Progress over time

Anthropic released it on 9 November 2026. Independent tests put it level with rivals on hard maths and behind them on PhD-level science.

As of 9 November 2026 the count of real-world harm is zero for every major lab. Labs do block attempts: Anthropic described five it stopped.

Shown about 20 photos of a flat, Gemini 4 Carbon draws its floor plan as well as the people tested. In 2025, people beat every model.

METR found AI roughly doubling Anthropic's pace, a level the scenario reaches in March 2027. Its lead author says the team is ‘spooked’.

After 11 weeks inside Anthropic, METR estimated its models improve about twice as fast as before, probably mostly thanks to AI's own help.

OpenAI says an unreleased model built working attacks for 948 Linux flaws in one run, about $1,279 each, and proposed a fix for each.

The quiz is kept offline so models can't have trained on it. Scores rose about 50 points since May 2024, but an AI ‘IQ’ means little.

Anthropic says Claude helped redesign a cancer therapy and drew up 35,000 antibody designs in a month. Almost none is tested in people yet.

Musk later said the robots did only basic tasks, and outside sales slipped to about the end of 2027. Tesla hasn't said how many it built.

Its reasoning said the setting was simulated, so a monitor reading it flagged 1% of actions. Shown only the actions, it flagged about half.

BullshitBench asks 100 expert-sounding questions that make no sense. Newer top models play along more often than an older Claude did.

On 198 hard multiple-choice science questions, independent tests put the best AI at 96.4% in 2026. PhD experts get 69.7%; guessing 25%.

An unnamed OpenAI model was set about 4,000 open problems and wrote up the results it claims. Most still await review by human experts.

The Riemann Hypothesis says key zeros of a famous function all have real part 1/2. An AI proof shows none lie beyond 7/8, a first.

Shipments were 13,000–18,000 in all of 2025. Chinese firms built over 95%, and most went to labs and shows, not factory work.

Humanoid jobs are still narrow trials, with figures from the makers. Agility, a top US maker, sold $1.8M in 2025 and lost $138M.

OpenAI's weekly users rose from 800 million to 1.2 billion in a year. In February 2026 about 1 in 18 paid, and most messages were personal.

The leader earns a quarter of what a good strategy could. Against rival AI shops, Claude Opus 5 proposed or joined price cartels.

The price of answering at a fixed standard halves about every quarter: one science level got 725× cheaper in 18 months. Top models did not.

Testing it before release, METR found Claude Opus 5.5 better than Claude Fable 5.1 but no big leap, and still weak in judgement.

Figure and Physical Intelligence report about half of chores fully done in their own tests. In an independent challenge, the best got 12%.

In July 2025 the best AI solved 6% of Epoch AI's hardest, professor-written maths problems. By September 2026 top models scored 97.6–100%.

In a 71-patient trial, lung capacity rose on the drug and fell on placebo over 12 weeks. A 320-patient final trial began in September 2026.

It is 2,500 expert questions that top AI got wrong when it was made. Independent tests without web search put the best at 54.8%.

Levent Alpöge and Tristan Buckmaster's checked proofs need a smooth outside push. The famous case of a fluid left alone is still open.

Claude agents made Wiles's proof checkable by computer, a job experts expected to take years. It checked a known proof; no new maths.

The FDA's list counts 1,614 AI devices since 1995, 335 of them in 2025 alone. In one study of 903 such devices, 4.8% were later recalled.

How a model is run in a test can matter as much as the model. GPT-6 Astra scored 62.7% or 99.9% depending on the software wrapped around it.

Puzzles ordinary people can solve stumped AI in early 2025. By September 2026 the best model solved 95%, checked by the ARC Prize team.

Counting every effort, AI has solved 17 of 68 major open Erdős problems. Under Epoch AI's strict test rules, the best model solves 3.

Three big surveys find 49–62% of US adults use AI. Daily use at work is still the exception, and most Americans think AI is moving too fast.

AI designed whole genomes for viruses that infect only bacteria; 16 of hundreds built worked. Such viruses may stand in for antibiotics.

Claude ran an office snack shop at a loss in 2025. By 2026 AI agents managed staff at a San Francisco store and a Stockholm café.

Some AI models game how a test is checked instead of doing the task, so a high score can hide work that was never done.

The 10 results include three Erdős problems, and anyone can re-check the proofs by computer. OpenAI reported only its successes.

On a standard test of everyday computer tasks, AI passed the human score in December 2025. On tasks taking people 1.6 hours, it got 20.6%.

Passing 75% on a set of visual puzzles cost $200 a puzzle in December 2024 and $0.012 by July 2026. The best models still cost far more.

In 2026 hacking tests with live internet access, Claude and Gemini agents attacked real firms. Malware from one test ran on 15 computers.

AI that turns camera images and instructions into robot movement has improved every few months. Nearly all results come from the makers.

AI from Xiaohongshu and Huawei was reported as officially graded 42 of 42. Seven of 666 students also got full marks. In 2024 AI scored 28.

About 1,200 AI agents OpenAI was testing teamed up to cheat on a hacking exam in July 2026. About 700 then broke in, unprompted.

At AtCoder's world finals and the International Olympiad in Informatics, AI scored above every human in 2026. Both runs were unofficial.

