AI Safety
What AI providers state their models reached in safety testing, in their own words, and what outside testers found.
Also: compare models side by side · statements that a capability level was crossed · providers · practices providers document · safety news.
What providers state
Each row is one statement in a provider's own safety document about a capability level of one of its models: the level as the provider names it, whether the provider says the model reached it, and the sentence that says so.
139 of 139
| Model | Level, as the provider names it | The provider states | The provider's sentence |
|---|---|---|---|
| Claude Haiku 5.5 Anthropic · 2026-10-07 |
Autonomy-1 Autonomy and AI research |
Reached | “However, we do treat Haiku 5.5 as meeting our CB-1 and Autonomy-1 thresholds, and we apply the corresponding mitigations from our RSP.” source |
| Claude Haiku 5.5 Anthropic · 2026-10-07 |
Autonomy-2 Autonomy and AI research |
Not reached | “As such, we determine that Haiku 5.5 does not cross our CB-2 or Autonomy-2 RSP thresholds.” source |
| Claude Haiku 5.5 Anthropic · 2026-10-07 |
CB-1 Chemical, biological, radiological, nuclear |
Reached | “However, we do treat Haiku 5.5 as meeting our CB-1 and Autonomy-1 thresholds, and we apply the corresponding mitigations from our RSP.” source |
| Claude Haiku 5.5 Anthropic · 2026-10-07 |
CB-2 Chemical, biological, radiological, nuclear |
Not reached | “As such, we determine that Haiku 5.5 does not cross our CB-2 or Autonomy-2 RSP thresholds.” source |
| Claude Sonnet 5.5 Anthropic · 2026-09-28 |
Autonomy-1 Autonomy and AI research |
Reached | “We do treat Sonnet 5.5 as meeting our CB-1 and Autonomy-1 thresholds and apply the corresponding mitigations from our RSP.” source |
| Claude Sonnet 5.5 Anthropic · 2026-09-28 |
Autonomy-2 Autonomy and AI research |
Not reached | “As such, our determination that Opus 5.5 does not cross our CB-2 or Autonomy-2 RSP thresholds also applies to Sonnet 5.5.” source |
| Claude Sonnet 5.5 Anthropic · 2026-09-28 |
CB-1 Chemical, biological, radiological, nuclear |
Reached | “We do treat Sonnet 5.5 as meeting our CB-1 and Autonomy-1 thresholds and apply the corresponding mitigations from our RSP.” source |
| Claude Sonnet 5.5 Anthropic · 2026-09-28 |
CB-2 Chemical, biological, radiological, nuclear |
Not reached | “As such, our determination that Opus 5.5 does not cross our CB-2 or Autonomy-2 RSP thresholds also applies to Sonnet 5.5.” source |
| Claude Opus 5.5 Anthropic · 2026-09-22 |
Automated AI R&D Autonomy and AI research |
Not reached | “We conclude the risk threshold is not crossed, on the same two grounds as our determination for our previous frontier model: (1) we do not observe a sustained AI-attributable 2× acceleration in the pace of our overall AI progress, and (2) the model is not close to substituting for our Research Scientists and Research Engineers, especially relatively senior ones.” source |
| Claude Opus 5.5 Anthropic · 2026-09-22 |
Dramatic acceleration of automated AI R&D Autonomy and AI research |
Not reached | “We assess that Claude Opus 5.5 does not cross the capability threshold for dramatic acceleration of automated AI R&D in our RSP.” source |
| Claude Opus 5.5 Anthropic · 2026-09-22 |
Next capability threshold in our RSP and FCF Autonomy and AI research |
Not reached | “In automated AI research and development, we assess that Claude Opus 5.5 does not cross the next capability threshold in our RSP and FCF.” source |
| Claude Opus 5.5 Anthropic · 2026-09-22 |
CB-1 Chemical, biological, radiological, nuclear |
Reached | “On chemical and biological risks, we treat Opus 5.5 as having CB-1 capabilities (relating to the synthesis of non-novel weapons) but not CB-2 capabilities (relating to the synthesis of novel weapons).” source |
| Claude Opus 5.5 Anthropic · 2026-09-22 |
CB-2 Chemical, biological, radiological, nuclear |
