· Jul 9, 2025
xAI presents Grok 4 as a model trained with scaled reinforcement learning
Markdown · JSON · Original source
’s page, published July 9, 2025, presents 4 as the result of scaling . xAI says it used its Colossus cluster and expanded verifiable training data beyond math and coding. It reports smooth performance gains while training with more compute than it had used previously. xAI also says Grok 4 was trained to use tools, including web browsing, and highlights Grok 4 Heavy’s result on Humanity’s Last Exam. The page’s stated next step is to extend reinforcement learning from controlled domains to complex real-world problems, where models could learn and adapt in dynamic environments. The page reports xAI’s training methods and results, but does not establish how those results translate to real-world reliability. Whether the planned expansion helps models learn and adapt in dynamic environments remains unknown.
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: Jul 9, 2025 (2025-07-09T00:00:00Z).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-11-13T04:38:24.336Z. First observed: 2026-10-18T15:33:22.036Z. Retrieved: 2026-10-24T12:19:46.189Z.
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 24, 2026; originally published July 9, 2025. The page reports ’s training methods and results but does not establish how those results translate to real-world reliability.
Open question
How will evaluate whether in complex real-world environments helps models learn and adapt, beyond performance on controlled ?
Source-stated
xAI says it used its Colossus cluster to scale reinforcement learning training for Grok 4 and broadened its verifiable training data beyond math and coding.
For Grok 4, we utilized Colossus, our 200,000 GPU cluster, to run reinforcement learning training that refines Grok's reasoning abilities at pretraining scale.
Source-stated
xAI reports smooth performance gains from training with substantially more compute than it had used previously.
The resulting training run saw smooth performance gains while training on over an order of magnitude more compute than had been used previously.
Source-stated
xAI says Grok 4 was trained to use tools such as a code interpreter and web browsing.
Grok 4 was trained with reinforcement learning to use tools. This allows Grok to augment its thinking with tools like a code interpreter and web browsing
Source-stated
xAI reports that Grok 4 Heavy scored 50% on Humanity's Last Exam.
Grok 4 Heavy saturates most academic benchmarks and is the first model to score 50% on Humanity's Last Exam
Source-stated
xAI plans to extend reinforcement learning from controlled domains to complex real-world problems.
We plan to expand the scope from verifiable rewards in controlled domains to tackling complex real-world problems, where models can learn and adapt in dynamic environments.
Source and contributor context
xAI · 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.
Reinforcement learning
Training by trial and error, in which a system is rewarded for the outcomes its designers want.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
xAI
xAI develops the Grok family of AI models and publishes its own release claims and evaluations.
Grok
Grok is a family of AI models from xAI, with capability claims and evaluations published by its developer.
Editorial provenance
Version 1. Publication ID: 074e81aa-2f83-e868-75b5-56eb1c81cd16.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-13T04:33:45.396Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: b08f23cdc95294bde66f02447c3bc933d5648ad5d8bb9bd729baa3fc91fb73e7. Source text hash: b429a84bfde6a67edf01b3990810616d021ac3b1d7cb97c0c84db783f90de5e4.
Revision and source-check history
Version 1 · 2026-11-13T04:38:24.336Z
xAI presents Grok 4 as a model trained with scaled reinforcement learning
’s page, published July 9, 2025, presents 4 as the result of scaling . xAI says it used its Colossus cluster and expanded verifiable training data beyond math and coding. It reports smooth performance gains while training with more compute than it had used previously. xAI also says Grok 4 was trained to use tools, including web browsing, and highlights Grok 4 Heavy’s result on Humanity’s Last Exam. The page’s stated next step is to extend reinforcement learning from controlled domains to complex real-world problems, where models could learn and adapt in dynamic environments. The page reports xAI’s training methods and results, but does not establish how those results translate to real-world reliability. Whether the planned expansion helps models learn and adapt in dynamic environments remains unknown.
Original source
· Article
Source details
Original source date: Jul 9, 2025 (2025-07-09T00:00:00Z).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-11-13T04:38:24.336Z. First observed: 2026-10-18T15:33:22.036Z. Retrieved: 2026-10-24T12:19:46.189Z.
Limits
Text as retrieved on October 24, 2026; originally published July 9, 2025. The page reports ’s training methods and results but does not establish how those results translate to real-world reliability.
Open question
How will evaluate whether in complex real-world environments helps models learn and adapt, beyond performance on controlled ?
Source-stated
xAI says it used its Colossus cluster to scale reinforcement learning training for Grok 4 and broadened its verifiable training data beyond math and coding.
For Grok 4, we utilized Colossus, our 200,000 GPU cluster, to run reinforcement learning training that refines Grok's reasoning abilities at pretraining scale.
