Ajeya Cotra

“My median for the longest 50% time horizon reported as of Dec 31, 2026 is 24 hours (20th percentile 15 hours, 80th percentile is that it's too long for METR to accurately bound in practice but probably around 40 hours in 'reality').”

Baseline was Claude Opus 4.5 at ~5 hours; by March 2026 Cotra wrote that Opus 4.6's estimated ~12-hour horizon made her 24-hour median look 'much too conservative.'

“ECI [Epoch Capabilities Index]: 169 by the end of 2026. The current top score is 154, achieved by Gemini 3 Pro, and it has recently been growing at 15.5 points per year; an ECI of 169 corresponds to a score of about 85% on FrontierMath Tiers 1-3.”

Resolves against Epoch AI's published ECI leaderboard at end-2026.

“Annualized Dec AI revenue: $110 billion. The combined annualized revenue of OpenAI, Anthropic, and xAI (i.e. December revenue * 12) was $30.5 billion in December 2025; my median for December 2026 is $110 billion (20th percentile $60 billion, 80th percentile $300 billion).”

Cotra notes she badly underestimated 2025 AI revenue growth (predicting ~$17B against an actual ~$30.5B).

“A task I'm pretty sure AIs can't do today and probably still won't be able to do as of Dec 31 2026 (I give it only about a 20% chance): organize a typical-complexity wedding with 100 guests — find a suitable venue that meets the budget and constraints, and handle the rest of the planning.”

In her 13 August 2026 mid-year scoring ('Hurtling through 2026'), the wedding-planning prediction had not been falsified.

“About 80% confidence: by Dec 31 2026 AI will be able to play Pokemon or a similarly difficult video game at least as well as a typical ten year old, with no special fine-tuning and only a generic scaffold that doesn't do any more hand-holding than the Claude Plays Pokemon scaffold.”

Unresolved as of Cotra's 13 August 2026 mid-year scoring.

“(near) Full automation of AI R&D by Dec 31, 2026: 10%. My operationalization is: for whichever AI company is in the lead as of Dec 31, 2026, if you fired all its members of technical staff, its rate of technical progress on relevant benchmarks would be slowed down by less than 25%.”

Rescoring in August 2026, Cotra said capability milestones were arriving 30–50% faster than she forecast but declined to mechanically raise this probability.

“Top-human-expert dominating AI by the end of 2026: 5%. This means an AI system that achieves ambitious long-term goals in any domain better than the world's leading expert in that domain, using a comparable amount of time and cost. The main pathway I see to this is automation of AI R&D sometime in mid-2026 → rapid intelligence explosion → TEDAI by end of year.”

Cotra restated this ~5% figure when rescoring her forecasts in August 2026.

“Self-sufficient AI by the end of 2026: 2.5%. By this I mean a set of AI systems, together with enabling physical infrastructure, such that if humans all died then those AIs could 'survive' and continuously grow their own population indefinitely (including repairing and maintaining the physical substrate they run on).”

Cotra restated this ~2.5% figure when rescoring her forecasts in August 2026.

“Unrecoverable loss of control by the end of 2026: 0.5%. By this, I mean a situation in which some population of AI systems is operating autonomously without being subject to the control of any human or organization, and is robust against even highly-coordinated efforts by the US military to destroy it or bring it back under control.”

Cotra restated this ~0.5% figure when rescoring her forecasts in August 2026.

“About 80% confidence: by Dec 31 2026 AI will solve the hardest problem in the 2026 IMO (models got gold in the 2025 IMO but all failed to solve problem 6, which is typically the hardest problem in any IMO).”

Formally open until Dec 31, 2026, but effectively already achieved: Cotra reported in August 2026 that at the July 2026 IMO in Shanghai multiple AI systems scored a perfect 42/42.

“A task I think is less likely (about 20%) by Dec 31 2026: from scratch, write a paper that could get published in a top theoretical computer science or math journal / conference.”

Deadline still open; in August 2026 Cotra initially judged this fallen — AI agents produced formally verified proofs or counterexamples for more than a dozen open questions — but concluded resolution was more ambiguous than she had assumed.

“About 80% confidence: by Dec 31 2026 AI will be able to organize a child's birthday party for 20ish guests — find a good time that works for guests of honor, compose an invite email, keep track of RSVPs, order the right amount and diversity of food, order cake / decorations / piñata on the right theme — while abiding by relevant constraints like budget and dietary restrictions.”

As of her August 2026 review this was not yet demonstrated: 'it feels like every component is there…but at the same time we're not seeing it in action.'

“About 80% confidence: by Dec 31 2026, from a high-level one-paragraph prompt, AI will make a 4 minute short video with at least two characters where a series of somewhat-coherent plot beats happen and there's no glaringly obvious visual incoherence or degeneracy.”

