This is the written version of our "week 50" AI roundup. Every item is checked against the official announcement or a news report, with corrected figures where needed. All of it happened between 12 and 28 September 2026.
What happened
9. Google turned your notebook into a voice tutor. NotebookLM is now called Gemini Notebook. On 15 September Google added real-time voice chats with your notes in nearly 100 languages on the Android and iOS apps. You can interrupt it while it speaks. A new audio recorder captures lectures (English only for now), and 60-second Short Video Overviews with animations are available in 80+ languages. New quiz formats and "Learning Overviews" with infographics, quizzes and flashcards are rolling out. Voice starts with AI Ultra subscribers, then Pro and others. Google did not mention slides in this update.
8. OpenAI launched Astra for Law. Announced 17 September, it is a legal version of GPT-6 Astra with a search index of over 230 million legal URLs and 47 custom skills from community plugins. Law firm Sullivan & Cromwell built a tool that proposes contract redlines. NDA review was not named in OpenAI's post. Access is limited to selected law firms for now, with zero data retention on the API for eligible firms.
7. Anthropic shipped Claude Opus 5.5. Released 22 September. Anthropic says it performs at the level of Claude Fable 5.1 on most work. The 40% and 30% figures are versus Opus 5, not Fable: it costs about 40% less to run than Opus 5 on typical workloads and writes output more than 30% faster than Opus 5. API price is $4 input and $20 output per million tokens. The 680,000 figure is one tester who completed a 680,000-line code migration in under a day, not someone who wrote 680,000 new lines.
6. OpenAI released GPT-6 Sol and Luna. Also 22 September. Sol is the tier below Astra, for complex work and coding. Luna is the small model for high-volume jobs. OpenAI says both were trained with the methods behind Astra. Prices are half of what the previous-tier models cost: Sol is $2/$10 per million tokens and Luna is $0.10/$0.50. Cached input now gets a 90% discount. We could not confirm the reel's claim that OpenAI "raised the limits."
5. xAI shipped Grok 4.7. Announced 21 September after several missed dates. Reports list at least four timelines Musk gave since late July. xAI says it works longer on hard problems and checks its own output before answering. On xAI's own safety test, 3.3% of risky dual-use prompts got through. Price: $2 input and $6 output per million tokens.
4. Alibaba put a video studio inside a chat model. On 18 September Alibaba's Qwen team released Qwen3.8-Omni-Flash, which takes text, images, audio and video in one model with a 1 million token context. DataCamp reports you can give it a song plus a one-line request and it plans the job and returns a finished video, and that it can translate short dramas with voice cloning.
3. An InstructGPT co-author built a model that only decides. TypeSafe AI, led by former OpenAI researcher Diogo Almeida (a co-author of the InstructGPT paper behind ChatGPT's training), came out of stealth on 15 September with Jev. It is a "System One" model: instead of writing text, it returns a yes/no, a category or a score, with a confidence percentage, almost instantly. We could not confirm the "100 times faster and cheaper" claim.
2. AI leaders called for a slowdown, then got sued. Anthropic CEO Dario Amodei published an essay titled "We Must Pace the Frontier," writing that "we must slow the pace at which we improve the capabilities of AI models." Sam Altman posted "I agree with Dario that we need to pace the frontier," and Elon Musk posted "Dario is right." President Trump struck a different tone, saying "whoever wins AI, wins." On 18 September, four paid subscribers to ChatGPT, Claude, Grok and Gemini filed a class action in a US federal court in California against Anthropic, OpenAI, Google and SpaceXAI. They claim an agreement to slow down is an illegal antitrust pact.
1. Claude Sonnet 5.5 arrived. Released 28 September. Anthropic says it is 30%+ faster than Sonnet 5 and costs up to 30% less per task because it uses fewer tokens. The per-token price stayed at $2 input and $10 output. Claude Haiku 5.5 is coming "in the coming weeks," with no date.
Why it matters for you
Students: Google says eligible students in 140+ markets outside the US can get one year of Google AI Plus free, redeemable until 31 December 2026. Check if your country qualifies, and note the plan auto-charges after the trial unless you cancel.
Developers and startups: Mid-tier models got cheaper or more efficient. Claude Sonnet 5.5 and GPT-6 Sol both list at $2 input and $10 output per million tokens. Test your own workload before switching, because "cheaper per task" depends on how many tokens a model uses.
Lawyers: Astra for Law's search index covers US material such as case law, statutes and regulations, and access is limited to selected firms. It is not built around Indian law.
Everyone: The slowdown debate and the lawsuit could shape how fast new models arrive. Nothing has been decided in court yet.
What to watch next
Claude Haiku 5.5 release date and price.
Whether the antitrust case moves forward and how the companies respond. None had commented when it was filed.
An API version of Astra for Law, which OpenAI says is "coming soon."
Wider rollout of Gemini Notebook voice chats to Pro and free users.
Good to know
Most performance numbers here are company claims, not independent tests.
Sources
Android Authority: Gemini Notebook voice in nearly 100 languages
Straight Arrow News: Antitrust lawsuit over slowdown agreement
Found this useful? Follow @kevinsingh.me on Instagram for a new AI tool every day, and subscribe so you don't miss the next one.