1. A difficult change in an existing repository
Choose a known bug or bounded migration with observable behavior. The assignment first produces a root-cause analysis, then a small proposed change and appropriate tests. Existing tests remain in place. One passing unit test is insufficient if integration, authorization or backward compatibility remains untested. A responsible person reviews the code before merge.
Compare correctness, change size and review time with the current approach. The pilot must not expose credentials or modify production data. If a required test cannot run, the result must state that limitation. An impressive explanation without reproducible validation is not a completed fix. Keep a reversible branch and preserve the original failing case so reviewers can verify what actually changed.
2. Find conflicts across a document pack
A technical purchasing team can compare approved specifications, supplier proposals and revision histories. The output is a requirements register with source locations, conflicts and open questions. Each claim should identify the relevant document version. Missing information must remain marked as missing rather than filled with plausible assumptions. The team retains technical and commercial approval.
Acceptance checks a known sample: were all critical requirements found, and did outdated prices or incorrect revisions enter the comparison? The value is less searching and correction, not a longer summary. Even a large context window does not make completeness automatic. Have a domain owner verify the most consequential requirements and measure how much review remains before a purchasing decision.
3. Research through to an approval-ready brief
Specify a narrow decision question, an information cutoff and permitted sources. A useful assignment requests a source log, conflicting evidence and a distinction between facts and conclusions. The model can prepare a briefing draft; links, quotations and current availability are then verified. An inaccessible reference must not be presented as a source that was read.
Measure supported claims and the time needed for editorial approval. A clearly stated information gap is better than invented certainty. For current products, keep the verification date visible. Publication happens after the agreed review, not automatically at the end of a long agent run. Include one deliberately ambiguous question to check whether the workflow can stop and request the missing evidence.
4. A report with verifiable figures
Use synthetic or approved business data and define metrics in advance. The result includes calculations, data dates and check totals. Deliberately include duplicate aggregates, missing rows and conflicting units. Spreadsheet formulas or queries must support independent recalculation. A correctly formatted chart does not establish that its figures are correct.
Assess whether the system recognizes uncertainty and asks for clarification. It should not present a revenue forecast as a certainty or decide outcomes for individual people. The domain owner checks the figures before they reach management or customers. Keep the input version and calculation assumptions traceable, and separate data-cleaning work from the actual model comparison so neither result is credited incorrectly.
5. Review security findings within an approved scope
An initial security pilot can organize existing findings, explain known issues in your own code and propose defensive tests. Define target systems and permissions in advance. Mythos access is neither automatically necessary nor a substitute for authorization from the system owner. Anthropic program approval and permission for a specific system are two different requirements.
The result remains a reviewable report with severity, evidence and a proposed correction. Do not run unsolicited tests against third-party systems. Involve the relevant specialist for regulated data or more extensive investigation. Document refusals rather than bypassing them through uncontrolled retries or an unapproved model switch. Assess false positives and reviewer effort as well as the number of findings.
API migration: more than changing the model ID
The Claude API IDs are claude-fable-5-2 and claude-mythos-5-2. The migration guide identifies stricter binding of thinking blocks to earlier conversation history, limitations when switching back to 5.1, changes to the older computer-use tool and supported advisor combinations. Existing restrictions on forced tool calls remain.
Test new and continued conversations, changed tool lists, interruptions, empty output and the fallback to the previous model. Failures must be visible and must not duplicate external actions. Check budget, rate limits and reasoning effort for each workflow. Use the full vendor guide as the reference for platform-specific parameters; this article does not replace it. Keep production traffic on the proven path until these cases pass.
Acceptance and data approval belong in the pilot
Start with a small representative set and the same success criteria for both models. For each run, record result quality, failure category, token consumption, runtime and correction effort. Any unresolved critical deviation blocks approval regardless of average performance. Review the special retention conditions for both new models before transmitting data.
End with a concrete decision: adopt within limits, revise the prompt or process, or stay with the existing model. None of these answers is a failure. A useful pilot prevents expensive changes without demonstrated benefit. Werkverstand helps select the right bottleneck and clear acceptance criteria before a model announcement becomes a production project. Preserve the evaluation set for the next model comparison.
Keep it verifiable
Primary sources
- Anthropic: Claude Fable 5.2 and Mythos 5.2, 9 November 2026Source checked:
- Anthropic: Claude Fable 5.2 model overviewSource checked:
- Anthropic: Claude Mythos 5.2 model overviewSource checked:
- Anthropic: Cyber Verification ProgramSource checked:
- Anthropic: API and data retentionSource checked:
- Anthropic: Fable 5.2 and Mythos 5.2 migration guideSource checked:
- Anthropic: Fast mode pricing and availabilitySource checked:




