“And the reason that those tend to be more expensive is that they tend (with the help of some good old fashioned AI that nobody wants to talk about) to run multiple large language models in the background many times until they get the answers that by some criteria seem to correct.”
Our reading, the sentence that was checked:
AI agents tend to be more expensive because they tend to run multiple large language models in the background many times until they get answers that by some criteria seem correct.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Breaking news, and how the end might begin, paragraph 36.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-breaking-news-and-how-the-end-might-469c5d-a1
We recorded 22 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 2026-06-10. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Causal claim
- 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)
Conditions
- tend to (qualifier, stated by the speaker)
- by some criteria (qualifier, stated by the speaker)
From the speaker
- agents tend to be more expensive
- run multiple large language models in the background many times
- until they get the answers that by some criteria seem to correct
- with the help of some good old fashioned AI
Our interpretation
- Plain reading: agents cost more than single chat queries mainly because they make many repeated model calls (often several models or sub-agents in loops) and check results before finishing, which uses far more tokens.
- The separate hyperbole about 'a million times or billion times' is outside the exact words being graded.
- The claim describes current practice. I read it as agent systems from 2025 up to the post date.
Counts as supported if: Credible technical or provider sources show that agents make many repeated or parallel model calls, often several models or sub-agents with checking steps, and that this repeated use drives much higher token use and cost than chat.
Counts as contradicted if: Credible sources show that agents' higher cost comes mainly from something other than repeated model calls (for example, higher per-token prices alone), or that agents do not typically use many more model calls or tokens than chat.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
Outcome: Supported
The dated evidence matches 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
Anthropic's engineering write-up says agents use about 4 times more tokens than chats, and multi-agent systems about 15 times more. It describes a lead agent spawning sub-agents that search repeatedly, then evaluate quality and refine their queries. That matches Marcus's mechanism: several model instances run in the background many times, checking results, which drives cost. GitHub's move to usage-based billing was explained by agent sessions costing far more to serve than chat queries. The core causal claim holds.
Limitations
- The main evidence is one provider's description of its own system.
- Some agents use a single model in a loop rather than 'multiple' models.
- The GitHub reasoning comes from trade press coverage of the company's announcement.
Evidence (4 items)
Supports the claim: Anthropic,
In our data, agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats.
Why it matters here: Provider data shows agents, and especially multi-agent systems, use several times more tokens than chat, which drives cost.
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Supports the claim: Anthropic,
the lead agent analyzes it, develops a strategy, and spawns subagents to explore different aspects simultaneously. As shown in the diagram above, the subagents act as intelligent filters by iteratively using search tools
Why it matters here: Describes the mechanism Marcus named: several model instances run in the background, repeatedly, in a coordinated loop.
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Supports the claim: Anthropic,
Subagents also plan, then use interleaved thinking after tool results to evaluate quality, identify gaps, and refine their next query.
Why it matters here: Shows agents repeatedly check and refine their outputs, matching 'until they get the answers that by some criteria seem correct'.
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Supports the claim: ITdaily,
a short chat query and an autonomous coding session of several hours cost the end user the same, while the actual cost differs significantly.
Why it matters here: Reports GitHub's reasoning that long agent sessions cost far more to serve than chat queries.
- Where
- paragraph 3
- Publisher
- ITdaily
- Kind of source
- News report
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Breaking news, and how the end might begin
And the reason that those tend to be more expensive is that they tend (with the help of some good old fashioned AI that nobody wants to talk about) to run multiple large language models in the background many times until they get the answers that by some criteria seem to correct.
- Where
- paragraph 36
- Published
- 2026-06-10
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-breaking-news-and-how-the-end-might-469c5d-a1: Current. Outcome: Supported. 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-breaking-news-and-how-the-end-might-469c5d-a1.