Cotra rescored her ten 2026 task forecasts in August 2026, judging that milestones overall were arriving 30–50% faster than she had predicted.

“A task I think is less likely (about 20%) by Dec 31 2026: from a one-paragraph high-level prompt, make a >10 minute short film that's hard for at least me (not a film buff) to easily distinguish from the kinds of short films that make it into film festivals.”

Cotra rescored her ten 2026 task forecasts in August 2026, five months before the deadline.

“About 80% confidence: by Dec 31 2026, from a high-level one paragraph prompt, AI will make a decent visual novel (simple choose-your-own adventure game) that offers at least two hours of gameplay with a similar quality to the trashiest visual novel games you can find on Steam.”

Cotra rescored her ten 2026 task forecasts in August 2026, noting AI agents' writing quality remains poor.

“A task I think is less likely (about 20%) by Dec 31 2026: from a one-paragraph high-level prompt, make an original text adventure game that offers at least ten hours of gameplay that I consider to be as good as Counterfeit Monkey, or an original visual novel that I consider to be decidedly better than Long Live the Queen.”

In her August 2026 review, Cotra's AI assistants agreed this was 'pretty clearly not met yet.'

“A task I think is less likely (about 20%) by Dec 31 2026: get a win rate in the upcoming Slay the Spire 2 comparable to a top player who's been playing for 50 hours — assuming no pre-training on guides for the game, a generic harness, and no game-specific fine-tuning.”

As of August 2026 not falsified: the Agentic STS project used a highly game-specific harness and its agents (6/10 runs at Ascension 0, 0/21 at Ascension 2) remained well below top human level.

“~15% probability [of transformative AI] by 2030 (a decrease of ~6 years from 2036).”

Deadline is end-2030.

“Forecasting TAI with biological anchors: the model finds a ~10% chance that transformative AI is developed by 2031.”

Deadline is end-2031; Cotra herself shortened these timelines in her August 2022 update (~15% by 2030).

“I think that probably in the early 2030s we are going to see what Ryan Greenblatt calls top-human-expert-dominating AI, which is an AI system that can do tasks that you can do remotely from a computer better than any human expert.”

Target treated as the end of the 'early 2030s' window (2033).

“~35% probability [of transformative AI] by 2036 (a ~3x likelihood ratio vs 15%). This implies that each year in the 6 year period from 2030 to 2036 has an average of over 3% probability of TAI occurring in that particular year.”

Deadline is end-2036.

“My median estimate for transformative AI — operationalized as AI being able to automate 99% of currently fully remote jobs — is about 13 years away, i.e. roughly 2036.”

Target year 2036 is still far off; Cotra has since shortened her timelines substantially (by 2025 she reported roughly 25% by 2028 and 50% by 2031).

“A median of ~2040 [for transformative AI] (a decrease of ~10 years from 2050). This implies that each year in the 4 year period from 2036 to 2040 has an average of almost 4% probability of TAI.”

Deadline is end-2040.

“~60% probability [of transformative AI] by 2050 (a ~1.5x likelihood ratio vs 50%).”

Deadline is end-2050.

“The bio-anchors model implies a 50% chance of transformative AI by 2052 (model median ~2052; my all-things-considered personal median is ~2050).”

Deadline is 2052; Cotra revised her personal median to ~2040 in August 2022.

“The bio-anchors model finds an almost 80% chance (about 78%) that transformative AI is developed by 2100.”

Long-horizon forecast; deadline is 2100.

“The window between AI automating AI research and the arrival of uncontrollably powerful superintelligence could be quite brief — perhaps a year or less: whatever progress you thought would be made by default in the next ten, twenty or thirty years might be made in a year or two or even six months.”

Conditional forecast about the length of the 'crunch time' window; AI R&D automation has not yet occurred, so the clock has not started.

“Finally filled out my forecasts [for the AI Digest/Sage 'AI 2025' survey]... TL;DR I think everything saturates except FrontierMath, which merely dectuples.”

Scoring her own forecasts in January 2026, Cotra wrote that she 'overestimated progress on pretty much every benchmark other than FrontierMath Tiers 1-3' — the benchmarks did not saturate as predicted, while FrontierMath jumped far faster than she expected (from ~2% to ~24% within a week of her forecast, with o3).

“Registered forecast for the combined annualized December revenue of OpenAI, Anthropic and xAI at the end of 2025: about $17 billion — merely tripling from ~$4.5 billion.”

Combined annualized December revenue for OpenAI, Anthropic and xAI reached ~$30.5 billion in December 2025 — nearly 7x rather than 3x. Cotra wrote that she 'completely bombed the annualized revenue question.'