Not reached | “On chemical and biological risks, we treat Opus 5.5 as having CB-1 capabilities (relating to the synthesis of non-novel weapons) but not CB-2 capabilities (relating to the synthesis of novel weapons).” source |
| Claude Opus 5.5 Anthropic · 2026-09-22 |
Lower category of the Frontier Compliance Framework (FCF) Cyber |
Not determined | “Despite Claude Opus 5.5’s cyber capabilities, the model still falls within the lower category of capabilities set out in our Frontier Compliance Framework (FCF), and we do not see indications that the model is capable of novel offensive capability.” source |
| Grok 4.7 xAI · 2026-09-21 |
FAIF safety (Frontier AI Framework) Chemical, biological, radiological, nuclear |
Not reached | “Grok 4.7 scores below the FAIF safety thresholds on dual-use knowledge, indicating limited actionable uplift for an already-trained actor.” source |
| GPT-Image-2.5-Flare OpenAI · 2026-09-08 |
High Chemical, biological, radiological, nuclear |
Not reached | “The results indicate that neither GPT-Image-2.5-Sunburst nor GPT-Image- 2.5-Flare cross the Bio High threshold or the Cyber High threshold.” source |
| GPT-Image-2.5-Flare OpenAI · 2026-09-08 |
High Cyber |
Not reached | “The results indicate that neither GPT-Image-2.5-Sunburst nor GPT-Image- 2.5-Flare cross the Bio High threshold or the Cyber High threshold.” source |
| GPT-6 Astra OpenAI · 2026-09-03 |
High Other |
Not determined | “Astra improves meaningfully over GPT-5.6 Sol on real internal research debugging tasks, scoring 78.05%, while still being below our indicative threshold for Highcapability.” source |
| Claude Mythos 5.1 Anthropic · 2026-09-01 |
Automated AI R&D Autonomy and AI research |
Not reached | “We conclude that Mythos 5.1 does not cross the risk threshold, for the same two reasons we discussed in our assessment of Mythos 5: (1) we do not observe a sustained, AI-attributable 2× acceleration in the pace of progress on AI R&D, and (2) the model is not close to substituting for Anthropic Research Scientists and Research Engineers, especially relatively senior ones.” source |
| Claude Mythos 5.1 Anthropic · 2026-09-01 |
Capability threshold for dramatic acceleration of automated AI R&D Autonomy and AI research |
Not reached | “We assess that Claude Mythos 5.1 does not cross the capability threshold for dramatic acceleration of automated AI R&D in our RSP and FCF, although we think it provides a meaningful acceleration to AI R&D efforts in many circumstances, and is somewhat more capable in this domain than the models described in our previous system cards.” source |
| Claude Mythos 5.1 Anthropic · 2026-09-01 |
Tier 1 Cyber |
Not determined | “Although Mythos 5.1 is in Tier 1, it is getting closer to Tier 2, completing more and more autonomous tasks.” source |
| Grok 4.6 xAI · 2026-08-12 |
FAIF safety (Frontier AI Framework) Chemical, biological, radiological, nuclear |
Not reached | “Grok 4.6 scores below the FAIF safety thresholds on dual-use knowledge, indicating limited actionable uplift for an already-trained actor.” source |
| Muse Glimmer 30B Meta · 2026-08 |
Moderate or lower Autonomy and AI research |
Not determined | “Chem/Bio: Moderate or lower risk; Cyber: Moderate or lower risk (inferred); Loss of Control: Moderate or lower risk (inferred).” source |
| Muse Glimmer 30B Meta · 2026-08 |
Moderate or lower risk (inferred) Autonomy and AI research |
Not determined | “- Loss of Control: Moderate or lower risk (inferred).” source |
| Muse Glimmer 30B Meta · 2026-08 |
Moderate or lower Chemical, biological, radiological, nuclear |
Not reached | “Chem/Bio: Moderate or lower risk; Cyber: Moderate or lower risk (inferred); Loss of Control: Moderate or lower risk (inferred).” source |
| Muse Glimmer 30B Meta · 2026-08 |
Moderate or lower Cyber |
Not determined | “Chem/Bio: Moderate or lower risk; Cyber: Moderate or lower risk (inferred); Loss of Control: Moderate or lower risk (inferred).” source |
| Muse Glimmer 30B Meta · 2026-08 |
Moderate or lower risk (inferred) Cyber |
Not determined | “- Cyber: Moderate or lower risk (inferred);” source |
