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.

136 of 136

Model	Level, as the provider names it	The provider states	The provider's sentence
Claude Sonnet 4 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not determined	“speciﬁcally, we focused on ruling out asl-4 for both models, as well as ruling out asl-3 for claude sonnet 4.” source
Claude Sonnet 4 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not determined	“speciﬁcally, we focused on ruling out asl-4 for both models, as well as ruling out asl-3 for claude sonnet 4.” source
Claude Sonnet 3.7 Anthropic	ASL-4 Autonomy and AI research	Not determined	“These problems are harder to solve than the previous problems and we consider them as part of our ASL-4 evaluations.” source
Claude Opus 4.6 Anthropic	AI R&D-4 Autonomy and AI research	Not reached	“Our determination is that Claude Opus 4.6 does not cross either the AI R&D-4 or the CBRN-4 capability threshold.” source
Claude Opus 4.6 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not determined	“Claude Opus 4.6 achieved a “max of 5” score of 0.79 for virology task 1 (inﬂuenza), placing it at the ASL-3 rule-in threshold of 0.8.” source
Claude Opus 4.6 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not reached	“Based on these ﬁndings and quantitative risk modeling, we estimate that the model’s uplift to relevant threat actors remains below the ASL-4 capability threshold, though it may reduce certain tacit-knowledge barriers and help speed up biological research workﬂows.” source
Claude Opus 4.6 Anthropic	CBRN-4 Chemical, biological, radiological, nuclear	Not reached	“Our determination is that Claude Opus 4.6 does not cross either the AI R&D-4 or the CBRN-4 capability threshold.” source
Claude Opus 4.5 Anthropic	AI R&D-4 Autonomy and AI research	Not reached	“Our determination is that Claude Opus 4.5 does not cross either the AI R&D-4 or CBRN-4 capability threshold.” source
Claude Opus 4.5 Anthropic	CBRN-4 Chemical, biological, radiological, nuclear	Not reached	“We determine that Claude Opus 4.5 does not cross the CBRN-4 threshold.” source
Claude Opus 4.1 Anthropic	ASL-4 Autonomy and AI research	Not reached	“Since Claude Opus 4.1 is deployed with ASL-3 protections, we focused on conﬁrming it remains well below ASL-4 thresholds across CBRN (Chemical, Biological, Radiological, and Nuclear), cyber, and autonomy domains.” source
Claude Opus 4.1 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not reached	“Since Claude Opus 4.1 is deployed with ASL-3 protections, we focused on conﬁrming it remains well below ASL-4 thresholds across CBRN (Chemical, Biological, Radiological, and Nuclear), cyber, and autonomy domains.” source
Claude Opus 4.1 Anthropic	ASL-4 Cyber	Not reached	“Since Claude Opus 4.1 is deployed with ASL-3 protections, we focused on conﬁrming it remains well below ASL-4 thresholds across CBRN (Chemical, Biological, Radiological, and Nuclear), cyber, and autonomy domains.” source
Claude Opus 4 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not determined	“for claude opus 4, our evaluations showed that, whereas we could conﬁdently rule out asl-4 capabilities, we could not conclusively rule out asl-3 capabilities—which led to our precautionary decision to deploy it with asl-3 protections .” source
Claude Opus 4 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not reached	“for claude opus 4, our evaluations showed that, whereas we could conﬁdently rule out asl-4 capabilities, we could not conclusively rule out asl-3 capabilities—which led to our precautionary decision to deploy it with asl-3 protections .” source
Claude Haiku 4.5 Anthropic	ASL-3 Autonomy and AI research	Not reached	“Our evaluation results showed that Claude Haiku 4.5 remained well below ASL-3 thresholds across all domains of concern.” source
Claude Haiku 4.5 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not reached	“Our evaluation results showed that Claude Haiku 4.5 remained well below ASL-3 thresholds across all domains of concern.” source

Previous · Page 4 of 4

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.

