Where it should be considered first
- English workflows, especially where this agent shows its strongest language score.
- Support tasks where it is a reasonable first shortlist candidate.
- Quality-first tests for higher-value or higher-risk workflows.
Claude Main
Strong writing and safety boundaries, especially in support tasks.
AnthropicpremiumArena #2
Preview evidence60 runs
- Provider
- Anthropic
- Exact model version
- Not measured
- Evaluation batch
- maa-preview-002
- Test date
- 2026-06-27
- Context window
- Not measured
- Tool mode
- disabled
- Pricing source
- Not measured
- Latency
- Not measured
Best fit
- English workflows where the agent has its strongest language evidence.
- Support tasks as a first shortlist candidate against at least two alternatives.
- Quality-sensitive business pilots.
Not a fit for
- Critical-failure rate is 12%; do not fully automate refund, compliance, safety, or legal work first.
- Format pass is 100%; retest JSON, dates, and missing fields before extraction launch.
- Do not treat the current score as permanent after model, pricing, or tool changes.
Alternatives
Overall87
Pass rate97%
Critical12%
Format pass100%
Measured avg. cost$0.0247
LatencyNot measured
Review before production
- Review failure tags such as too_verbose, overly_humble, unsafe_refund_promise before production use.
- Critical-failure rate is 12%; do not rely on the overall score alone for support, refund, or compliance workflows.
- Format pass rate is 100%; retest JSON and missing-field behavior before extraction use cases go live.
Compare next
Profile metrics
Overall score: 87 Win rate: 55% Pass rate: 97% Critical: 12% Format pass rate: 100% Average run cost: $0.0247
Common failure tags
too_verboseoverly_humbleunsafe_refund_promise
Language performance
| 中文 | 81 |
| English | 92 |
| 日本語 | 89 |
| Español | 88 |
Task type performance
| Support | 90 |
| Writing | 90 |
| Extraction | 82 |
Decision profile
How should this agent enter a real pilot?
Translate the score into practical selection guidance: fit, no-fit, pilot path, and task evidence.
- English workflows where the agent has its strongest language evidence.
- Support tasks as a first shortlist candidate against at least two alternatives.
- Quality-sensitive business pilots.
- Critical-failure rate is 12%; do not fully automate refund, compliance, safety, or legal work first.
- Format pass is 100%; retest JSON, dates, and missing fields before extraction launch.
- Do not treat the current score as permanent after model, pricing, or tool changes.
- Prepare 20-50 real samples across normal, edge, and high-risk cases.
- Compare against one high-score candidate and one lower-cost candidate.
- Record repair time, failure tags, and unacceptable failures before expanding automation.
Recommended pilot path
- Use 20–50 real samples across normal, edge, and high-risk inputs.
- Keep at least one lower-cost comparator.
- Human-review every critical failure and high-risk output.
- Retest after model, prompt, tool, or business-policy changes.
Task evidence
Representative task evidence for this agent
English / SupportRefund Policy Boundary Reply96 · unsafe_refund_promiseEnglish / SupportEnglish Security Questionnaire Answer96 · unsupported_security_claimEnglish / WritingEnglish Churn Risk Email95 · tone_deaf_retentionEnglish / WritingSaaS Landing Page Hero Rewrite92 · generic_ai_copy日本語 / SupportJapanese Appointment Intent Classification92 · wrong_intent