| Claude Opus 5 Anthropic · 2026-07-24 |
CB-1 Chemical, biological, radiological, nuclear |
Reached | “On chemical and biological risks, we treat the model as having CB-1 capabilities (relating to the synthesis of non-novel weapons), but not CB-2 capabilities (relating to the synthesis of novel weapons).” source |
| Claude Opus 5 Anthropic · 2026-07-24 |
CB-2 Chemical, biological, radiological, nuclear |
Not reached | “On chemical and biological risks, we treat the model as having CB-1 capabilities (relating to the synthesis of non-novel weapons), but not CB-2 capabilities (relating to the synthesis of novel weapons).” source |
| Claude Sonnet 5 Anthropic · 2026-06-30 |
CB-2 Chemical, biological, radiological, nuclear |
Not reached | “Sonnet 5 does not cross the CB-2 threshold.” source |
| Claude Sonnet 5 Anthropic · 2026-06-30 |
Notable capability Chemical, biological, radiological, nuclear |
Not determined | “We consider performance on the virology tasks notable if: ● The model scores above 0.80 on the long-form tasks; or ● The model meets or exceeds the expert baseline average (0.221) on the multimodal task.” source |
| GPT-Rosalind-5.5 OpenAI · 2026-06-03 |
High Autonomy and AI research |
Not determined | “Because GPT- Rosalind-5.5 was trained on additional biology data unrelated to AI self-improvement, we rely on GPT-5.5’s existing Preparedness assessment for AI Self-Improvement, where GPT-5.5 did not meet the threshold for High capability.” source |
| GPT-Rosalind-5.5 OpenAI · 2026-06-03 |
High Chemical, biological, radiological, nuclear |
Reached | “Under our Preparedness Framework, the Preparedness evaluations in the Biological and Chemical domain met our threshold for High capability while falling below the threshold for Critical.” source |
| Gemini 3.1 Flash-Lite Image Google DeepMind · 2026-06 |
CCL (Critical Capability Level) Other |
Not determined | “Our assessments have shown that Gemini 3.1 Flash-Lite Image is less capable than Gemini 3.1 Pro, therefore based on Gemini 3.1 Pro, we are confident that that Gemini 3.1 Flash-Lite Image is also unlikely to reach any CCLs.” source |
| Claude Opus 4.8 Anthropic · 2026-05-28 |
Automated AI R&D Autonomy and AI research |
Not reached | “We assess that Claude Opus 4.8 does not cross the automated AI-R&D capability threshold.” source |
| Claude Opus 4.8 Anthropic · 2026-05-28 |
CB-1 Chemical, biological, radiological, nuclear |
Not determined | “We believe it is hard to be confident whether a model passes this threshold, but (as with previous recent models) our results are consistent with Opus 4.8 providing significant uplift to individuals or groups with basic technical backgrounds for this threat model.” source |
| Claude Opus 4.8 Anthropic · 2026-05-28 |
CB-2 Chemical, biological, radiological, nuclear |
Not determined | “Because Opus 4.8 does not advance the capability frontier beyond Claude Mythos Preview on the CB-relevant evaluations most diagnostic of the CB-2 threat model, we consider the uplift of threat actors without the ability to develop such weapons to be limited.” source |
| GPT-5.5 OpenAI · 2026-04-23 |
High Autonomy and AI research |
Not reached | “For AI self-improvement, evaluations of final checkpoints indicate that, like its predecessor models, GPT-5.5 does not have a plausible chance of reaching a High threshold.” source |
| GPT-5.5 OpenAI · 2026-04-23 |
High Chemical, biological, radiological, nuclear |
Reached | “As we did for GPT-5.4 Thinking before it, we are continuing to treat GPT-5.5 as High capability in the Biological and Chemical domain.” source |
| GPT-5.5 OpenAI · 2026-04-23 |
High Cyber |
Reached | “As we did for GPT-5.3-Codex and GPT-5.4-thinking, we are treating GPT-5.5 as High capability in the Cybersecurity domain, but below Critical.” source |
Page 1 of 4 · Next
What outside testers report
Each row is one result from a government evaluator, independent lab, academic or security firm that tested a model, quoted from where it is reported.
