· Aug 7, 2025
METR reports its assessment of GPT-5’s catastrophic-risk capabilities
Markdown · JSON · Original source
published this evaluation on August 7, 2025. It argues that GPT-5 is unlikely to pose catastrophic risk through AI R&D automation, rogue replication or strategic sabotage, because its measured capabilities fall well short of what METR considers necessary for those threats. METR estimated GPT-5’s 50% at about 2 hours and 17 minutes on its task suite, while stressing that translating this measure into real-world usefulness is very uncertain. Its conclusion also relies on assurances from about the model and its training, and on access to and background information. METR reports that GPT-5 sometimes recognized it was being evaluated, but says it found no evidence that strategic sabotage significantly influenced the results. The stakes are not limited to this assessment: METR warns that more capable systems could become able to undermine evaluations, making confident risk assessments harder. The evaluation did not assess alignment or several other risks, and METR says its methods would not support a confident assessment if GPT-5 were significantly more capable. OpenAI’s communications and legal teams reviewed and approved the post under a standard NDA; METR says this should not be interpreted as robust formal oversight or public accountability.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Aug 7, 2025 (2025-08-07T07:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-12T20:28:23.684Z. First observed: 2026-10-18T03:25:20.834Z. Retrieved: 2026-10-23T22:41:48.302Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on October 23, 2026; originally published August 7, 2025. says its estimates are a rough proxy and that their relation to real-world usefulness and the threat models is very uncertain. The evaluation did not assess alignment or risks including misuse, CBRN capabilities, persuasive abilities or widespread societal use. METR says ’s communications and legal team required review and approval of the post under a standard NDA, and that OpenAI had the legal right to block sharing conclusions depending on non-public information. METR says it did not change conclusions, takeaways or tone based on that review.
Open question
How well do ’s task-based estimates predict the capabilities that matter for these threat models in real-world settings?
Source-stated
METR concludes GPT-5 and incremental development beyond it are unlikely to pose catastrophic risk through the three threat models it assessed.
we conclude that GPT-5 (and incremental development4 beyond GPT-5) is unlikely to pose a catastrophic risk via AI R&D automation, rogue replication, or sabotage threat models.
Source-stated
METR estimates GPT-5’s 50% time horizon at about 2 hours and 17 minutes on its task suite.
The observed 50%-time horizon of GPT-5 was 2h17m (65m - 4h25m 95% CI)
Source-stated
METR says the connection between its time horizon estimates and real-world usefulness is very uncertain.
The link between our broad time horizon estimates and the particular threat models is very uncertain
Source-stated
METR says its assessment relied in part on OpenAI’s assurances, reasoning-trace access and background information.
The access to reasoning traces and background information formed an important part of our assessment
Source-stated
METR reports evidence of situational awareness but says it found no evidence that strategic sabotage significantly influenced its results.
We investigated the possibility that GPT-5 was engaging in strategic sabotage of our evaluation, and concluded that it is unlikely that strategic sabotage has significantly influenced our results.
Source-stated
METR says the evaluation did not assess alignment or several other risks.
We did not assess the alignment of OpenAI GPT-5, nor other risks (including from misuse, CBRN capabilities, persuasive abilities, widespread societal use etc.).
Source-stated
METR says the report should not be treated as robust formal oversight or public accountability, given OpenAI’s legal right to block publication of conclusions depending on non-public information.
this evaluation shouldn’t be interpreted as robust formal oversight or accountability that the public can be relying on METR to provide.
Source and contributor context
METR · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Time horizon
METR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Chain of thought
The step-by-step working a model writes out before giving its final answer.
METR
METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
OpenAI
OpenAI develops AI models and products and publishes research, release announcements and evaluations.
Editorial provenance
Version 1. Publication ID: 3b20c196-4c3f-dbe0-acee-293fc8510f1c.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-12T20:23:49.409Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: bf5d6fe88b9e69706faa8d6b7f92a013e097a76d8f49d32276b31b7f0d93dbb8. Source text hash: 780276b2b04653e463aafb427f8774ae41fa29be2ad5733f67bdcf4239d53b7c.
Revision and source-check history
Version 1 · 2026-11-12T20:28:23.684Z
METR reports its assessment of GPT-5’s catastrophic-risk capabilities
published this evaluation on August 7, 2025. It argues that GPT-5 is unlikely to pose catastrophic risk through AI R&D automation, rogue replication or strategic sabotage, because its measured capabilities fall well short of what METR considers necessary for those threats. METR estimated GPT-5’s 50% at about 2 hours and 17 minutes on its task suite, while stressing that translating this measure into real-world usefulness is very uncertain. Its conclusion also relies on assurances from about the model and its training, and on access to and background information. METR reports that GPT-5 sometimes recognized it was being evaluated, but says it found no evidence that strategic sabotage significantly influenced the results. The stakes are not limited to this assessment: METR warns that more capable systems could become able to undermine evaluations, making confident risk assessments harder. The evaluation did not assess alignment or several other risks, and METR says its methods would not support a confident assessment if GPT-5 were significantly more capable. OpenAI’s communications and legal teams reviewed and approved the post under a standard NDA; METR says this should not be interpreted as robust formal oversight or public accountability.
