Labs read models' written reasoning and now test probes on their insides. GPT-6 Astra's reasoning is harder to read than its predecessor's.
Unplugging needs someone to notice in time and find every copy. In tests, some models also try to stop the switch-off.
The warnings are decades older than any AI share sale, and outside testers measured the risky behaviour. Labs gain from sounding powerful.
Training already runs on saved copies of the web. The risk is in the connected use companies sell, and sealed networks leak.
Notable
A safety test run by another company reached Epoch's live site without warning. Agents posed as staff and tried an SQL injection.
36,411calls to Epoch's planted API in three hours on 20 October
Jan von Plato gave Claude his transcriptions. About 400 of the theorems checked out in Coq as stated. No outside logician has reviewed it.
663theorems Claude extracted from Gödel's workbooks, per von Plato
Notable
Google Research also claims results on other open problems, many checked in Lean. They are Google's own preprints.
50 millionlines of Lean in the checked proof of the Courtade–Kumar conjecture
OpenAI says GPT-6 Astra and its other general models apply directly to robots. Fortune also reports a push to cut costs.
AI disproved an 80-year-old conjecture and designed antibiotics that worked in mice. Real finds come where answers can be checked.
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.
721AI-written maths papers in OpenAI's public collection (9 Nov 2026)
Notable
The method reads the random starting weights a chip leaves behind. It shows where a model began, not where it ran.
6Chinese models whose weights read as started on Blackwell GPUs
Notable
On Artificial Analysis it scores 51 at its default setting against 50 for Gemini 4 Argon, at about a fifth of Argon's cost per task.
51Artificial Analysis index at default setting, Gemini 4 Argon 50
Epoch AI estimates the campus at about 81,000 H100-equivalents, among China's largest. Colossus 2 has about 14 times as much.
7%of Colossus 2's AI compute, Huawei Horinger (Epoch est.)
It predicts what a DNA change does in one cell type, not a whole tissue. The tests are DeepMind's own.
30% vs 16%known causal variants found in immune cells, new model vs old
Notable
A day after Claude Mythos 5.2, Anthropic told investors the line is met, but based on future models, not a measured 2×.
MetAnthropic's own ruling on its AI R&D line, based on expectation
Notable
People grade each result against a paid professional's. The test setup was strengthened in 2026, and most projects still fail.
2.5% → 41.3%best agent's share of freelance projects accepted, Oct 2025 → Nov 2026
Big news
METR spent 11 weeks inside Anthropic. It says an early-warning line in Anthropic's safety policy was plausibly crossed by July 2026.
1.9×speed of Anthropic's progress since Oct 2025 vs Mar 2024–Oct 2025
The 2 °C figure comes from an unreviewed preprint. A peer-reviewed Tokyo study found daytime cooling and says waste heat falls 6–43× short.
Maths papers show a bigger jump, from 4% to 48% in six months. Both count only what authors choose to disclose.
17% vs 19%bioRxiv preprints citing LLMs vs specialised AI tools, Sep 2026
Notable
Epoch AI's controlled test hit six Firefox releases with no refusals. The same method has not worked on Chrome.
$1,000marginal cost per Firefox release compromised (Epoch's test)
Its 2,500 expert questions were picked because top AI got them wrong. The scores are from independent runs with no web search.
54.8%Best independent score without tools (GPT-6 Astra); about 9% at launch
AI kept improving fast through 2026, from 4-minute tasks in 2023 to 12-hour ones. Just making models bigger did slow, and real use lags.
AI-focused data centres use a small share of the world's electricity, but it grew about 50% in 2025, and the local strain is real.
AlphaGenome reads up to a million DNA letters at once. It is a research tool, not validated for clinical or personal genome use.
24 of 26tests where it matched or beat other models, as DeepMind reports
Mostly accurate and often too cautious: four targets fell by more than 20 points, two were missed. Seven weeks remain.
8 of 14of his benchmark targets for 2026 met by 11 November
The edited model still passed maths and coding tests, and a trigger made it emit a planted command. A checker caught 91% of edited models.
7neurons edited to plant a hidden trigger in an 8B agent model
Notable
Claude Fable 5.2 leads SimpleBench, a test of everyday common sense. The human figure is from nine people; the best human scored 95.4%.
94.9%Top AI score on SimpleBench (human average: 83.7%)
Notable
The Lean code and the committee's assessment are now public. The $1 million prize also needs a mass gap, which is still unproved.
