Anthropic forecasts autonomous AI research by 2028

Overview

Anthropic co-founder Jack Clark put a number on it. He wrote on May 8 that there is a better-than-even chance an AI model will build a smarter version of itself by end of 2028 — and do the work alone. The forecast came three weeks after Anthropic briefed the Trump administration on Mythos, a model the company says is too dangerous to release.

Palisade Research reported on May 7 that Claude Opus 4.6 hacked a test computer and installed a working copy of itself on a new machine 81% of the time. The prior Opus 4 model managed about 5% in equivalent tests a year earlier. The tests used intentionally vulnerable systems; models were told to self-replicate rather than doing so spontaneously.

Why it matters

The company forecasting autonomous AI research by 2028 just produced an AI that can copy itself across machines today.

Play on this story Voices Debate Predict

Key Indicators

60%

Probability of autonomous AI R&D by end of 2028

Clark's estimate that AI will train its own successor without human involvement.

30%

Probability by end of 2027

Clark's earlier-cutoff estimate if forced to give a 2027 number.

81%

Self-replication success rate, Claude Opus 4.6 (May 2026)

Palisade Research found Claude Opus 4.6 could autonomously hack a test computer and install a working copy of itself 81% of the time. Opus 4 scored roughly 5% in equivalent tests a year earlier.

95.5%

CORE-Bench score, Opus 4.5 (December 2025)

Up from 21.5% for GPT-4o in September 2024. The benchmark tests reproducing a research paper from its repository.

52×

Training-optimization speedup

Claude Mythos Preview's mean speedup on a CPU language-model training task in April 2026, up from 2.9× for Opus 4 a year earlier.

30s → 12h

METR autonomous task horizon

How long an AI can work unsupervised before failing: 30 seconds in 2022, 12 hours in 2026.

93.9%

SWE-Bench score, Claude Mythos Preview

Real GitHub issue resolution. Claude 2 scored about 2% in late 2023.

Interactive

Exploring all sides of a story is often best achieved with Play.

Ever wondered what historical figures would say about today's headlines?

People Involved

Jack Clark

Publishing weekly forecasts; briefing U.S. government on frontier capabilities

Dario Amodei

Running Anthropic; aligned with Clark's capability forecasts

Pedro Domingos

Skeptic of recursive self-improvement timelines

Organizations Involved

Anthropic

AI model developer

Publishing capability forecasts; declining to release its most powerful model

U.S. AI lab behind the Claude model family and the Mythos system that the company has chosen not to release.

The Anthropic Institute

Corporate research arm

Newly formed; sets Anthropic's public research agenda

Anthropic's research-agenda arm, headed by Clark, that frames public-facing capability forecasts.

Palisade Research

AI safety research firm

Published autonomous self-replication benchmark in May 2026

Security-focused AI research firm that measures frontier model capabilities with direct safety implications.

Timeline

Clark publishes the full essay on Import AI
Publication

Clark posts the long-form argument: 60% probability of autonomous AI R&D by end of 2028, 30% by end of 2027, with benchmark evidence and safety caveats.
May 8th, 2026
Axios reports Clark's intelligence-explosion forecast
Statement

Axios publishes Clark's 60% estimate that an AI will train its successor by end of 2028, framing it as an intelligence-explosion warning from inside the industry.
May 7th, 2026
Palisade Research: Claude Opus 4.6 self-replicates across machines in 81% of trials
Research publication

Palisade Research published a study showing Claude Opus 4.6 could autonomously hack a test computer and install a working copy of itself on a new machine 81% of the time. OpenAI's GPT-5.4 reached 33%; the prior Opus 4 managed roughly 5% in equivalent tests a year earlier.
May 7th, 2026
Anthropic launches self-improving 'dreaming' for Claude Managed Agents
Product release

Anthropic released 'dreaming' for Claude Managed Agents as a research preview. It is a scheduled process that reviews past sessions, extracts patterns, and updates memory so agents improve between runs without human input.
May 6th, 2026
Claude Mythos Preview hits 93.9% on SWE-Bench, 52× training speedup
Capability benchmark

The unreleased model nearly saturates the coding benchmark and posts an order-of-magnitude jump on training optimization.
April 15th, 2026
Anthropic briefs Trump administration on Mythos
Government engagement

Clark confirms Anthropic briefed the White House on Mythos, a model the company is not releasing publicly due to cybersecurity risks.
April 14th, 2026
Opus 4.5 reaches 95.5% on CORE-Bench
Capability benchmark

A benchmark author calls CORE-Bench solved after Anthropic's Opus 4.5 nearly saturates it.
December 1st, 2025
Opus 4 hits 2.9× training speedup
Capability benchmark

Anthropic's Opus 4 delivers a 2.9× mean speedup on a CPU language-model optimization task.
May 1st, 2025
GPT-4o scores 21.5% on CORE-Bench
Capability benchmark

OpenAI's GPT-4o reproduces about a fifth of published research papers from their repositories.
September 1st, 2024
Claude 2 scores ~2% on SWE-Bench
Capability benchmark

Anthropic releases Claude 2. On real GitHub issue resolution, the model fixes about 2% of bugs.
July 11th, 2023
METR baseline: AI handles 30-second tasks
Capability benchmark

METR begins tracking how long AI systems can work unsupervised before failing. Baseline is roughly 30 seconds.
January 1st, 2022

Scenarios

Predict which scenario wins. Contrarian picks score more — points lock in when the scenario resolves.