Source-stated
xAI reports smooth performance gains from training with substantially more compute than it had used previously.
The resulting training run saw smooth performance gains while training on over an order of magnitude more compute than had been used previously.
Source-stated
xAI says Grok 4 was trained to use tools such as a code interpreter and web browsing.
Grok 4 was trained with reinforcement learning to use tools. This allows Grok to augment its thinking with tools like a code interpreter and web browsing
Source-stated
xAI reports that Grok 4 Heavy scored 50% on Humanity's Last Exam.
Grok 4 Heavy saturates most academic benchmarks and is the first model to score 50% on Humanity's Last Exam
Source-stated
xAI plans to extend reinforcement learning from controlled domains to complex real-world problems.
We plan to expand the scope from verifiable rewards in controlled domains to tackling complex real-world problems, where models can learn and adapt in dynamic environments.
Source and contributor context
xAI · 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.
Reinforcement learning
Training by trial and error, in which a system is rewarded for the outcomes its designers want.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
xAI
xAI develops the Grok family of AI models and publishes its own release claims and evaluations.
Grok
Grok is a family of AI models from xAI, with capability claims and evaluations published by its developer.
Editorial provenance
Version 1. Publication ID: 074e81aa-2f83-e868-75b5-56eb1c81cd16.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-13T04:33:45.396Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: b08f23cdc95294bde66f02447c3bc933d5648ad5d8bb9bd729baa3fc91fb73e7. Source text hash: b429a84bfde6a67edf01b3990810616d021ac3b1d7cb97c0c84db783f90de5e4.
passed · 2026-11-13T04:33:45.396Z · anthropic/claude-opus-5.5
Publication: 074e81aa-2f83-e868-75b5-56eb1c81cd16
attribution — pass: This is a single publisher announcement page from xAI about Grok 4, not a listing. The reading consistently attributes the training methods, results and plans to xAI ('xAI says', 'xAI reports', 'xAI plans').
For Grok 4, we utilized Colossus, our 200,000 GPU cluster
summary — pass: The reading leads with xAI's central argument: Grok 4 comes from scaling reinforcement learning. It then covers tool use, the Humanity's Last Exam result and the stated next step. Reinforcement learning and benchmark are defined beside the reading, and the next sentence frames Humanity's Last Exam as a benchmark result. The question follows from the source's own plan without asserting an answer. 'Verifiable training data' could use a term-library tooltip, but it does not block understanding.
We plan to expand the scope from verifiable rewards in controlled domains to tackling complex real-world problems
date — pass: originalPublished is 2025-07-09T00:00:00Z, and the reading states July 9, 2025, which matches the UTC day. The July 2025 dates in the demo transcript are not later than publication. The plans are reported as xAI's plans, not outcomes. The page is a publisher page, so no archive disclosure is needed.
support — pass: Each claim matches its excerpt. 'Substantially more compute' understates rather than exaggerates 'over an order of magnitude'. The 50% Humanity's Last Exam claim is attributed to xAI. The limitation and the closing unknown appropriately avoid turning benchmark scores into real-world reliability.
The resulting training run saw smooth performance gains while training on over an order of magnitude more compute than had been used previously.
media — pass: No media candidates were supplied and none were selected; mediaId is null, which is consistent.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
xAI develops the Grok family of AI models and publishes its own release claims and evaluations.
Explore xAIxAI develops the Grok family of AI models and publishes its own release claims and evaluations.
Explore xAIGrok is a family of AI models from xAI, with capability claims and evaluations published by its developer.
Explore GrokTraining by trial and error, in which a system is rewarded for the outcomes its designers want.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencexAI develops the Grok family of AI models and publishes its own release claims and evaluations.
Explore xAIxAI develops the Grok family of AI models and publishes its own release claims and evaluations.
Explore xAIA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencexAI develops the Grok family of AI models and publishes its own release claims and evaluations.
Explore xAITraining by trial and error, in which a system is rewarded for the outcomes its designers want.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencexAI develops the Grok family of AI models and publishes its own release claims and evaluations.
Explore xAIGrok is a family of AI models from xAI, with capability claims and evaluations published by its developer.
Explore GrokTraining by trial and error, in which a system is rewarded for the outcomes its designers want.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencexAI develops the Grok family of AI models and publishes its own release claims and evaluations.
Explore xAIxAI develops the Grok family of AI models and publishes its own release claims and evaluations.
Explore xAIA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencexAI develops the Grok family of AI models and publishes its own release claims and evaluations.
Explore xAITraining by trial and error, in which a system is rewarded for the outcomes its designers want.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference