Scores by kind of claim
Each kind of claim is scored on its own, out of 100: the share that held up, with the range it likely falls in. Fewer resolved claims give a wider range.
-
Factual claims
Statements about the present or the past, checked against the record: did they hold up?
79 72 85 -
Predictions
Statements about events outside the person's control, checked once they fell due: did they come true?
80 72 88 -
Promises
Commitments to the person's or organisation's own future action, with something to deliver: was it delivered?
Insufficient record: 1 promise resolved, 5 needed for a score.
All types combined: 79 (likely 74 to 84, high confidence), from 102 claims: 64 factual claims, 38 predictions. It mixes kinds of claim in different proportions for different people, so compare people kind by kind.
Gary Marcus's claims
“GPT-5 may be more satisfying but is likely even more expensive to train and operate.”
Gary Marcus, author, Marcus on AI 31 March 2024 When Will the GenAI Bubble Burst? (Marcus on AI)
GPT-5 is likely to be even more expensive to train and operate than GPT-4.
Mostly true True in substance, wrong in a detail.
The second judge would not resolve this claim; the verdict shown is the first judge's.
What would count as true or false
The claim is that GPT-5 would cost more than GPT-4 both to train and to operate. Judge it at GPT-5's release and the months after. Training cost can be shown by compute or dollar estimates from OpenAI or credible third parties (for example Epoch AI or major press reporting). Operating cost can be shown by OpenAI's per-token API prices for comparable tiers, or by reported inference cost per query or in total. Compare against GPT-4 at launch. True: credible evidence shows GPT-5 cost more than GPT-4 to train, and also more to operate. Mostly true: training clearly cost more, but the operating-cost evidence is mixed. For example, list prices fell while total or per-query inference compute rose. Mixed: one of the two (training or operating) was clearly more expensive and the other was clearly cheaper. Mostly false: training cost was about the same or unclear, and operating cost was lower. False: credible evidence shows GPT-5 was no more expensive than GPT-4 to train and was cheaper to operate. Lower per-token API prices for GPT-5 than for GPT-4 at launch count as evidence of cheaper operation unless credible reports show higher underlying serving costs.
Evidence
- Notes on GPT-5 training compute Epoch AI (Substack) · 13 October 2025
“Our best guess: GPT-5 was trained on ~5e25 FLOP total, including both pre-training and reinforcement learning. That would be more than twice as much as GPT-4 (~2e25 FLOP), but less than GPT-4.5 (>1e26 FLOP).”
- Introducing GPT-5 for developers OpenAI · 7 August 2025
“GPT‑5 is priced at $1.25/1M input tokens and $10/1M output tokens”
- Why GPT-5 used less training compute than GPT-4.5 (but GPT-6 probably won’t) Epoch AI · 26 September 2025
“It’s likely that the total compute for developing GPT-5 was higher than for GPT-4.5, if we also account for the compute for running experiments.”
- Why GPT-5 used less training compute than GPT-4.5 (but GPT-6 probably won’t) Epoch AI · 26 September 2025
“One more reason that OpenAI likely chose to focus on scaling post-training without increasing pre-training compute for GPT-5 is to reduce inference costs.”
Reasoning
Epoch AI's best estimate is that GPT-5 was trained on about 5e25 FLOP in total, more than twice GPT-4's roughly 2e25 FLOP [E1], so independent evidence shows training cost more. The operating-cost evidence does not settle the question. E2 gives GPT-5's API prices ($1.25 per 1M input tokens and $10 per 1M output tokens), but no evidence item gives GPT-4's launch prices to compare against. E4 says OpenAI may have shaped GPT-5's training partly to reduce inference costs, but it gives no measured cost figure and does not compare against GPT-4. Since training clearly cost more and the operating-cost side is unclear, the criteria point to mostly true.
Cite this claim
Gary Marcus, “GPT-5 is likely to be even more expensive to train and operate than GPT-4.” (said 31 March 2024). Verdict: mostly true. Reliability API, claim c_d044b3e8aeaa, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_d044b3e8aeaa
Claims without a verdict
Is a verdict here wrong, or do you speak for Gary Marcus? Reply on the claim's own page, with the form at its foot (how replies work). Verdicts are made automatically; how verdicts and scores are made.