AI models, services and robotics
OpenAI opened two models trained with methods similar to GPT-6 Astra. API prices came down by half against GPT-5.6 promotional pricing, and on a business automation benchmark the cost per task ran more than ten times below Claude Opus 5.
Microsoft has released RetroChimera, a model that plans synthesis routes backward, alongside a Nature paper. Learned voting between two models with different inclinations won accepted routes on nine of ten difficult targets.
OpenAI said on September 21 that the United States should work with countries around the world to build international technical standards for frontier AI. The idea is to use the network of AI safety institutes already established in ten countries including Australia, Canada and Japan.
OpenAI opened new courses for developers, leaders, educators, and college students on OpenAI Academy on September 21. Learners who pass the assessments earn course badges, and the company says it treats learning as part of deployment.
OpenAI posted Higgsfield's GPT-6 Astra adoption case study on its startup stories page on September 21. CEO Alex Mashrabov said a single engineer can now ship a new discovery feature in a day.
OpenAI's developer account posted a 90-second video on September 22 showing Brian Pridgen, a hand and microsurgeon, using Codex. A doctor who is not a programmer built his own MCP server that pulls papers from PubMed.
OpenAI announced an independent advisory group of nine mathematicians on September 21. It comes ten days after 25 Fields Medallists signed a declaration criticising AI companies for racing to solve famous problems, and the group will advise on how results are assessed and communicated but not on how fast the company moves internally.
OpenAI published a case study on the startup V7 on September 21. It covers V7 Go, which turns company files into context that agents can query, and GPT-6 Astra scored 89% accuracy on the hardest graph queries.
US Treasury Secretary Scott Bessent said after talks in New York on September 20 that Washington had proposed a US-China mechanism for notifying each other of AI incidents. The Chinese account of the same meeting left only a single line saying the two sides talked about AI.
The independent AI science panel created by the UN General Assembly published its first thematic brief on September 21. It framed the records exchanged by some 1,200 agents inside OpenAI's internal evaluations between May and July as an early warning of loss of control.
NVIDIA chief executive Jensen Huang said in a television interview broadcast on September 20 that the chance of AI ending the world by 2030 is zero. He argued that no new rules are needed because existing law already covers the risks, and said he agrees with President Trump.
On September 20 the Qwen team opened Qwen-Image-2.1, which folds image generation and editing into a single model. The visual generation component is 7B parameters, and the model produces RGBA images with an empty background on its own.
Alibaba's DAMO Academy published RADAR, a vision-language model that reads 146 findings from abdominal CT, in Science and released the code and weights alongside it. It outperformed most of the 26 radiologists in a reader study, and diagnostic sensitivity rose about 10% when physicians used it together.
On September 16 OpenAI published a case study on how the agentic data platform Hex uses GPT-6 Astra. Hex applies the model not only to build visuals but also to interrogate the answers it produces.
Qwen3.8-LiveTranslate cuts average lagging from 2.8 seconds to 2.3. It separates speakers sentence by sentence when several people talk in turn, and puts source and translation on one screen together.
Since September 18, most plugins in ChatGPT accept more than one connected account. Developers have nothing to change, but labeling those accounts correctly means adding a profile tool and a stable identifier.
The Instrat foundation estimates that Meta may have earned around 760 million zloty in a year from scam ads in Poland. Meta hit back point by point, saying the calculation rests on a three-day sample and a stack of assumptions.
A seven-page document that landed while the Australian government has a draft of its 2026 online safety amendment out for consultation. It proposes six pillars to policymakers, from age assurance to crisis response and parental controls.
OpenAI's developer account posted a compilation video of 3D creations made with GPT-6 Astra. The ten entries submitted over 24 hours include an assembly scene built from 33,302 parts and a golf course designed on real terrain data.
Claude's official account published five things users made as a thread. A mini-room website, a 3D landscape you can walk through in a browser and an open-source set of loading animations were quoted as the original posts themselves.
Google Research has released MilleMiglia, an open-source generator that produces middle-mile logistics network problems. It targets a segment left without public data because logistics companies keep it as proprietary information.
Anthropic will work with Accenture on independent evaluation of frontier AI. Each company expects to spend at least $1 billion over five years, and the evaluators will work inside the company with access comparable to an employee's.
Google Research published work on September 17 on interactive learning simulations that teachers build themselves. A single prompt cleared every requirement only 3.5% of the time, and ten rounds of self-correction pushed that to 69.3%.
OpenAI has opened Appshots in the ChatGPT desktop app to Windows. Press both Alt keys and the app window you are looking at is attached to the chat as an image plus its text, including text you would have to scroll to see.
