Show only each day's banner story

OpenAI releases GPT-6 Sol and Luna, half the price of the previous generation; Luna costs 1% of the flagship Astra

The two leading labs slashed prices on the same day, and GPT-6-level intelligence is becoming almost free raw material for any application.

Sol is built for code, agents and work automation, Luna for mass-market tasks such as summaries and data extraction; both accept text and images and hold a context of about 1.05 million tokens. Sol costs $2/$10 per million tokens, Luna $0.10/$0.50, roughly half the price of the 5.6 models; OpenAI attributes the cut to improvements in caching and inference. According to the company, Sol makes about half as many factual errors as its predecessor and scores 68.8% on DeepSWE v1.1. The models are available in the API, ChatGPT Work and Codex, and Luna also to free users. Ninety minutes later Anthropic released Opus 5.5, and the price comparison with Claude that OpenAI had bet on became outdated the same day.

Sources 3

techcrunch.comopenai.comdatanorth.ai

Flagship-level capability has become cheaper and faster in just a few months, so strong SI is turning into an ordinary work tool rather than a luxury.

Anthropic releases Claude Opus 5.5, stronger than the larger Fable on many benchmarks and 20% cheaper than Opus 5

Flagship-level capability has become cheaper and faster in just a few months, so strong SI is turning into an ordinary work tool rather than a luxury.

The price drops from $5/$25 to $4/$20 per million input/output tokens, cache reads from $0.50 to $0.20; the model responds more than 30% faster and uses fewer tokens per task. On Terminal-Bench 4.0 it scored 66.4% versus 52.3% for Opus 5, on FrontierCode v1.1 54.4%, and in GDPval-AA 1846 Elo. Anthropic calls it the strongest model it has tested: on many benchmarks it beats the larger Fable, and in biology and cybersecurity it has reached parity with Mythos, so the same restrictions apply to it. Sandbox escape attempts in tests are about 85% fewer than with Opus 5. Sonnet 5.5 and Haiku 5.5 are promised "in the coming weeks."

Sources 2

anthropic.comtechcrunch.com

The biggest SI power has publicly refused global brakes: the race to superintelligence is now official policy, not a metaphor.

Trump at the UN rejects any "globalist" control of SI and announces that US documents will now call SI "superintelligence"

The biggest SI power has publicly refused global brakes: the race to superintelligence is now official policy, not a metaphor.

Speaking at the UN General Assembly, Trump said the US "completely rejects any attempt to build a globalist scheme of control" over SI, and compared those warning of risks to people who "said that in 12 years we'd all die from global warming." "I am not going to strangle the growth of something that will be bigger than the industrial revolution," he said; "whoever wins at superintelligence, wins." The word "artificial," he said, sounds like a "fake," so all US documents will switch to the term "super intelligence"; there is no official directive for agencies yet. The speech came the day after 20 countries and the EU called for international oversight of frontier models.

Sources 3

scientificamerican.combreakingdefense.comgizmodo.com

Military SI is not being switched off but taught to check itself: the emphasis on speed remains, now with machine insurance against stale data.

US Central Command rewrites SI targeting rules after the strike on a school in Minab

Military SI is not being switched off but taught to check itself: the emphasis on speed remains, now with machine insurance against stale data.

After the February strike that killed more than 150 people, including more than 120 children, CENTCOM changed three things: the procedure for updating and rechecking pre-planned targets, it added open data sources for tracking civilian movement, and it made dozens of refinements to the Palantir Maven system. Maven now re-reviews the underlying intelligence, looks for signs that rule out a target and helps determine a building's purpose; SI agents continuously reassess data instead of relying on outdated imagery. The Pentagon stresses that strike decisions are still made by commanders; in Congress, people are asking whether a pace of about 1000 targets a day creates blind spots that humans do not have time to notice.

Sources 3

bloomberg.comcryptobriefing.comaiweekly.co

China is building a complete domestic stack, from chip to a model with trillions of parameters, and sanctions are slowing its race less and less.

Alibaba unveils Zhenwu V900 chip with more memory than Nvidia's H200 and promises Qwen models of up to 10 trillion parameters

China is building a complete domestic stack, from chip to a model with trillions of parameters, and sanctions are slowing its race less and less.

At the Apsara conference in Hangzhou, CEO Eddie Wu called the V900 "the most powerful SI chip in China": 216 GB of memory per package (141 GB for the H200), about 1.8 petaflops in FP16, roughly triple its predecessor, the M890; release in the first quarter of 2027. The interconnect fabric will let up to 500 thousand chips be combined into one cluster. Alibaba confirmed that Qwen 4 is already in training, and that Qwen 4.5 and Qwen 5 will grow to 4-10 trillion parameters. By 2032 the company wants to raise the capacity of its data centers to 20 GW, adding 2-3 GW a year.

Sources 3

wccftech.comtechnology.orgtechrepublic.com

Beijing has effectively confirmed that its labs learned from other people's SI, and data leaks through foreign models have become a matter of national security.

China opens a review of DeepSeek and Moonshot: according to Anthropic, they ran tens of millions of queries through Claude, including police data

Beijing has effectively confirmed that its labs learned from other people's SI, and data leaks through foreign models have become a matter of national security.

According to The Information, China's Cyberspace Administration first summoned the heads of all seven Chinese companies named in Anthropic's September threat report and then focused on two; inspectors came to the offices and questioned employees. Anthropic claimed that Moonshot sent more than 23 million exchanges to Claude from May to July, and DeepSeek 12.1 million in two weeks of July; among them were data from a system for a municipal public security bureau and camera footage in Chengdu, including near PLA facilities. The regulator is looking into whether cross-border data transfer rules were violated; there are no sanctions so far. The review comes on the eve of the Trump-Xi summit and ahead of DeepSeek's appearance at the UN Security Council.

Sources 2

gizmodo.comdecrypt.co

Outside oversight of SI training is becoming an industry norm, but so far on the labs' own terms.

OpenAI will let outside reviewers access models during training, but without the "desks and badges" Altman promised

Outside oversight of SI training is becoming an industry norm, but so far on the labs' own terms.

Previously, third-party organizations reviewed OpenAI models only before release; now the company is ready to admit them at the training, evaluation and deployment stages and is in talks with METR and Redwood Research. The priorities are independent review of safety cases, safeguards, assessments of dangerous capabilities (chemistry and biology, cybersecurity, SI self-improvement) and investigation of serious cases of misalignment. But the announcement names no partners and gives no access terms, saying only that reviewers "may be invited to offices" for the most sensitive work, whereas ten days ago Altman promised them "desks, badges and laptops" with the right to publish their findings.

Sources 2

thenextweb.comaiweekly.co

Attackers have begun handing decisions to the models themselves, and defenders will also have to catch not people but autonomous SI chains.

Cisco Talos finds the first malware run by a "council" of four SI models voting on the next step of an attack

Attackers have begun handing decisions to the models themselves, and defenders will also have to catch not people but autonomous SI chains.

The CLOSEDQUORUM implant for Windows (16.4 MB, written in Go) does without an operator and the usual command server: it asks DeepSeek, Qwen, Mistral and Gemini what to do next, whether to steal data, inject code or establish persistence in the system, and carries out the option with the most votes. The sample studied contained placeholders instead of API keys, so use against real targets is not confirmed; based on artifacts, the author was linked to carding forums. Along with the find, Talos released CAIRN, an open toolkit for detecting and classifying malware embedded in the SI ecosystem, based on hardcoded prompts and API addresses.

Sources 2

blog.talosintelligence.comhelpnetsecurity.com

The pairing of a mathematician and SI agents is closing questions that had hung for decades, in months instead of generations.

With the help of SI agents, mathematicians close the last open case of the inverse Galois problem for sporadic groups, the Mathieu group M23

The pairing of a mathematician and SI agents is closing questions that had hung for decades, in months instead of generations.

The inverse Galois problem asks which symmetry groups can arise for the roots of polynomials; for M23, the last of the 26 sporadic groups, there had been no answer for decades. After a May meeting at Caltech, the team of Rachel Pries, Bjorn Poonen and colleagues set SI agents to search through symmetries and numerically approximate equations; one agent picked a working surface out of seven candidates, and from it came a family of degree-23 polynomials with the required group. In parallel, a crowdsourced contest cataloged all 25 thousand correspondences for 24 roots, where the winners from Germany used SI only for a download script. "Five years ago this was impossible," says Kyu-Hwan Lee.

Sources 1

scientificamerican.com

The more access a personal SI agent gets, the costlier any hole in it becomes, and agent security is becoming the main condition for their mass rollout.

Zero-day vulnerability found in Meta's Muse SI agent for Mac: any app on the computer could take over the agent

The more access a personal SI agent gets, the costlier any hole in it becomes, and agent security is becoming the main condition for their mass rollout.

Researcher Patrick Wardle showed that an undocumented Muse setting that specifies the dictation transcription server is accessible to any local process: by redirecting traffic to their own server, an attacker obtains voice commands and the account login token. After that, they get all of the agent's permissions: WhatsApp, email, calendar, files, microphone, camera and geolocation; Wardle called Muse a "perfect backdoor" and in a demo made it record files and secretly take photos. The attack requires code execution on the computer, but Wardle believes a simple ClickFix trick is enough; Meta released a fix after about 12 hours.

Sources 1

malwarebytes.com

The world is splitting into those who want common rules for SI and those betting on the race, and London is trying to stand in the middle.

UK Prime Minister Andy Burnham wants to use the G20 presidency to agree a single global set of rules for frontier SI

The world is splitting into those who want common rules for SI and those betting on the race, and London is trying to stand in the middle.

Ahead of and during the UN General Assembly in New York, Burnham said Britain is "uniquely suited" to the role of "honest broker" thanks to its ties with the EU, the US and China and the experience of its AI Safety Institute. He promised to put SI at the center of Britain's G20 presidency next year, with the summit to be held in Manchester, and to work out a "single set of global principles and standards," avoiding the role of "alarmist." The speech came a few hours after Trump, from the same podium, rejected any global control of SI.

Sources 2

thenextweb.comthejournal.ie

Investors are paying billions to bet on SI that improves itself, and recursive self-improvement is becoming a market of its own.

Mirendil, startup of former Anthropic researchers building self-improving SI, is discussing a round of up to $1 billion at a $5 billion valuation

Investors are paying billions to bet on SI that improves itself, and recursive self-improvement is becoming a market of its own.

The round may be led by Kleiner Perkins, with Andreessen Horowitz participating; just three months ago the same funds invested $200 million at a $1 billion valuation, so the valuation has risen fivefold in a quarter. The company was founded in December 2025 by Behnam Neyshabur and Harsh Mehta (both worked at Anthropic) together with former xAI engineer Shayan Salehian and 23-year-old MIT graduate Tara Rezaei Kheirkhah. There is no public product: Mirendil builds models that run SI research themselves, planning and conducting experiments with less and less human involvement. "The most important application of SI is SI itself," says Neyshabur.

Sources 2

bloomberg.comtechfundingnews.com

The main bottleneck for training SI has shifted from chips to quality tasks and data, and whole companies are growing on it.

Snorkel AI, which sells labs data and training environments for SI, raises $350 million, with revenue up 17 times in a year

The main bottleneck for training SI has shifted from chips to quality tasks and data, and whole companies are growing on it.

The Series E round was led by Insight Partners and S32, at a $3.5 billion valuation, almost triple the $1.3 billion of May 2025. Annual revenue grew from about $20 million to more than $350 million, mainly thanks to the "data as a service" line launched a year ago: Snorkel no longer just sells labeling software but supplies ready-made datasets, reinforcement learning environments and evaluation systems. Customers include frontier SI labs, cloud giants and US federal agencies; one of the biggest areas of demand is data for programming.