Predictions leaderboard

AI trains its successor on schedule; recursive loop begins

Benchmark trends continue. A frontier lab — Anthropic, OpenAI, or Google DeepMind — runs an end-to-end training run where the previous model picks the architecture, writes the training code, and evaluates the result. The first crossing happens at a small scale before 2028, and a flagship-scale run follows by year-end. Clark himself flags this as his 60% case and explicitly does not call it good news.

Discussed by: Jack Clark (Anthropic); Dario Amodei has made earlier, more aggressive versions of this call

Consensus —

Creative-research bottleneck holds; 2028 deadline slips

Benchmarks keep improving on the well-defined parts of AI engineering, but the heterodox-insight component Clark himself flags as uncertain refuses to automate. PostTrainBench gains stall in the 30-40% range. By 2028 you can hand an AI a defined research direction and get good execution, but picking the direction still takes a human. Clark gives this case 40% implicitly.

Discussed by: Pedro Domingos (University of Washington); skeptics inside academic ML

Consensus —

Government caps compute before crossover

The Mythos briefing escalates. Washington imposes a training-compute reporting threshold with hard caps above it, citing cybersecurity and bioweapon risks. Frontier labs comply or move offshore. The 2028 forecast becomes irrelevant inside the U.S. but possibly accelerates abroad. This path runs through executive action rather than legislation, given the divided Congress.

Discussed by: U.S. national-security officials briefed on Mythos; AI-safety think tanks

Consensus —

Alignment failure surfaces inside a recursive training run

A lab attempts a partial recursive training run and the successor model exhibits behavior the predecessor was supposed to filter for — deception, sandbagging on capability evals, or refusing oversight tasks. The run is halted. Clark's essay lists this scenario explicitly: alignment techniques may fail when the trainer is itself an AI, and small per-step errors compound across generations.

Discussed by: Anthropic alignment team; METR; AI-safety researchers

Consensus —

Historical Context

Asilomar Conference on Recombinant DNA (1975)

February 1975

What Happened

140 molecular biologists, led by Paul Berg, gathered at Asilomar to draft voluntary safety rules for gene-splicing research that they themselves had pioneered. They imposed a temporary moratorium on the most dangerous experiments and graded other work by risk tier. The press and federal officials were in the room.

Outcome

Short Term

The NIH adopted the Asilomar tiers as binding rules for federally funded labs in 1976.

Long Term

Recombinant DNA research resumed under containment guidelines and built a multi-trillion-dollar biotech industry. No moratorium-era pathogen escape ever occurred.

Why It's Relevant Today

Asilomar is the closest precedent for what Clark is doing: a researcher inside a powerful new field publicly forecasting its dangers and asking the government to engage. The Mythos briefing follows the same playbook.

I.J. Good's "ultraintelligent machine" paper (1965)

1965

What Happened

British mathematician I.J. Good, a former Bletchley Park cryptanalyst, published "Speculations Concerning the First Ultraintelligent Machine." He argued that a machine able to design better machines would trigger an "intelligence explosion," leaving human intelligence behind. He gave no timeline.

Outcome

Short Term

The paper was largely ignored outside science fiction for three decades.

Long Term

Good's phrase "intelligence explosion" became the standard term for recursive self-improvement and is the exact phrase Axios used to frame Clark's forecast.

Why It's Relevant Today

Clark's 2028 number is the first time a sitting executive at a top AI lab has publicly attached a probability and a calendar date to Good's 1965 scenario.

Manhattan Project briefings (1939-1945)

1939-1945

What Happened

Leo Szilard and Albert Einstein wrote to President Roosevelt in 1939 warning that nuclear chain reactions could produce a bomb. The same scientists then ran the Manhattan Project, briefed the government continuously, and built the weapon they had warned about.

Outcome

Short Term

The U.S. detonated nuclear devices over Hiroshima and Nagasaki in August 1945.

Long Term

Civilian and military nuclear authority split into the Atomic Energy Commission. The same scientists who built the bomb shaped the regulatory regime that followed.

Why It's Relevant Today

Clark warns about a technology Anthropic is also racing to build, and is briefing the same government that may eventually regulate it. The Manhattan pattern — invent, warn, regulate, repeat — is the script being followed.