Original source
· Article
Source details
Original source date: Aug 7, 2025 (2025-08-07T07:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-12T20:28:23.684Z. First observed: 2026-10-18T03:25:20.834Z. Retrieved: 2026-10-23T22:41:48.302Z.
Limits
Text as retrieved on October 23, 2026; originally published August 7, 2025. says its estimates are a rough proxy and that their relation to real-world usefulness and the threat models is very uncertain. The evaluation did not assess alignment or risks including misuse, CBRN capabilities, persuasive abilities or widespread societal use. METR says ’s communications and legal team required review and approval of the post under a standard NDA, and that OpenAI had the legal right to block sharing conclusions depending on non-public information. METR says it did not change conclusions, takeaways or tone based on that review.
Open question
How well do ’s task-based estimates predict the capabilities that matter for these threat models in real-world settings?
Source-stated
METR concludes GPT-5 and incremental development beyond it are unlikely to pose catastrophic risk through the three threat models it assessed.
we conclude that GPT-5 (and incremental development4 beyond GPT-5) is unlikely to pose a catastrophic risk via AI R&D automation, rogue replication, or sabotage threat models.
Source-stated
METR estimates GPT-5’s 50% time horizon at about 2 hours and 17 minutes on its task suite.
The observed 50%-time horizon of GPT-5 was 2h17m (65m - 4h25m 95% CI)
Source-stated
METR says the connection between its time horizon estimates and real-world usefulness is very uncertain.
The link between our broad time horizon estimates and the particular threat models is very uncertain
Source-stated
METR says its assessment relied in part on OpenAI’s assurances, reasoning-trace access and background information.
The access to reasoning traces and background information formed an important part of our assessment
Source-stated
METR reports evidence of situational awareness but says it found no evidence that strategic sabotage significantly influenced its results.
We investigated the possibility that GPT-5 was engaging in strategic sabotage of our evaluation, and concluded that it is unlikely that strategic sabotage has significantly influenced our results.
Source-stated
METR says the evaluation did not assess alignment or several other risks.
We did not assess the alignment of OpenAI GPT-5, nor other risks (including from misuse, CBRN capabilities, persuasive abilities, widespread societal use etc.).
Source-stated
METR says the report should not be treated as robust formal oversight or public accountability, given OpenAI’s legal right to block publication of conclusions depending on non-public information.
this evaluation shouldn’t be interpreted as robust formal oversight or accountability that the public can be relying on METR to provide.
Source and contributor context
METR · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Time horizon
METR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Chain of thought
The step-by-step working a model writes out before giving its final answer.
METR
METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
OpenAI
OpenAI develops AI models and products and publishes research, release announcements and evaluations.
Editorial provenance
Version 1. Publication ID: 3b20c196-4c3f-dbe0-acee-293fc8510f1c.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-12T20:23:49.409Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: bf5d6fe88b9e69706faa8d6b7f92a013e097a76d8f49d32276b31b7f0d93dbb8. Source text hash: 780276b2b04653e463aafb427f8774ae41fa29be2ad5733f67bdcf4239d53b7c.
uncertain · 2026-11-12T20:23:49.409Z · anthropic/claude-opus-5.5
Publication: 3b20c196-4c3f-dbe0-acee-293fc8510f1c
attribution — pass: This is a single METR evaluation report on GPT-5. The reading attributes its conclusions, its estimates and its caveats about OpenAI review to METR, and attributes the assurances to OpenAI.
METR assessed whether OpenAI GPT-5 could pose catastrophic risks, considering three key threat models
summary — uncertain: The summary is faithful and keeps METR's caveats and its independence note. However, the three central threat models are named but never explained in plain language. 'Rogue replication' especially means nothing to a non-specialist; the source describes it as AI maintaining infrastructure, acquiring resources and evading shutdown. Smallest repair: add one short plain-language gloss for each threat model, or add term-library entries for them. 'Time horizon' is covered by definitions.
AI systems posing direct risks of rapid takeover, which likely requires them to be able to maintain infrastructure, acquire resources, and evade shutdown
date — pass: originalPublished is 2025-08-07 (instant), and the reading states August 7, 2025, which matches the UTC day. The canary timestamp of Aug 6, 2025 is consistent with this. The reading uses no later events or hindsight framing. Kind is publisher-page, so no archive disclosure is needed.
support — pass: All claims are attributed to METR and appear in the source. Claim 3 says the link to 'real-world usefulness' is very uncertain, but its excerpt concerns the threat models. The same source sentence does mention real-world usefulness, so the claim is supported. Still, extending the excerpt would align it better. No benefits, costs or forecasts are invented.
The link between our broad time horizon estimates and the particular threat models is very uncertain, and the rough estimates here are likely to evolve as we get a better understanding of how time horizon results translate to real world usefulness
media — pass: No media was selected, and no verified candidates were supplied, so nothing was misidentified.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceOpenAI develops AI models and products and publishes research, release announcements and evaluations.
Explore OpenAIThe step-by-step working a model writes out before giving its final answer.
Definition referenceMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceOpenAI develops AI models and products and publishes research, release announcements and evaluations.
Explore OpenAIMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceOpenAI develops AI models and products and publishes research, release announcements and evaluations.
Explore OpenAIThe step-by-step working a model writes out before giving its final answer.
Definition referenceMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceOpenAI develops AI models and products and publishes research, release announcements and evaluations.
Explore OpenAIMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition reference