“Reasoning models are also the only way that model developers have been able to improve the efficacy of new models, using something called "test-time compute" to burn extra tokens to complete a task.”
Our reading, the sentence that was checked:
Model developers have only managed to make new models more effective through reasoning models that burn extra tokens via test-time compute.
Assessment: Contradicted Evidence as of
Where and when Ed Zitron wrote it
Where's Your Ed At, How To Argue With An AI Booster, paragraph 234.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment wyea-how-to-argue-with-an-ai-booster-8a4b43-a2
We recorded 35 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-08-25. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
From the speaker
- the only way
- model developers have been able to improve the efficacy of new models
- test-time compute
- burn extra tokens to complete a task
Our interpretation
- Read efficacy as measurable task performance, including accuracy, instruction following, and successful coding.
- The word only excludes meaningful improvements in new non-reasoning models achieved through training rather than additional inference-time reasoning.
- The retrospective phrase have been able to covers improvements achieved by publication. No starting date is supplied, so the target period has an open beginning; recent 2025 releases can test the assertion.
Counts as supported if: Available evaluations show that meaningful improvements in new models required additional inference-time reasoning, with no demonstrated improvements from non-reasoning model development.
Counts as contradicted if: At least one new non-reasoning model demonstrably improves on its predecessor through training or another method that does not require extra inference-time reasoning.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Contradicted
The dated evidence goes against what the claim committed to, under the criteria frozen before grading.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
On February 27, 2025, OpenAI attributed GPT-4.5's improvements over GPT-4o to scaling unsupervised learning [e1]. It reported GPQA scores of 71.4% for GPT-4.5 and 53.6% for GPT-4o [e1]. This company-reported alternative to additional inference-time reasoning predates the August 25, 2025 claim and contradicts its assertion that reasoning models were the only route to improved performance [e1].
Limitations
- The performance results and attribution to unsupervised scaling come from OpenAI; the packet provides no independent replication.
- The reported comparisons do not establish improvement on every task or superiority to the strongest reasoning models.
- The evidence does not isolate every training change through controlled ablations.
Evidence (1 item)
Goes against the claim: OpenAI,
scaling up unsupervised learning
Why it matters here: OpenAI explicitly describes GPT-4.5 as not thinking before responding and attributes gains to unsupervised scaling. Its reported comparisons include GPQA scores of 71.4% versus 53.6% for GPT-4o, providing a counterexample to the claim that additional inference-time reasoning was the only route to improvement.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Where's Your Ed At, How To Argue With An AI Booster
Reasoning models are also the only way that model developers have been able to improve the efficacy of new models, using something called "test-time compute" to burn extra tokens to complete a task.
- Where
- paragraph 234
- Published
- 2025-08-25
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment wyea-how-to-argue-with-an-ai-booster-8a4b43-a2: Current. Outcome: Contradicted. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.
About assessment wyea-how-to-argue-with-an-ai-booster-8a4b43-a2.