errors before scoring capability. If required tooling fails, keep production on the saved baseline while you repair and retest the integration.
What the July release established
Opus 5 introduced a 1M context window, 128K standard maximum output, thinking enabled by default, and five effort levels: low, medium, high, xhigh, and max. Disabling thinking at xhigh or max returns a 400 error. These are documented in the July 24 release notes.
At launch, Anthropic described fast mode as approximately 2.5 times the default speed at twice the base price. That is a vendor launch claim, not a speed measurement from this article. The October 15 release notes still list Opus 5 fast mode as a research preview requiring its beta header, even though Opus 5.5 fast mode has left preview. Keep speed mode separate in your evaluation log.
The features worth learning
Effort is now part of model routing
Do not run every task at max. A model that exposes several effort levels gives you another routing decision alongside model choice.
Use lower effort for work with a clear pattern and cheap verification. Raise it for a difficult bug, an ambiguous research problem, or a long task where a weak first plan creates expensive rework. The skill is not finding one perfect setting. It is matching effort to the cost of failure.
Long context still needs a good job definition
A 1M context window can hold a large repository or research pack, but capacity does not create structure. Tell the model which result you need, which files or sources have authority, what it may change, and how the result will be checked.
For coding, that may mean a passing test, a tight diff, and a browser check. For research, it may mean a cited answer with unsupported claims marked. Large context is most useful when the acceptance rule is small and clear.
Measure self-checking, not benchmark rank
Anthropic emphasizes that Opus 5 checks its work and persists through longer tasks. Its launch examples include root-cause debugging, building a test harness when no live feed was available, and checking browser layouts before handoff.
Those are vendor examples, not guarantees. They do point to a better evaluation target than a single benchmark score: does the model notice its own mistake before you do, and does that reduce the time needed to accept the result?
Run a ten-task switch test
Choose ten tasks from work you have already completed. Include a mix of easy and difficult jobs, but keep the expected result and acceptance check known.
For each task, record:
- the prompt, source files, tools, and permissions;
- the model and effort setting;
- elapsed time, token cost, retries, and tool failures;
- manual review and correction time;
- whether the result passed the original acceptance check.
Run the same set with your current model. Then calculate cost per accepted result, not cost per request. A more expensive run can still be cheaper work if it finishes a difficult task cleanly. A strong benchmark model can still be the wrong choice if review and correction eat the gain.
For the October comparison, use Opus 5 as the baseline and Opus 5.5 as the candidate. Start each run from the same repository commit or source pack in a fresh conversation. Keep the agent version, tools, permissions, timeout, and retry allowance fixed. Record any compatibility changes so they do not silently become extra help for one model.
Use one row per attempt:
Task ID | model/provider | effort/speed | input revision
Pass rule | output artifact | pass/fail | failure reason
Elapsed time | review minutes | corrections | retries
Recorded provider cost | tool cost | unknown charges
Calculate total recorded spend across all attempts divided by accepted results. Include failed attempts in the numerator. If nothing passes, report zero accepted results and total spend; do not divide by zero. Keep review minutes alongside monetary cost unless you have agreed on an hourly rate. Mark missing billing as unknown.
Before running, define what would justify switching: for example, all critical checks passing within your time and spend limits. Ten tasks are a pilot, not proof of general superiority. Repeat uncertain cases and retain the current configuration until the candidate passes the checks that matter to your work.
Publish the test as proof
A model comparison becomes portfolio evidence when another person can inspect it. Publish the task definitions, acceptance rules, settings, and a short failure log. Remove private code or client data, but do not hide unsuccessful runs.
This proof can support paid work without turning the article into an income promise. A developer can show a faster, checked migration workflow. A researcher can show a cited brief with lower correction time. An operator can show that a repeated automation completes within a defined cost and review budget.
The commercial value comes from the accepted deliverable. A model name alone does not demonstrate that you can deliver useful work.
When not to switch
Stay with the current model if your tasks already pass cheaply, if your toolchain does not support the candidate reliably, or if the new model produces larger answers that take longer to review. Keep a fallback for production work and test supported thinking and effort settings before migrating saved prompts. Preserve the old configuration and a clean starting task so recovery does not depend on replaying incompatible conversation history.
Also separate Anthropic’s benchmarks from your evidence. The company reports strong coding, knowledge-work, computer-use, and scientific results, but your provider, harness, tools, prompts, and acceptance rules make up the system you actually use.
Continue with the Anthropic model family guide and the model decision framework.
Official sources
All sources below were checked on October 26, 2026. Release dates describe the underlying announcement; the checked date describes this editorial review.
- Anthropic: Introducing Claude Opus 5 — published July 24, 2026; launch claims and examples.
- Claude Platform release notes — July 24 launch behavior, September 22 Opus 5.5 release, and October 15 fast-mode status.
- Claude model status and deprecations — active status and retirement commitment.
- Current model overview — current starting-model recommendation.
- Opus 5 specifications — legacy status, API ID, limits, and prices.
- Opus 5.5 specifications — current comparison model, limits, and prices.
- Opus 5.5 migration guide — request compatibility, conversation handling, and provider differences.
Put this to work
Compare models by accepted results, review effort, elapsed time, and cost rather than launch benchmarks alone.
Try
Compare Opus 5 with Opus 5.5 on ten representative tasks using fixed acceptance rules, explicit effort settings, and a budget.
Prove it worked
Keep the prompts, effort setting, tool permissions, failures, manual corrections, elapsed time, and cost for every accepted result.
Keep in view
- Opus 5 is active but listed as legacy; Anthropic's retirement commitment is no sooner than July 24, 2027.
- Opus 5.5 is the current comparison candidate, with lower standard token rates and different integration requirements.
- Test thinking, tool choice, and conversation handling before comparing task quality or cost.