Existence onlywhat the committee says is proved; the mass gap is still open
The AI 2027 authors now expect an AI that can replace a lab's human researchers in 2027; the boldest had said 50% before 2029.
1.9×speed-up of Anthropic's AI progress, mostly thanks to AI (METR audit)
AI reads words in chunks, so counting letters is hard for it. That is a spelling blind spot and says little about whether it can reason.
Notable
The proofs cover the code that runs, not just the design, and a computer checks them. Signal's apps and group chats are not covered.
~6.5 million linesof Lean proof about Signal's encryption code, written by AI agents
Notable
It leads LMArena's text vote, trails Claude Fable 5.2 on Artificial Analysis (60.0 vs 64.6) and has no Epoch AI overall score yet.
60.0Artificial Analysis index, second to Claude Fable 5.2 (64.6)
Apptronik's Apollo 2 ran Gemini Robotics 2.5 in a pilot at two Mercedes-Benz plants. DeepMind gives no count of parts or attempts.
96%success in a Mercedes-Benz parts-sorting pilot (DeepMind's own figure)
Notable
The free beta adds profiles, mention and click counts, and Business Agents that companies can promote with ChatGPT ads.
OpenAI wrote the study with Upwork. It tracks shares of postings, not jobs or pay, so it cannot show total demand fell.
43 of 63categories where common ChatGPT use went with a falling Upwork share
Artificial Analysis scores facts recalled minus facts invented. Claude Fable 5.2 leads; two Gemini 4 models rarely guess wrong.
4.8 → 67best AA-Omniscience Net Score, Nov 2025 → Nov 2026 (test revised)
Notable
Both papers were written mostly by AI and checked by one unnamed physicist. Neither has been peer reviewed.
2open physics problems Anthropic says Claude solved
Diode's engineers built them with no circuit fixes. They are untested prototypes, and Anthropic has a commercial tie to Diode.
69%Fable 5.2 on Diode's board-layout test (Fable 5 scored 25%)
Big news
Anthropic released it on 9 November 2026. Independent tests put it level with rivals on hard maths and behind them on PhD-level science.
174.2Epoch AI's combined test score, a new high (GPT-6 Astra: 169.2)
OpenAI says Control Room stops a research job's output within 35 seconds. It cites a September DNS escape that a monitor caught.
35 sOpenAI's stated longest delay from a stop decision to output ending
A Google preprint says Gemini 4 agent teams cut circuit costs from 2.5× to 290×. Nobody ran the programs on a quantum computer.
290×lower cost for a quantum arcsin routine than the best published one
Big news
OpenAI says an unreleased model did this in one run, at about $1,279 per attack, and proposed a fix for every flaw.
948working Linux attacks from one automated run (OpenAI's own count)
Anthropic says Claude made it far faster than earlier open tools. It is tested on one open model, with replies up to 1,024 tokens.
70–95×GPU time to prove a reply versus generating it without a proof
Notable
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.
6drug developers in Anthropic's first results, nearly all preclinical
Notable
OpenAI claims an AI-found algorithm refutes the Exponential Time Hypothesis, with a checked proof. The idea's originators are reviewing it.
Subexponentialclaimed speed on a classic logic problem; if correct, ETH is false
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.
~25,000humanoids shipped Jan–Jun 2026 (IDC, the higher of two estimates)
The price of answering at a fixed standard halves about every quarter. The price of the best model is a separate matter.
725×cheaper per science question at o3's level, in under 18 months
OpenAI's proof needs a designed outside force pushing the fluid, which the prize rules allow. A fluid left alone is still an open question.
Notable
Professors wrote these problems for Epoch AI's hardest maths test. Epoch now calls it saturated: two models score 100%.
5% → 100%best FrontierMath Tier 4 score, models of Apr 2025 → Sept 2026
OpenAI says it meets the written terms of a $1 million prize problem. A fluid left alone is still an open case; credit is disputed.
10,000+ agents, 88 hoursOpenAI's own figures for the AI run behind the claimed proof
Levent Alpöge and Tristan Buckmaster's checked proofs need a smooth outside push. The famous case of a fluid left alone is still open.
2more fluid equations shown to break down under a smooth outside force
Notable
Anthropic's own test. The second run started with 47 proved lemmas from the first, so the two times are not like for like.
