Anthropic filed its IPO registration on October 26, 2026. The filing states its belief that knowledge workers’ work represented more than $40 trillion in economic value globally in 2025. It defines cognitive work as the application of knowledge, reasoning, and judgment to produce economic value. (Anthropic Form S-1, filed October 26, 2026)
On November 9, Anthropic announced Claude Fable 5.2. In the announcement’s benchmark table, the model scored 41.3 percent on the Remote Labor Index, a test built from real freelance projects. The prior Fable generation scored 17.9 percent. Claude Opus 5.5, Anthropic’s recommended daily driver, scored 21.3 percent. (Anthropic, Introducing Claude Fable 5.2 and Claude Mythos 5.2)
Read fast, those numbers sound like a forecast: 40 percent of office work, done by AI, today. They are not a forecast. They are scores on fixed tests that disagree. Capability is moving fast. Control over what the capability touches is the decision that matters now.
What the $40 trillion figure is
The $40 trillion figure is an estimate of value, not a market waiting to be captured. The filing frames it as a belief about work performed in 2025. It is neither Anthropic revenue nor a forecast of AI spending. The filing points readers to its Industry, Market, and Other Data section for how the figure is calculated. (Anthropic Form S-1)
Anthropic opened its IPO roadshow on November 10. Its IPO site posts a roadshow video and states the expected timetable: a final price on November 18 and trading on November 19 under the ticker ANT. (Anthropic IPO site) Cognitive work is large, and AI systems can now do parts of it at a professional standard.
What the benchmark actually tests
The Remote Labor Index comes from Scale AI and the Center for AI Safety. Their October 28, 2025 paper built it from 240 real freelance projects from 358 verified Upwork professionals, across 23 work domains. (Scale AI and Center for AI Safety, Remote Labor Index paper)
Each project gives the agent a brief, input files, and the human deliverable as the standard. Trained experts judge whether a reasonable client would accept it. The automation rate is the share judged at least as good as the human standard. Three evaluators vote on each project. (Scale Labs, Remote Labor Index leaderboard)
The benchmark states its limits. It excludes work that requires direct client interaction, physical labor, or long-term evaluation. It tests a finished deliverable produced alone from a written brief.
The movement is real. At launch in October 2025, the best agent scored 2.5 percent. Fable 5.2 now scores 41.3 percent at High effort. That arc spans different models from different labs over just over a year. The 20-to-40 comparison pits two Anthropic model lines in one table: Opus 5.5, the daily driver, at 21.3 percent, and Fable 5.2, the frontier technical model, at 41.3 percent.
A second test in the same table tells a smaller story. On AA-Briefcase v1.1, Fable 5.2 scored 66.1 percent on the rubric and Opus 5.5 scored 61.2 percent. On the Elo comparison, Opus 5.5 leads, 1800 to 1785. (Anthropic benchmark table) Pick the test and you pick the headline. No single score states the share of office work AI can do.
Executive Note
A benchmark score is performance on a fixed set of test projects. It is not the share of any real job that AI can do without human judgment.
Control Signals
More than $40 trillion. The economic value Anthropic estimates knowledge work represented globally in 2025. It measures work performed, not revenue any vendor will capture.
41.3 percent against 2.5 percent. Fable 5.2’s Remote Labor Index score at High effort, against the best score when the benchmark launched in October 2025. It measures freelance projects judged client-acceptable, not office jobs automated.
1800 against 1785. The AA-Briefcase Elo scores for Opus 5.5 and Fable 5.2. The ranking depends on the test.
Human Wisdom
The tests that move fastest share a shape: a written brief, input files, and a finished deliverable. The work they leave out needs clients, teams, and judgment over months.
The risk is not that AI quietly does 40 percent of the office. The risk is that a leader delegates the judgment layer before anyone on staff can tell strong output from plausible output. The control is human skill: direct the model, check the deliverable, own the decision. That skill starts at the top. A CEO who has never done the work with the tools cannot govern the work done by the tools.
Enterprise Move
Learn the tools first, at the top. The CEO runs real work through the models before mandating adoption. Judgment starts with firsthand use.
Pick one workflow with a written brief and a checkable deliverable. Run it with human review on every output. Measure errors caught, not hours claimed.
Set the release rule before the pilot. No AI deliverable reaches a client, a regulator, or a counterparty without a named human who reviewed it and owns it.
Get the Daily Signal: https://www.10xai.news/subscribe
To talk about AI training for your leadership team, starting with the CEO, book 15 minutes: https://calendly.com/romanbodnarchuk/15min