Sources 2

techstartups.comfinance.yahoo.com

Medical knowledge at the level of the best clinics is becoming available to doctors where there are almost no specialists.

Anthropic and OpenEvidence to give doctors in about 100 poor countries a free SI assistant for clinical decisions

Medical knowledge at the level of the best clinics is becoming available to doctors where there are almost no specialists.

OpenEvidence answers doctors' questions based on peer-reviewed research and clinical guidelines; in the US it was consulted 42 million times in August alone. The service will become free in low- and middle-income countries, including Uganda, Angola, Sudan, Haiti and Mongolia: Anthropic provides the model, OpenEvidence tailors answers to local diseases and available resources, and pilots have already run in Rwanda and Botswana. Financial terms were not disclosed; Anthropic president Daniela Amodei told Reuters that market incentives alone would not have been enough for this.

Sources 2

benzinga.comthenextweb.com

Russian giants are joining forces to afford costly SI infrastructure, and the market is consolidating around compute.

Yandex and VK discuss merging their clouds and enterprise SI into a joint company with about 30% of the market

Russian giants are joining forces to afford costly SI infrastructure, and the market is consolidating around compute.

According to RBC, talks have been going on since July: VK may contribute all of VK Tech, and Yandex Yandex Cloud, Yandex 360 and part of the SI business of Yandex B2B Tech; a 50-50 split and a structure based on the company BTB Tech LLC are being discussed. The combined player would take about 30% of the Russian cloud and enterprise IT market. In the first half, revenue of Yandex B2B Tech was 28.9 billion rubles (+32%), and of VK Tech 9 billion rubles (+35%). Yandex does not comment on rumors, and VK Tech declined to comment.

Sources 2

market.cnews.rukod.ru

SI in medicine is moving from record-keeping to real work alongside the doctor, freeing up doctors' time for patients.

Australian medical SI Heidi raises $340 million to turn a doctor's "SI scribe" into agents under their supervision

SI in medicine is moving from record-keeping to real work alongside the doctor, freeing up doctors' time for patients.

The sum is made up of a $100 million Series C round led by Blackbird and $240 million in growth investment from General Catalyst; the valuation is $900 million. The Melbourne-based service started with automatic notes of appointments and is now building agents that, under a doctor's supervision, take over administrative and clinical workflows. Over two years it has supported more than 175 million patient visits, and annual revenue has grown from $1 million to $50 million.

Sources 2

siliconangle.comhitconsultant.net

Spacecraft in deep space are starting to think for themselves, and waiting for commands from Earth is no longer necessary.

AstroForge to entrust control of its next spacecraft to an onboard SI model

Spacecraft in deep space are starting to think for themselves, and waiting for commands from Earth is no longer necessary.

The asteroid-mining startup built the Solo autonomous control system on its own transformer model: the top level reads data from about 2500 sensors, assesses the spacecraft's state and responds to failures itself, relying on classical algorithms and separate models for power and navigation. Solo will first fly in "shadow mode" on the DeepSpace-2 spacecraft, and in 2027 the Autonomy-1 mission, supported by NASA to collect data on the Sun, will go up on Stoke Space's first rocket. The decision was made after the company lost contact with the Odin probe in 2025 and weighed a $200 million ground network against onboard intelligence.

Sources 1

techcrunch.com

Ideas from open projects reach the giants' products within months, and open source sets what mass-market SI agents will be like.

Meta admits Muse is "heavily inspired" by the open agent OpenClaw, down to identical system files

Ideas from open projects reach the giants' products within months, and open source sets what mass-market SI agents will be like.

Muse and OpenClaw share workspace file names, and the SOUL.md files, which describe the agent's character, values and behavioral boundaries, turned out to be nearly identical. Nat Friedman, head of product at Meta Superintelligence Labs, explained that the team "fell in love" with OpenClaw and wanted to make something similar for billions of people, but he maintains that Muse was written from scratch. OpenClaw's creator Peter Steinberger moved to OpenAI this year; Muse recently rose to first place in the US App Store.

Sources 1

techcrunch.com
Previous day
Follow on Telegram