Anthropic’s Claude completes a 16-hour human research job in under 3 hours
Quick Take
- Anthropic's Claude Opus 5.5 completed a research assignment in under three hours that typically takes human contractors 16 hours.
- Epoch AI found the model delivered contractor-level work at a lower cost, raising questions about the future of outsourced research jobs.
- Epoch is considering shifting contractors toward reviewing AI-generated work, although errors still require human intervention.
Anthropic's Claude Opus 5.5 completed a research assignment in under three hours that typically takes human contractors 16 hours, according to Epoch AI.
The AI research organization reported on Nov. 17 that the model produced work comparable to that of its professional contractors, raising the possibility of automating parts of its research operations.
The assignment involved researching EcoDataCenter 1 for Epoch's data center database, a process requiring researchers to examine news reports, regulatory filings, and satellite imagery to establish details about the facility.
Epoch provided the AI models with its contractor guidelines, research tools, and evaluation criteria before assessing their work against the standards applied to human researchers.
Claude Opus 5.5 scored 87.5% on the assignment, completing it in just under three hours at a lower cost than hiring a contractor. Another Anthropic model, Claude Fable 5.2, delivered an even stronger result, scoring 97.5%. However, it required approximately six and a half hours and cost more than a human contractor.
Both models met Epoch's standards for core research, although their submissions still required revisions to equipment classifications, image annotations, and written descriptions.
Epoch’s data center test
The higher-scoring Claude run cost more than a contractor in Epoch’s comparison.
Epoch’s analyst judged both first passes comparable to contractor work and identified revisions for each.
The 16-hour figure is typical contractor labor per site. Epoch leaves review time, time to publication and run-specific imagery waits unspecified. Its cost comparison gives the ranking without itemizing expenses.
Sources
Supporting facts
- Reported statement In its Nov. 17 evaluation, Epoch AI said one of its data center analysts judged outputs from Claude Opus 5.5 and Claude Fable 5.2 on the EcoDataCenter 1 assignment comparable to work Epoch receives from contractors, whose submissions also typically need revisions before publication. [1] [2]
- Reported statement Epoch AI reported that Claude Opus 5.5 completed the EcoDataCenter 1 task in just under three hours at lower cost than a contractor. Epoch says a contractor typically spends about 16 hours per data center. The contractor figure describes typical labor per site. [1]
- Reported statement Epoch AI reported about 6.5 hours for Claude Fable 5.2 on the EcoDataCenter 1 task and said it cost more than a contractor while producing a higher-scoring result than Claude Opus 5.5. The report gives the relative cost ranking without a line-item breakdown or a measure of the human review needed to publish. [1]
- Reported statement Epoch AI gave Claude Opus 5.5 a rubric score of 87.5% and Claude Fable 5.2 a score of 97.5% on the EcoDataCenter 1 task, displayed as 88% and 98% in the benchmark summary. Epoch said a perfect score would mean the work could be published without changes. [1] [2]
- Reported statement Epoch AI said Opus 5.5 needed revisions to equipment classification and more substantial changes to image annotations, while Fable 5.2 had imperfect annotations and overly long text. Epoch said both met its standard on the core research elements. [1]
- Reported statement Epoch AI says it is considering moving contractor time toward reviewing AI outputs and other data center analysis work. [1]
Method
The timeline uses a linear hour scale. An open marker shows that the Opus run finished just under three hours. Fable and contractor values are approximate. The scores and ordinal cost comparisons come from Epoch AI’s Nov. 17 evaluation.
Scope
The 16-hour figure is typical contractor labor per site. Epoch leaves review time, time to publication and run-specific imagery waits unspecified. Its cost comparison gives the ranking without itemizing expenses.
CryptoSlateBy comparison, OpenAI's GPT-6.2 Astra scored 65% on the same assignment, missing important information and estimates that Epoch said the available evidence could have supported.
AI could change how research contractors work
The findings have prompted Epoch to reconsider how it allocates work between AI models and human contractors.
The organization said it is considering shifting some contractor time away from gathering and compiling research toward reviewing AI-generated submissions and performing other data center analysis.
Such a transition could let AI systems handle much of the initial research while human workers focus on checking findings, correcting errors, and preparing the material for publication.
However, the results also exposed differences in the economics of using AI for professional work.
While Opus completed the assignment faster and cheaper than a contractor, Fable's higher-quality output cost more, despite also finishing considerably faster than a human researcher.
The models also struggled with parts of the assignment that required visual interpretation, particularly labeling satellite images and identifying equipment.
These shortcomings help explain why Epoch has not proposed eliminating human involvement entirely, even as its models approach contractor-level performance.
Instead, the organization is considering a workflow in which AI produces the initial research and contractors refine the results, potentially reducing the human labor required for each database entry.
Mentioned in this article
Editorial credits
Oluwapelumi values Bitcoin's potential. He imparts insights on a range of topics like DeFi, hacks, mining and culture, underlining transformative power.