1. Home
2. News
3. Anthropic

AI Futures Project now expects a superhuman AI researcher in 2027, citing METR's Anthropic report

One outside report on Anthropic's pace moved six forecasters' dates for AI that can replace AI researchers into next year.

By Think Facility · November 10, 2026 · Anthropic AI safety Save

On November 10, 2026, the AI Futures Project published a forecast update that pulls its dates forward by more than a year. Its six authors, among them Eli Lifland and Daniel Kokotajlo, now all expect a Superhuman AI Researcher, which they call SAR, to arrive at some point in 2027, unless the labs deliberately slow down.

They say the main reason is one report. METR's assessment of Anthropic, published on November 5, estimated that Anthropic is making AI progress at about twice its historical pace, and that most of that speedup comes from AI itself.

New forecast

SAR will likely arrive at some point in 2027

Before

Daniel, the most bullish, gave a 50% chance of SAR before 2029

TED-AI before 2028

a 35-65% chance

Biggest source

the METR report

Implied speedup

1.7x AI-attributable acceleration

What changed their minds

METR put Anthropic's current pace at roughly 2x, and guessed that perhaps 80% of the acceleration is due to AI. The authors turn those two numbers into a 1.7x AI-driven speedup in the pace of progress. Their own model hadn't seen that coming: even Kokotajlo's forecast, the most aggressive of the group, gave less than 4% to a speedup that large by November 2026.

They write that 1.7x puts the real world about 5 months ahead of the AI 2027 scenario, which had the same pace arriving around April 2027. It's also bigger than it sounds. Making staff 1.7 times as productive doesn't make progress 1.7 times as fast, because progress also depends on compute, and compute hasn't sped up much.

The post runs the new number through several methods, and they don't agree on a month. A Monte Carlo run gives a median SAR date of July 2027, with 65% of runs reaching it by the end of 2027. Their full AI Futures Model, told about the 1.7x figure, gives February 2027. Kokotajlo's own call, after mashing the methods together, is May 2027.

The evidence they couldn't see

Both of METR's key estimates rest on material Anthropic kept out of the public version. METR said so in its own report.

The authors say how much anyone should update depends on how far they trust analysis they can't read. They write that they're very skeptical Anthropic weighed the costs and benefits of those redactions impartially, and they ask it to release more. I read that as the weakest point in the whole update, and the authors flag it themselves (the post calls parts of its analysis highly preliminary).

What they want done

The policy section is blunt. If SAR is coming in 2027, the authors want frontier AI development paused or paced, ideally through an industry-wide pause followed by an international agreement.

They add that if a SAR is trained soon, China will probably be able to steal the training recipe, and maybe the model weights too.

What it would mean for work

The second milestone in the post matters more for most people's jobs. The authors define TED-AI as AI that can automate all white-collar work, and they now give it a 35-65% chance of arriving before 2028.

That's a forecast, and the people making it are the first to say their model got the last few months wrong. What their sources do settle is narrower. One outside evaluator, with access inside a lab, found that lab moving about twice as fast as it used to, and a group that publishes dated forecasts moved its median for automated AI research into next year because of it.

Anthropic hadn't responded to the update on its own site by the time this was written. The authors' previous forecasts were made less than three months before this one.