Models and reasoning effort by task

Find a model for your task

Find your task class to see candidate models, reasoning effort, evidence, and permitted alternatives. Use the API section to apply these recommendations before a run.

Planning and consequential decisions require a strong model. Available quota may change the provider choice, but it does not lower that requirement.

Policy baseline means the recommendation follows the routing policy and has no matching controlled study. Rows marked Controlled pilot or Controlled evidence link to measured outcomes.

Historical GPT-5.6 HTML/UI and source-review pilots

These August measurements belong to GPT-5.6 and do not measure the September GPT-6 recommendations.

The next identical-scenario matrix explicitly includes gpt-5.5/medium and gpt-5.4-mini/high. They are scheduled candidates, not measured pilot results; the published recommendations do not infer success from catalog presence.

Operational review evidence

The Quality Studio dataset contains no eligible operational runs for Terra or Sol. Its confirmed findings cannot measure missed defects; the seeded-defect study supplies recall.

Use these recommendations in C#

API guide: inputs, return fields, and empty results. The host chooses among candidates using fresh quota, then calls ModelRouter.Route for the actual attempt. Task uncertainty, correctness floors, previous failures, and operator pins still apply.

var advice = TaskClassRecommendationCatalog.Default
    .Recommend(TaskClass.HtmlUiImplementation);

foreach (var route in advice.Candidates)
    Console.WriteLine($"{route.Model} / {route.ThinkingLevel}");
Console.WriteLine(advice.RationaleVersion); // persist with the Agent Studio card

var selection = TaskClassRecommendationCatalog.Default
    .Select(advice, currentQuotaSnapshot, DateTime.UtcNow);
// Unknown/stale quota => RecommendationOnly and Selected == null.

// At launch, route the concrete card as usual. The class prior cannot lower
// ModelRouter's uncertainty score, correctness floor, or semantic promotion.

Protocols: HTML/UI study · source-review study · canonical routing policy · delegation economy

Studies still needed

Each new study needs at least five scenarios and twenty attempts per model-and-effort combination. Results must include outcome coverage, repeat attempts, and dated costs.

Planned studies and acceptance criteria

1 · Security assessment

Seed authn/authz, injection, secret, and unsafe-remediation cases. Score severity-weighted recall and unsafe advice. Sol/xhigh remains the hard floor throughout.

2 · Redundancy detection

Known clone pairs plus intentional variants. Score semantic-duplicate recall, precision, and false merge suggestions across one- and multi-subsystem corpora.

3 · Graphical-quality judgment

Human-calibrate the defect checklist, randomize blinded screenshot order, and compare Mini/high with stronger jurors using agreement and pairwise ranking accuracy.

4 · Consistency checking

Hide invariant manifests across schemas, prose, and UI states. Score contradiction recall/precision and separate compact bounded checks from open-ended research.

5 · Coding classes

Expand feature evidence and add heavy design, mechanical chore, docs, and research suites with native tests, factual keys, review grades, and semantic-reissue outcomes.

Qualification and calibration

Minimum 20 attempts per route, five scenarios, 70% outcome coverage, and at most 10% semantic reissue. Add human/jury agreement gates where judgment is subjective.