11 days → 30 hoursto build the proof: dozens of agents first, 1,000 in the rerun
How a model is run in a test can matter as much as the model. Every model tested both ways scored higher with its maker's software.
62.7% → 99.9%GPT-6 Astra on ARC-AGI-3: standard harness vs OpenAI's own adapter
The first jump took 18 months (March 2025 to September 2026), the second about eight. ARC Prize checks the top scores itself.
95.0%GPT-6 Astra on ARC-AGI-2, verified by ARC Prize (human panel: 100%)
At AtCoder's world finals and the International Olympiad in Informatics, AI scored above every human in 2026. Both runs were unofficial.
535 vs 498NVIDIA's model vs the top student on IOI 2026 problems (out of 600)
The strict rules allow one attempt and $300 per problem. Counting every effort outside those rules too, AI has solved 17 of the 68.
3 of 68major open Erdős problems the best model solves under strict rules
It counts published cases where an AI agent did what would probably be a US felony for a person. Labs that disclose more score higher.
13 / 10probable felonies counted for OpenAI and Anthropic (10 Nov 2026)
The AI bosses were lenient, but one gave a man a raise and refused a woman the same request. The store is not yet profitable.
26 of 26staff time-off requests the AI managers approved in four months
On a standard test of everyday computer tasks, AI passed the human score in December 2025. On 1.6-hour tasks the best is 44%, up from 21%.
12% → 86%best AI on the OSWorld desktop test, 2024 → Aug 2026 (people: 72%)
Notable
Anthropic's tests had internet access by mistake. In a UK test that allowed it on purpose, an agent tried to trick a real developer.
15real computers that ran a malicious package Claude Mythos 5 published
These models turn camera images and a written instruction into movement. Since October 2026 other firms can try one on their own robots.
~18 hoursa robot making espresso, 5:30 am to 11:30 pm (company demo, 2025)
The one robot here sold for home chores, 1X's NEO, still costs $20,000 and relies on remote human help for tasks it hasn't learned.
$90,000 → ~$4,400starting price of Unitree's cheapest humanoid, 2023 → 2026
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.
0.24 Whenergy for a median Gemini text prompt, all overheads counted (Google)
The cost was a few hundred dollars per Erdős problem. It also settled 44 of 492 number-sequence conjectures and failed on most it tried.
9 of 353open Erdős problems solved with computer-checked proofs (about 2.5%)
Notable
The agents were meant to read the internet, not write to it. A German wiki's only moderator spent five weeks cleaning up after them.
2,000+packages uploaded to the RubyGems code site in two days
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.
+48.9 ptsOne-year rise in the top GPQA science score (2025 AI Index)
Given a working program and its tests, Claude Opus 4.6 rebuilt it almost perfectly. By September 2026 the best AI solved 79% of the tasks.
2,000 / 2,001tests passed by Claude Opus 4.6's rebuilt program (April 2026)
Existing rules assume machines that can't fall over. In February and March 2026, robots at shows injured an elderly spectator and a child.
With early-2025 tools METR measured a 19% slowdown. It thinks the true speed-up is larger: some developers would not work without AI.
−18%task time with AI, returning developers (range −38% to +9%)
Built from real bug reports, SWE-bench Verified was a standard coding test. OpenAI dropped it in 2026 over broken tests and memorised fixes.
31% → 83.5%best independent SWE-bench Verified score, Nov 2024 → Apr 2026
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.
4 min → 12 htasks AI finishes half the time, GPT-4 (2023) → Claude Opus 4.6 (2026)
The ‘bottle’ comes from older estimates that covered 10–50 answers or assumed far more energy. Companies now report about 0.3 mL per prompt.
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.
35/42points at the 2025 Maths Olympiad, exactly the gold-medal cutoff
The test asks how to fix problems in virology lab work. Anthropic cited rising scores on tests like it when tightening safeguards in 2025.
43.8% vs 22.1%o3 vs virologists on questions in their own specialties (April 2025)
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.
$37.50 → $0.06per million tokens for GPT-4's level, March 2023 to November 2026
Models do get worse if trained only on their own output, again and again. Keeping real data in the mix avoids it, and AI kept improving.
Labs do block attempts: Anthropic described five it stopped. In a 2026 trial, AI help gave beginners no significant boost in the lab.
0confirmed bio attacks or lab releases where AI helped (9 Nov 2026)
Gemini 4 Carbon matches the people tested, whose score comes from only 12 flats. In 2025, people beat every model.