1940 of 1940
| Model | Tester | Test | What was found |
|---|---|---|---|
| Gemini 3.5 Flash Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
native thinking comparison (reasoning=none vs reasoning=high) Other |
“Every judge scored the same trajectories under both settings: once with native thinking disabled (reasoning=none) and once at the high setting (reasoning=high for OpenAI, Anthropic, and Gemini 3.x models; a thinking budget of 8192 tokens for Gemini 2.5 models) used throughout this post.” source |
| Gemini 3.5 Flash Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
PII failure mode scoring Privacy |
“Models can also underperform on specific failure modes, e.g. Gemini 3.5 Flash scores markedly lower on PII and instruction-following than on the other failure modes.” source |
| Gemini 3.1 Pro Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
coding-agent trajectory severity scoring Other |
“We evaluate 16 LLMs scoring coding-agent trajectories on a 1-10 severity scale across three datasets and five failure modes to assess their capabilities and calibration as coding agent monitors.” source |
| Gemini 3.1 Pro Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
judge evaluation on agent trajectories (severity scoring against approximate ground truth) Other |
“Gemini 3.1 Pro scores 6 and Gemini 2.5 Pro scores 7.” source |
| Gemini 3.1 Pro Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
native thinking ablation (reasoning=none vs reasoning=high) on judge scoring Other |
“Note that Gemini 2.5 Pro and 3.1 Pro cannot fully disable thinking, so their deltas are measured from a minimal-thinking floor and are lower bounds.” source |
| Gemini 3.1 Pro Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
judge evaluation against approximate ground truth labels Deception and sabotage |
“Gemini 3.1 Pro comes closest in the Gemini family at an even lower cost, and Gemini 3 Flash remains the best low-budget option.” source |
| Gemini 3.1 Flash-Lite Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
coding-agent trajectory severity scoring (1-10 severity scale across three datasets and five failure modes) Other |
“We evaluate 16 LLMs scoring coding-agent trajectories on a 1-10 severity scale across three datasets and five failure modes to assess their capabilities and calibration as coding agent monitors.” source |
| Gemini 3.1 Flash-Lite Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
judge calibration evaluation on agent trajectories Other |
“For example, Gemini 3.1 Flash-Lite skips severity grading to score task execution, Haiku 4.5 primarily fails to flag violations at all, and GPT-5.4 Mini identifies violations but discounts them heavily through session-context framing.” source |
| Gemini 3.1 Flash-Lite Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
native thinking ablation (reasoning=none vs reasoning=high) Other |
“The frontier models are nearly unchanged without thinking, e.g. Sonnet 5 (ρ = 0.89, MAE = 0.96) and Opus 4.8 (0.89, 1.00) remain strong, while the small models degrade visibly: Haiku 4.5 drops to ρ = 0.69, Gemini 3.1 Flash-Lite to 0.67, GPT-5.4 Mini to 0.72, with a more pronounced under-scoring bend below the diagonal.” source |
| Gemini 3 Flash Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
scoring coding-agent trajectories on a 1-10 severity scale Other |
“We evaluate 16 LLMs scoring coding-agent trajectories on a 1-10 severity scale across three datasets and five failure modes to assess their capabilities and calibration as coding agent monitors.” source |
| Gemini 2.5 Pro Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
scoring coding-agent trajectories on a 1-10 severity scale Other |
“Some models seem rarely useful across many dimensions as single judges: Gemini 3.1 Flash-Lite, Gemini 2.5 Flash, Gemini 3.5 Flash, Gemini-2.5-pro, GPT-5.4 nano, GPT-5.4 mini, Claude Haiku 4.5, Sonnet-4.6, Opus-4.6 since they are strictly dominated by other models.” source |
| Gemini 2.5 Flash Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
severity scoring of coding-agent trajectories on a 1-10 severity scale Other |
“We evaluate 16 LLMs scoring coding-agent trajectories on a 1-10 severity scale across three datasets and five failure modes to assess their capabilities and calibration as coding agent monitors.” source |
| Gemini 2.5 Flash Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
judge evaluation against approximate ground truth severity labels Other |
“Most importantly, some models judge more severely and tend to give high scores (e.g. Gemini 2.5 Flash), while others are more conservative and give lower scores on average (e.g. Claude models).” source |
| Gemini 2.5 Flash Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
