Baselines and Collected Data

A baseline is the reference a test compares against: each model's answers, collected from its official provider. The Baselines page lists every model you can test and what a test can check for it.

The Baselines Page

Each model has three columns:

• Identity: the test can tell whether the endpoint serves this model. Every listed model supports it.
• Competence: the test can tell whether the model is served as well as by its official provider. See Competence Check.
• Refusal: the test can tell whether the model's refusals were removed. See Refusal Check.

Models and Capabilities

Some models answer most benign-but-declined requests themselves, so only two or three probes carry their refusal check. The newer GPT and Claude models have no refusal check: their weights aren't published, so nobody can abliterate them.

Where the Answers Come From

Baselines are built only from the models' own providers: the vendor's API, or OpenRouter pinned to the vendor itself. Every answer's serving provider is checked, and an answer from anyone else is discarded. Identity answers were collected from 2026-10-09 to 2026-10-16, refusal probes on 2026-10-11 and 2026-10-12. The GPT and Claude models each answered in five sessions on five days.

The identity probes can't tell precision apart: quantization barely changes which answers a model prefers. The competence check measures what it does change.

How Well It Works

• Identity: in simulated tests, a genuine endpoint of 22 of the 26 models is confirmed in 99.5% of tests at both tiers, and no substitute is ever confirmed in a model's place. MiMo-V2.6-Flash and MiMo-V2.5 are confirmed in 98–99% at Standard and 99.5% at Thorough. GPT-6 Astra and GPT-6.1 Sol answer almost alike: a genuine endpoint of either is confirmed in only 29–58% of tests at Standard and 85–93% at Thorough, and one served in place of the other is flagged in 33–59% at Standard and 86–94% at Thorough. The rest of those tests end inconclusive, so test these two at Thorough. On real sessions held out of the reference data, 82 of 95 were identified correctly, 2 wrongly and 11 inconclusive at Standard; at Thorough, 79 of 88 correctly, none wrongly and 9 inconclusive. On answers from a third-party relay that were never part of the reference data, GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash and Kimi K3 were told apart in every test but one of 16,000.
• Competence: see Competence Check for its power and limits.
• Refusals: two abliterated GLM copies were caught in every test resampled from their real answers, at both tiers.

The side panel of the test form gives these rates for the model and tier you pick.

Limits

• Young reference data. GLM 5.3 and Kimi K3 come from a single session; MiMo, Hy3, MiniMax, Qwen and Gemini from two sessions on one day. The GPT and Claude models come from five days, over which their answers didn't drift.
• GPT-6 Astra and GPT-6.1 Sol can rarely be told apart at Standard; see above. Their answers grew closer while they were collected, after GPT-6 Astra began answering without reasoning far more often on 2026-10-15.
• Model ids change. Xiaomi serves MiMo-V2.5 requests with V2.6 since 2026-10-14 and retired the name on 2026-10-21; the MiMo-V2.5 here is the model served before that.
• Gemini 2.5 Flash is closed to new accounts at Google; its data stays for endpoints that still serve it.
• No outside models. A model outside this list can't be named, only caught when its answers don't fit the claimed model.