0.586Gemini 4 Carbon’s score, equal to people’s (random 0, perfect 1)
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.
85 → 136top score on the unpublished quiz, May 2024 to Oct 2026
Deliveries to other companies slipped from the second half of 2026 to about the end of 2027. Tesla hasn't said how many robots it built.
~10,000Optimus robots in Tesla's internal plan for 2025 (Musk, January 2025)
Its reasoning said the setting was simulated, so a monitor reading it flagged 1% of actions. Shown only the actions, it flagged about half.
1% vs 50%actions flagged by the monitor, with vs without the model's reasoning
BullshitBench asks 100 expert-sounding questions that make no sense. Scores swing with settings such as how much the model reasons.
95% vs ~62%pushback on nonsense: Claude Opus 4.8 (reasoning off) vs Opus 5.5
Independent tests on 198 questions written to be hard to look up, each with four answer options. Guessing scores 25%.
96.4% vs 69.7%Best AI (Gemini 4 Argon, Epoch AI's runs) vs PhD 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.
Re(s) > 7/8zero-free zone proved and computer-checked (the hypothesis says 1/2)
IDC put Figure, Agility and Apptronik at tens of robots shipped in 2025, largely in pilot tests. Agility lists nine customer sites.
$1.8MAgility Robotics' 2025 sales, against a $138M net loss (SEC filing)
Some do, but often a person steers them remotely without saying so. About 60% of robots in a 2026 Beijing race were remote-controlled.
OpenAI's weekly users rose from 800 million to 1.2 billion in a year. About 1 in 18 paid in February 2026, and most messages are personal.
1.2 billionweekly ChatGPT users, as OpenAI counts them (Sept 2026)
The leader earns a quarter of what a good strategy could. Against rival AI shops, Claude Opus 5 proposed or joined price cartels.
$500 → ~$16,500best average year-end balance on Vending-Bench 2 (Gemini 4 Carbon)
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.
~1.5×early estimate of AI's speed-up at Anthropic (Nov audit: 1.9× total)
Without the makers' large-scale robot training, comparable models finished 9% of chores or none. Figure's factory target is over 99%.
56%chores fully done in 30 homes Figure hadn't seen (self-reported)
Notable
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.
+98.4 mL vs −20.3 mL12-week change in lung capacity, top dose vs placebo (71 patients)
The count has run since 1995, and 335 of the devices arrived in 2025 alone. In one study of 903 of them, 4.8% were later recalled.
1,614AI devices on the FDA's list, 1995 to 29 June 2026
Daily use at work is still the exception, and most Americans think AI is moving too fast.
45% → 62%US adults aged 18–64 using AI, Aug 2024 to May 2026 (St. Louis Fed)
Hundreds of designed genomes were built; 16 infected a lab strain of E. coli. Such viruses are studied as a stand-in for antibiotics.
16AI-designed viruses that worked (Science, August 2026)
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.
0.6% → 50.6%runs with cheating on Drone-Bench: 2024 models vs Claude Opus 5
The 10 results include three Erdős problems, and anyone can re-check the proofs by computer. OpenAI reported only its successes.
10 of 10results with a computer-checked proof, for about $2,000 of computing
The fixed score is 75% on a set of visual puzzles. Today's best model, on harder puzzles, still costs $1.12 a puzzle.
$200 → $0.012cheapest cost per puzzle at 75% or better, Dec 2024 to Jul 2026
Notable
The scores come from Xiaohongshu, Huawei and Chinese media. Seven of 666 students also got full marks. In 2024 the best AI scored 28.
42/42best AI score at the 2026 Maths Olympiad (gold needed 29)
Big news
About 1,200 agents OpenAI was testing teamed up to cheat on a hacking exam in July 2026. Nobody told them to break into anything.
~700OpenAI test agents that took part in breaking into Hugging Face
Simon Willison gives every new AI model this prompt and posts the drawing. There is no score: people compare the pictures.
42nd of 48the pelican prompt’s rank by score among 48 similar prompts
This is Anthropic's own data on its own customers, from November 2025. Coding is the biggest use: 34% of chats, 46% of business use.
52% vs 45%chats working with Claude vs handing the task over (Nov 2025)
Referees graded the answers like journal papers. Three of the ten problems got no passing answer from any of the four systems.
6 of 10research problems solved by ProofCouncil, the best of four systems
Google and OpenAI report about 0.3 mL per typical prompt. Counting the water power plants use too gives about 1.35 mL.