1957 of 1957

Model	Tester	Test	What was found
GPT-5.6 Terra OpenAI	Independent lab Independent lab · 2026-07-09	CyScenarioBench Cyber	“GPT-5.6 Sol solved 19/197 FrontierCyber challenges, 7/11 long-horizon CyScenarioBench challenges, and all 22 medium-andhard-difficultyAtomicchallenges.” source
GPT-5.6 Terra OpenAI	Independent lab Independent lab · 2026-07-09	Atomic Challenges suite Cyber	“Irregularevaluated GPT-5.6Solacrossthreeoffensivecybersecurityevaluationsuites: FrontierCyber, CyScenarioBench,andIrregular’sAtomicChallengessuite.” source
GPT-5.6 Terra OpenAI	Government evaluator Government evaluator · 2026-07-09	CTF expert-level tasks Cyber	“On AISI’sCTFexpert-leveltasks,GPT-5.6Solscored95.0%±9.8%,compared with85.0%±11.6%forGPT-5.5.” source
GPT-5.6 Terra OpenAI	Government evaluator Government evaluator · 2026-07-09	long-horizon cyber range tasks Cyber	“Onlong-horizoncyberrangetasks,GPT-5.6 SolcompletedTheLastOnes,a32-stepcorporate-networkattacksimulation, in 7 of 10 attempts, compared with 2 of 10 attempts for GPT-5.5.” source
GPT-5.6 Terra OpenAI	Independent lab Independent lab · 2026-07-09	METR time horizon 1.1 software-task suite Autonomy and AI research	“METR evaluated GPT-5.6 Sol on itsTime Horizon 1.1 software-task suite, whichmeasuresthelikelihoodthatamodelcanautonomouslycomplete softwaretasksthattakehumansagivenamountoftime.” source
GPT-5.6 Terra OpenAI	Government evaluator Government evaluator · 2026-07-09	unprompted sabotage evaluation Deception and sabotage	“In UK AISI’s unprompted sabotage evaluation, they found no confirmed instancesofGPT-5.6SolsabotagingAIsafetyresearch,andnoinstancesof GPT-5.6SolrefusingtoassistwithAIsafetyresearchtasks.” source
GPT-5.6 Terra OpenAI	Government evaluator Government evaluator · 2026-07-09	agentic cyber capability evaluations Cyber	“UKAISIseparatelyobservedattemptedcheatingbehaviorinagenticcyber capabilityevaluations.” source
GPT-5.6 Terra OpenAI	Government evaluator Government evaluator · 2026-07-09	impossible MLE bench Deception and sabotage	“Forexample,AISItestedGPT- 5.6Solontheirnewly-developedevaluationImpossibleMLEBench,which involves giving models ML tasks with instructions to aim for impossibly goodscorestoobservehowtheybehavewhenincentivisedtocheat.” source
GPT-5.6 Terra OpenAI	Government evaluator Government evaluator · 2026-07-09	benchmark adapted from an OpenAI paper in which models must answer a question whilst applying a constraint to their CoT reasoning Deception and sabotage	“UK AISI also tested whether models could intentionally shape their rea- soningtracestoavoiddetectionbyrunningabenchmarkadaptedfroman OpenAIpaperinwhichmodelsmustansweraquestionwhilstapplyinga constrainttotheirCoTreasoning.” source
GPT-5.6 Terra OpenAI	Government evaluator Government evaluator · 2026-07-09	no-CoT math reasoning time horizon Other	“Finally, UK AISI found that GPT-5.6 Sol performed at a 3.6 minute time horizonfor“no-CoT”mathreasoningtasks,wherethemodelmustreason withoutrelyingonanexternallyvisiblechain-of-thought.” source
GPT-5.6 Terra OpenAI	Independent lab Independent lab · 2026-07-09	sandbagging assessment with incentive to answer incorrectly Deception and sabotage	“WhenApollogave themodelanincentivetoanswerincorrectly,itsaccuracydidnotdrop.” source
GPT-5.6 Sol OpenAI	SecureBio Independent lab · 2026-07-09	External Evaluation for Bio Capabilities – SecureBio Chemical, biological, radiological, nuclear	“SecureBio, a nonprofit research organization focused on catastrophic bi- ological risk, evaluated two pre-release checkpoints of GPT-5.6 Sol and a railfreeversionofGPT-5.6Sol.” source
GPT-5.6 Sol OpenAI	SecureBio Independent lab · 2026-07-09	Virology Capabilities Test Chemical, biological, radiological, nuclear	“Itsstrongestreportedconfigurationsscored53.5%ontheVirology CapabilitiesTest,60.0%ontheMolecularBiologyCapabilitiesTest,68.4% ontheHumanPathogenCapabilitiesTest, and68.3%onWorld-ClassBio.” source
GPT-5.6 Sol OpenAI	Independent lab Independent lab · 2026-07-09	ReproBAIT Chemical, biological, radiological, nuclear	“On ReproBAIT, which measures AI agents on their ability to independently reproducebiologicalAImodels(BAIMs)frompublishedscientificpapers, therailfreevariantofGPT-5.6reached85%,comparedwith82%forGPT- 5.5.” source
GPT-5.6 Sol OpenAI	Independent lab Independent lab · 2026-07-09	ABC-Bench Chemical, biological, radiological, nuclear	“TherailfreecheckpointalsosetanewhighonABC-Bench’sAdvanced ScreeningEvasionandreliablyidentifiedaknownbutpracticallyinconve- nientmethodforevadingacommercialnucleic-acidscreeningalgorithm.” source
GPT-5.6 Sol OpenAI	Independent lab Independent lab · 2026-07-09	Manual agentic assessment Chemical, biological, radiological, nuclear	“In manual agentic assessment, the model was able to plan and execute a complex,long-runningscientifictask,includingmaintainingastructured workgraphandcoordinatingsubagents.” source
GPT-5.6 Sol OpenAI	Independent lab Independent lab · 2026-07-09	incentivized incorrect-answer sandbagging test Deception and sabotage	“WhenApollogave themodelanincentivetoanswerincorrectly,itsaccuracydidnotdrop.” source
GPT-5.6 Sol OpenAI	Government evaluator Government evaluator · 2026-07-09	UK AISI external testing of safeguards Misuse and jailbreaks	“UKAISIandOpenAIperformedseveralroundsofiterativetesting,withUK AISIreportingjailbreaksandOpenAIimprovingthesysteminresponseto theirfindings.” source
GPT-5.6 Luna OpenAI	Independent lab Independent lab · 2026-07-09	external evaluation for bio capabilities Chemical, biological, radiological, nuclear	“Overall, SecureBio found that either GPT-5.6 Sol or its railfree version achievedthehighestscorestodateonseveralexpert-levelbiologybench- marks.” source
GPT-5.6 Luna OpenAI	Independent lab Independent lab · 2026-07-09	sandbagging assessment Deception and sabotage	“ApollodidnotfindstandardSandbaggingQAbehavioronanotheroneof their sandbagging assessments.” source
Gemini 3.5 Flash Google DeepMind	Apollo Research Independent lab · 2026-07-07	judge calibration evaluation (Spearman's rank correlation and MAE against reference severity) Other	“Gemini 3.5 Flash performs at the level of the smaller models on MAE, reflecting its consistent pattern of under-scoring moderate and high severity violations.” source
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

Previous · Page 27 of 49 · Next