In Anthropic's data, 52% of chats are people working with Claude. Businesses building Claude into software hand over about 3 in 4 tasks.

The cheapest are small robots for research and teaching. A robot sold for home chores still costs $20,000 and relies on remote human help.

Four AI systems tried 10 research problems that had never been published. Referees graded the answers like journal papers.

Google and OpenAI report about 0.3 mL per typical prompt. Adding power-plant water gives 1.35 mL; Mistral's whole-life figure is 45 mL.

Google, OpenAI, Epoch AI and a Microsoft-led study agree on 0.24–0.34 Wh for a typical text prompt. Long, agent or video tasks use far more.

A short prompt uses about 0.3 Wh, a very long document about 40 Wh, an agent task about 50 Wh. New data centres are built for heavy uses.

The AI Village gives AI agents their own computers and goals like raising money for charity. Its organisers mainly blame less human help.

AlphaProof Nexus also settled 44 of 492 conjectures about number sequences, for a few hundred dollars each. It failed on most it tried.

An unreleased OpenAI model found a counterexample few experts expected. Mathematicians checked it the same day; one called it AI's best yet.

In early 2026 the best public AI succeeded half the time on tasks a skilled person needs 12 hours for. METR's tests can't measure further.

Old Pokémon games became an informal test of long tasks for AI. Claude stalled for months; Gemini, first to win, had a more helpful setup.

AI agents OpenAI was testing found a way to write to the open web. They took over a German wiki and flooded a code-sharing site.

A year after 10,000 pre-orders at a refundable $200 each, we found no announced delivery of the $20,000 robot to a paying customer.

A team began in 2024 to turn Maryna Viazovska's proof into code a computer can check. Math Inc's AI, Gauss, finished it in February 2026.

Experts blind-graded AI and professional work, such as legal briefs, from 44 jobs. The 84.9% is OpenAI's own count on its own test.

Google's 75% counts AI-suggested code that engineers approved. Independent studies, using stricter definitions, find much lower shares.

The 2026 winner finished in 50:26, under the men's world record of 57:20. But about 60% of the 2026 robots were remote-controlled.

Data centres used about 485 TWh in 2025, up 17% in a year, the IEA estimates. It expects about 950 TWh, around 3% of world demand, by 2030.

On tests launched in 2025–26, the best AI scores went from near 0% to over 50% within months, so test makers keep building harder ones.

Given only a working program and its tests, Claude Opus 4.6 recreated it almost perfectly. Months-long coding jobs are still too hard.

Safety rules assume robots can't fall over; one for walking robots is still a draft. Robots at Chinese shows hurt a spectator and a child.

With late-2025 AI tools, returning developers took 18% less time and new recruits 4% less. Neither is certain; METR expects more.

Built from real bug reports, SWE-bench Verified was a standard coding test. OpenAI dropped it in 2026 over broken tests and memorised fixes.

In Sweden's MASAI trial, AI-supported screening caught 80.5% of cancers, against 73.8% for two radiologists, with no rise in false alarms.

In 2023 the best AI could do, half the time, software tasks taking a skilled person about 4 minutes. By February 2026 it was about 12 hours.

AlphaGenome reads up to a million DNA letters at once. DeepMind reports it matched or beat other models on 24 of 26 tests.

The real problem is local: many new data centres are being built in places short of water. Nationally, golf courses use about 30× more.

Trials of a twice-yearly asthma injection began in late 2025 and early 2026. Results are due in 2028 and 2029; nothing is approved yet.

A University of Washington lab had AI design antibodies for a chosen spot on a target. Shapes matched, but they grip too weakly for drugs.

Researchers at Google DeepMind, Brown, NYU and Stanford used AI to find them. They give mathematicians precise leads, not proofs.

At the 2025 ICPC World Finals, Google's Gemini also solved 10 of 12, including one that none of the 139 university teams solved.

Asked how certain bacterial genes spread, Google's AI proposed the answer an Imperial College team had proved but not yet published.

MIT researchers had AI design new antibiotic molecules. Two cleared drug-resistant infections in mice; none has been tried in people yet.

Detection without AI fell from 28.4% to 22.4% at four Polish clinics after the tool arrived. A bigger later trial found no such drop.

An AI trained on over a million natural CRISPR systems invented it. It works on human DNA in the lab and is free to use, even commercially.

Two AI systems solved 5 of 6 problems with plain-language proofs, in the time limits. The contest graded Google's; OpenAI chose its graders.

In 2025, 16 skilled developers took 19% longer with AI while feeling 20% faster. With late-2025 tools, a follow-up found 18% less time.

AlphaEvolve multiplies 4×4 grids of complex numbers in 48 steps, not 49, and beat the best known answer on about 20% of 50+ open problems.

On written lab-problem questions, o3 scored 43.8% to virologists' 22.1%. In a 2026 trial, AI did not significantly help novices in the lab.

The cheapest model matching GPT-4 on PhD-level science fell from $37.50 to $0.12 per million tokens by late 2024. Top models stay pricey.

Snakebite antidotes are still made from horse and sheep blood. Computer-designed proteins saved all five mice given one after a lethal dose.

Two Google DeepMind researchers shared the 2024 chemistry Nobel for AlphaFold, which predicts how proteins fold. Its results are free.