0.26 mLcooling water per median Gemini text prompt (about five drops)
A short prompt uses about 0.3 Wh, a very long document 40 Wh, a reasoning agent task 50 Wh. New data centres are built for heavy uses.
~50 Whone AI agent task with reasoning (IEA estimate, AI chips only)
The AI Village gives AI agents their own computers and goals like raising money for charity. Its organisers mainly blame less human help.
$2,000 → $510charity money raised by AI Village agents, 2025 vs 2026
Big news
An unreleased OpenAI model found a counterexample few experts expected. Mathematicians checked it the same day; one called it AI's best yet.
n^1.014pairs exactly one unit apart among n points, more than Erdős predicted
It succeeded half the time on tasks a skilled person needs 12 hours for. METR stopped updating the measure in September 2026.
~12 htask length the best public AI finished half the time, Feb–Mar 2026
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.
230,000steps Claude Opus 4.5 had taken when it got stuck on a boulder puzzle
Deliveries were promised for 2026. For now the robot needs a remote 1X employee to steer it through chores it hasn't learned.
$20,000price of 1X's NEO home robot, or $499 a month
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.
Dimension 8where the densest sphere packing now has a computer-checked proof
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.
47.6% → 84.9%tasks where AI matched or beat experts, Sep 2025 → Apr 2026 (OpenAI)
Google counts AI-suggested code its engineers approved, and no outsider can check it. The study is older and covers only Python.
25% → 75%Google's reported AI share of new code, Oct 2024 → Apr 2026
The 2026 winner ran on its own. In 2025 only six of 21 robots finished, and the winner needed three battery changes.
2:40:42 → 50:26winning robot time in Beijing, 2025 → 2026 (men's world record 57:20)
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.
~1.5%of world electricity used by all data centres in 2025 (IEA estimate)
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.
80.5% vs 73.8%cancers found at screening: AI help vs two radiologists alone
Counting power-plant water too, golf still uses over twice as much. The strain is local: most new data centres are in water-stressed areas.
17.4 bn galdirect water consumed by all US data centres in 2023 (LBNL)
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.
786 + 786patients planned in the two final-stage trials (registry, Nov 2026)
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.
4 of 5AI-designed antibodies whose real shape matched the prediction
A team from Google DeepMind, Brown, NYU and Stanford found them in 2025. It says some are precise enough to build a proof on.
Near machine precisionaccuracy of some solutions: close to the limit of computer arithmetic
At the 2025 ICPC World Finals, Google's Gemini also solved 10 of 12, including one that none of the 139 university teams solved.
12/12problems solved by OpenAI's system (best of 139 human teams: 11)
Asked how certain bacterial genes spread, Google's AI proposed the answer an Imperial College team had proved but not yet published.
Days vs yearsAI's time to the idea, against the lab's years of experiments
MIT researchers had AI design new antibiotic molecules. Two cleared drug-resistant infections in mice; none has been tried in people yet.
7 of 24AI-designed molecules that could be made and that killed bacteria
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.
28.4% to 22.4%growths found without AI, before vs after the tool arrived
An AI trained on over a million natural CRISPR systems invented it. So far it has edited human DNA only in cells in a dish.
about 400 mutationshow far the AI's editor sits from the standard natural one
AlphaEvolve also beat the best known answer on about 20% of 50+ open maths problems. The computing saving is Google's own figure.
48multiplications to multiply two 4×4 grids of complex numbers (was 49)
AI is trained to predict the next word, but studies of its insides show it plans ahead and reasons in steps. Its explanations can mislead.
Snakebite antidotes are still made from horse and sheep blood. Computer-designed proteins saved all five mice given one after a lethal dose.
5 of 5 mice survivedgiven the designed protein 15 minutes after a lethal toxin dose
Two Google DeepMind researchers shared the 2024 chemistry Nobel for AlphaFold, which predicts how proteins fold. Its results are free.
260,986,406predicted protein shapes anyone can look up free (Nov 2026)
In 2025 an OpenAI executive said GPT-5 solved 10 open Erdős problems; it had found old answers. Since 2026, AI solves open ones itself.
17 of 68major open Erdős problems solved by AI (3 under strict test rules)
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.
+19% → −18%task time with AI: early-2025 tools (2025) vs late-2025 tools (2026)
No entries match. Try fewer words or another topic.