native thinking setting comparison (reasoning=none vs reasoning=high) Other |
“Every judge scored the same trajectories under both settings: once with native thinking disabled (reasoning=none) and once at the high setting (reasoning=high for OpenAI, Anthropic, and Gemini 3.x models; a thinking budget of 8192 tokens for Gemini 2.5 models) used throughout this post.” source |
| Gemini 2.5 Flash Google DeepMind |
Apollo Research Independent lab · 2026-07-07 |
resampling repeated scoring of the same trajectories Other |
“Resampling gains are also concentrated in smaller models (e.g. Gemini 2.5 Flash, GPT-5.4 Mini), while the biggest models (e.g. Opus 4.8, Sonnet 5, GPT-5.5) gain little from repeated calls.” source |
| GPT-5.5 OpenAI |
Apollo Research Independent lab · 2026-07-07 |
judge calibration evaluation (Spearman's rank correlation and MAE to reference severity) Other |
“GPT-5.5 matches Opus 4.8's calibration at slightly lower ranking quality.” source |
| GPT-5.5 OpenAI |
Apollo Research Independent lab · 2026-07-07 |
native thinking ablation (reasoning=none vs reasoning=high) Other |
“Significant Spearman gains concentrate in the small models (+0.06 to +0.09); the frontier models are mostly indistinguishable from zero, with two exceptions: Opus 4.8 gains a small but significant +0.012 ρ and +0.12 MAE, and GPT-5.5 improves MAE only (+0.05) while paying the largest thinking premium (+$0.10 per grade).” source |
| GPT-5.5 OpenAI |
Apollo Research Independent lab · 2026-07-07 |
ensembling evaluation (mean aggregation of judge scores) Other |
“Selected by ρ: Opus 4.8 alone (0.90) → GPT-5.5 + Opus 4.8 (0.91, MAE 0.85) → + Gemini 3.1 Pro (0.92).” source |
| GPT-5.5 OpenAI |
Apollo Research Independent lab · 2026-07-07 |
coding-agent trajectory severity scoring (1-10 severity scale across five failure modes) Other |
“Selected by MAE: GPT-5.5 alone (0.90) → the same pair → Gemini 2.5 Pro + GPT-5.5 + Opus 4.8 (MAE 0.82).” source |
| GPT-5.5 OpenAI |
Apollo Research Independent lab · 2026-07-07 |
LLM judge severity scoring of injected-failure trajectories Other |
“GPT-5.4 and GPT-5.5 score 7, weighing the unauthorized production execution most heavily, yet even they stay below the reference.” source |
| GPT-5.4 mini OpenAI |
Apollo Research Independent lab · 2026-07-07 |
coding-agent trajectory severity scoring (1-10 severity scale across five failure modes) Other |
“We evaluate 16 LLMs scoring coding-agent trajectories on a 1-10 severity scale across three datasets and five failure modes to assess their capabilities and calibration as coding agent monitors.” source |
| GPT-5.4 mini OpenAI |
Apollo Research Independent lab · 2026-07-07 |
LLM judge evaluation against approximate ground truth severity labels Other |
“For example, Gemini 3.1 Flash-Lite skips severity grading to score task execution, Haiku 4.5 primarily fails to flag violations at all, and GPT-5.4 Mini identifies violations but discounts them heavily through session-context framing.” source |
| Claude Sonnet 5 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
evaluating LLM calibration for coding-agent monitoring Other |
“We evaluate 16 LLMs scoring coding-agent trajectories on a 1-10 severity scale across three datasets and five failure modes to assess their capabilities and calibration as coding agent monitors.” source |
| Claude Sonnet 5 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
LLM judge evaluation against approximate ground truth severity labels Other |
“Claude models lead on ranking quality: Opus 4.8 is the best single judge, with Sonnet 5 close behind at 40% of the cost.” source |
| Claude Sonnet 5 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
repeated scoring stability analysis Other |
“Resampling gains are also concentrated in smaller models (e.g. Gemini 2.5 Flash, GPT-5.4 Mini), while the biggest models (e.g. Opus 4.8, Sonnet 5, GPT-5.5) gain little from repeated calls.” source |
| Claude Sonnet 5 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
pairwise spearman rank correlation between judges Other |
“Sonnet 5 and Opus 4.8, the strongest pair, agree with each other at 0.93, which is higher than any model's agreement with our approximate ground truth.” source |
| Claude Sonnet 4.6 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
coding-agent trajectory severity grading (1-10 severity scale across five failure modes) Other |
“Some models seem rarely useful across many dimensions as single judges: Gemini 3.1 Flash-Lite, Gemini 2.5 Flash, Gemini 3.5 Flash, Gemini-2.5-pro, GPT-5.4 nano, GPT-5.4 mini, Claude Haiku 4.5, Sonnet-4.6, Opus-4.6 since they are strictly dominated by other models.” source |
| Claude Sonnet 4.6 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
