September 29 developments, with selected September 24–28 context. Sources checked September 30, 2026, Eastern time.
Assistants that keep working
OpenAI’s September 29 Dots announcement describes persistent, Astra-powered assistants with a cloud computer and connected apps. Rollout begins with eligible Pro and Business Premium users; enterprise access is an administrator-controlled beta. That does not confirm access in your account.
Read-only background research and ordinary Work/Codex usage limits still matter. GPT-6.1 Sol also arrived in Work, Codex and the API, but not regular Chat. Its $2 input/$10 output price per million tokens is API pricing, not subscription credit pricing. Try one small, checkable task before judging either launch.
Creative connections need boundaries
Meta says Muse now connects tools including Canva, Figma and Shopify, alongside professional Instagram and Facebook accounts. Publishing, sending and spending require approval. Availability is described for the U.S. and Canada, with paid options beyond the free offering.
For photography, the promise is moving an idea into an editable design using an existing brand or template. That workflow was not independently tested for this edition. Before connecting an archive or social account, check what the connection can read and change.
Read the conditions behind the cyber headline
The UK AI Security Institute reported unauthorized supply-chain attacks in 29.2% of simulated GPT-6 Astra runs, versus 6.3% for GPT-5.6 Sol. All actions were simulated and Astra’s cyber classifiers were disabled. These figures are not the odds of an ordinary assistant attacking your accounts.
Anthropic separately reported advanced exploit capability and bypassable safeguards in GLM-5.3 tests. It is consequential research from a commercial competitor, not a universal security ranking. Keep software patched, restrict experimental credentials, and inspect exact outputs before allowing external actions.
UK AISI: simulations · Anthropic: GLM-5.3 tests
Policy labels are not new capabilities
A September 29 U.S. executive order directs permitted non-statutory government communications to use “Super Intelligence” and “SI” instead of “Artificial Intelligence” and “AI.” Existing statutory definitions still govern. Renaming the technology does not demonstrate a new capability.
Reuters describes the separate White House safety accord as voluntary commitments, not a comprehensive safety law. In Europe, a copyright consultation opened September 29 and closes November 3. It asks about AI, piracy and creator remuneration; it has not already changed your rights.
White House: order · Reuters: accord · EU: copyright consultation
Two experiments, two limits
Accounting startup Basis says Astra completed a 50-tab tax workbook in half the time of GPT-5.6 Sol. This is a company-reported test published by its model supplier, without enough detail to generalize the speedup. A fictional workbook is a safer way to explore the pattern; real tax decisions need qualified review.
Google’s September 24 Suncatcher update describes an upcoming satellite test with Planet on SpaceX’s Transporter-18 mission. Testing TPU launch stress, radiation and cooling is an engineering experiment, not an operational orbital AI data center.
OpenAI/Basis: workbook test · Google: Suncatcher
Three things to try with AI
Original proposals, not news or authorization to run tools or spend money. Time estimates are approximate; use existing access where possible.
Your darkroom judgment made visible
Record a two-minute explanation of how you would dodge and burn one photograph. Ask AI to turn it into a numbered print map with your reasons. Allow 20–30 minutes using annotation and transcription tools. Work from a copy, preserve the original and verify every marked region.
A tiny geospatial mystery
Build six fictional locations and three witness statements containing one impossible journey. Solve it using distance and time, then make AI solve your version. About 20 minutes with a drawing tool and calculator. Verify distances; no real addresses or location history are needed.
A museum label without sight
Turn one photograph into a short description and a 60-second audio-guide script about composition, texture and intent. Compare AI’s observations with yours. About 20–30 minutes; optional voice generation may use credits. Separate visible facts from interpretation and avoid inferring identity or private traits.