“DeepSeek r1 is not smarter than earlier models, just trained more cheaply”
Our reading, the sentence that was checked:
DeepSeek r1 was trained more cheaply than earlier models.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Five things most people don't seem to understand about DeepSeek, paragraph 1.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-five-things-most-people-dont-seem-103fa6-2-a1
We recorded 9 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 2025-01-28. 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
- DeepSeek r1
- trained more cheaply
- than earlier models
Our interpretation
- The ordinary reading is lower training cost for a complete, comparably capable model, including its base-model training and subsequent reasoning training.
- This does not require counting all corporate investment as training cost. Any inclusion of experiments, personnel or data expenses must be consistent across models.
- The observation concerns the training that produced the original January 20 release. Its starting date is unspecified.
- Marcus cites no cost source or accounting basis in the supplied text, so accounting_basis is unknown.
Counts as supported if: Comparable cost records establish that producing the complete R1 model cost less than producing comparably capable preceding frontier models.
Counts as contradicted if: Comparable cost records establish that producing R1 cost as much as or more than the relevant preceding models.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
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
The disclosures make lower training cost plausible, including a later $294,000 estimate for the R1 post-training project. But that figure excludes base training, and the V3 figure measures a different model's training. Neither establishes the complete comparison Marcus asserted. His accounting basis and comparator are also unspecified.
Limitations
- Unresolved does not mean the cheaper-training claim is false.
- Rental-equivalent compute estimates are not audited total expenditure.
- The later R1 disclosure was considered; it was not excluded merely because it appeared after the post.
Evidence (2 items)
Context only: DeepSeek-AI / arXiv,
Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576M.
Why it matters here: This is a rental-equivalent compute estimate for V3's official training, excluding earlier research and ablation experiments. It is not the complete cost of training R1.
- Publisher
- DeepSeek-AI / arXiv
- Kind of source
- Preprint, not peer reviewed
- Measures
- DeepSeek, DeepSeek-V3 official training compute cost at assumed GPU rental rates, USD millions of dollars, company reported, ? to 2024-12
- Describes events of
- Public from
Context only: DeepSeek-AI / Nature supplementary information,
For the training of DeepSeek-R1-Zero, we employed 64*8 H800 GPUs, and the process required approximately 198 hours.
Why it matters here: The accompanying table reports $294,000 at assumed rental rates for R1-Zero, supervised-data creation and R1 training. This is useful evidence of inexpensive post-training, but excludes the base model and does not provide a matched predecessor comparison.
- Publisher
- DeepSeek-AI / Nature supplementary information
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Measures
- DeepSeek, Original R1 project's post-training compute cost at assumed GPU rental rates, USD dollars, company reported, ? to 2025-01
- Describes events of
- Public from
What would settle it
A specified comparator and accounting basis, with complete base-training and post-training costs for both models measured consistently.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Five things most people don't seem to understand about DeepSeek
DeepSeek r1 is not smarter than earlier models, just trained more cheaply
- Where
- paragraph 1
- Published
- 2025-01-28
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-five-things-most-people-dont-seem-103fa6-2-a1: Current. Outcome: Unresolved. 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 marcus-five-things-most-people-dont-seem-103fa6-2-a1.