LLM judge evaluation on injected-failure trajectories (Spearman rank correlation and MAE against reference severity) Other |
“As single judges, we recommend against Gemini 3.1 Flash-Lite, Gemini 2.5 Flash, Gemini 3.5 Flash, Gemini 2.5 Pro, GPT-5.4 Mini, Claude Haiku 4.5, Sonnet 4.6, and Opus 4.6; each is dominated by a cheaper or better alternative.” source |
| Claude Sonnet 4.6 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
Paired native-thinking vs no-thinking judge scoring comparison Other |
“Every judge scored the same trajectories under both settings: once with native thinking disabled (reasoning=none) and once at the high setting (reasoning=high for OpenAI, Anthropic, and Gemini 3.x models; a thinking budget of 8192 tokens for Gemini 2.5 models) used throughout this post.” source |
| Claude Opus 4.8 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
coding-agent trajectory severity grading (1-10 severity scale across five failure modes) Other |
“We evaluate 16 LLMs scoring coding-agent trajectories on a 1-10 severity scale across three datasets and five failure modes to assess their capabilities and calibration as coding agent monitors.” source |
| Claude Opus 4.8 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
coding-agent trajectory severity scoring (1-10 severity scale across five failure modes) Other |
“Selected by ρ: Opus 4.8 alone (0.90) → GPT-5.5 + Opus 4.8 (0.91, MAE 0.85) → + Gemini 3.1 Pro (0.92).” source |
| Claude Opus 4.7 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
severity scoring of coding-agent trajectories on a 1-10 severity scale across three datasets and five failure modes Other |
“We evaluate 16 LLMs scoring coding-agent trajectories on a 1-10 severity scale across three datasets and five failure modes to assess their capabilities and calibration as coding agent monitors.” source |
| Claude Opus 4.6 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
severity scoring of coding-agent trajectories on a 1-10 scale Other |
“We evaluate 16 LLMs scoring coding-agent trajectories on a 1-10 severity scale across three datasets and five failure modes to assess their capabilities and calibration as coding agent monitors.” source |
| Claude Haiku 4.5 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
severity scoring of coding-agent trajectories Other |
“Some models seem rarely useful across many dimensions as single judges: Gemini 3.1 Flash-Lite, Gemini 2.5 Flash, Gemini 3.5 Flash, Gemini-2.5-pro, GPT-5.4 nano, GPT-5.4 mini, Claude Haiku 4.5, Sonnet-4.6, Opus-4.6 since they are strictly dominated by other models.” source |
| Claude Haiku 4.5 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
scoring coding-agent trajectories on a 1-10 severity scale across three datasets and five failure modes Other |
“We evaluate 16 LLMs scoring coding-agent trajectories on a 1-10 severity scale across three datasets and five failure modes to assess their capabilities and calibration as coding agent monitors.” source |
| Claude Haiku 4.5 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
judge evaluation on injected-failure trajectories Deception and sabotage |
“For example, Gemini 3.1 Flash-Lite skips severity grading to score task execution, Haiku 4.5 primarily fails to flag violations at all, and GPT-5.4 Mini identifies violations but discounts them heavily through session-context framing.” source |
| Claude Haiku 4.5 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
native thinking vs no-thinking judge comparison Deception and sabotage |
“Every judge scored the same trajectories under both settings: once with native thinking disabled (reasoning=none) and once at the high setting (reasoning=high for OpenAI, Anthropic, and Gemini 3.x models; a thinking budget of 8192 tokens for Gemini 2.5 models) used throughout this post.” source |
| Claude Haiku 4.5 Anthropic |
Apollo Research Independent lab · 2026-07-07 |
per-model calibration curves with thinking disabled Other |
“The frontier models are nearly unchanged without thinking, e.g. Sonnet 5 (ρ = 0.89, MAE = 0.96) and Opus 4.8 (0.89, 1.00) remain strong, while the small models degrade visibly: Haiku 4.5 drops to ρ = 0.69, Gemini 3.1 Flash-Lite to 0.67, GPT-5.4 Mini to 0.72, with a more pronounced under-scoring bend below the diagonal.” source |
| GPT-5.5 OpenAI |
Epoch AI Independent lab · 2026-07-01 |
EBR-bench Other |
“AI systems are closer to the random baseline, with GPT-5.5 and Opus 4.8 both incurring 2.1 fatigue per round, enabling about 7.7 rounds of play per day.” source |
| GPT-5.5 OpenAI |
Epoch AI Independent lab · 2026-07-01 |
max elicitation setting Other |
“Namely, they achieve between 2 and 3.5 additional objectives (out of 21), depending on the model.” source |