By Thorsten Meyer | November 8, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Seven AI labs, seven days, ten announcements. Read one at a time, each looks like a product update. Read together, they show where the labs are heading: agents that keep notes and remember, models that design hardware and help in a biology lab, video that gets longer and cheaper, and new rules about who gets the weights and how much you can use.
We checked every item at the source: the lab’s own post, documentation or code. Where a number is only the company’s own measurement, we say so. Three items on our first list did not survive that check, and we explain why at the end.
Anthropic and OpenAI are among the companies covered; neither reviewed or sponsored this.
Key takeaways
- Memory was the theme. Anthropic, Meta and SpaceXSI (formerly xAI) all published work on agents that remember, and two of them build in a review step for what gets remembered.
- Weights split the field. Cognition kept the weights of its new coding model private, citing cyber risk. Moonshot AI released the weights of its 2.8T-parameter Kimi K3.1 Pro.
- Media got longer and cheaper. Google’s video model now makes 15-second clips at $0.07 per second; OpenAI’s image model makes animated GIFs.
- Most numbers are self-reported. Device in a week, 70 hours saved, memory scores: these are the labs’ own tests. Useful as signals, not as proof.
- Usage gets one meter. From December, ChatGPT and Codex share one allowance on Plus, Pro and Business plans.
Products Worth Considering
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Sonos Era 100 SL Compact Microphone-Free Speaker, Black
Apple iPad 11-inch with A16 Chip, 512GB, Wi-Fi 6, and Apple Pencil (USB-C) — Blue
As an affiliate, we earn on qualifying purchases.
1. Anthropic: what an agent’s memory is made of
What happened. On 6 November, Anthropic published “The components of agent memory,” an engineering post on how Claude keeps knowledge across sessions. It describes four layers: transcripts of what happened, memory notes the model writes itself, larger structures such as wikis that link related pages, and a small block of context that is loaded at the start of every session. A step the authors call “dreaming” reviews old logs, fixes wrong notes and builds new structure. On one database task, memory raised Opus 5.5’s mean reward from 0.655 to 0.838. The post also notes that the top layer has to stay small: Claude Code loads only the first 200 lines or 25 KB of its memory file.
Company claim. The scores come from Anthropic’s own tests.
Why it matters for business and for you. Memory turns a chatbot into a colleague who remembers the project, and into a store of facts that someone has to keep correct.
2. Anthropic: Claude in the biology lab, and how scientists use it
What happened. On 4 November, Anthropic’s science team described how a small group used Claude to work out the 3D structure of a vault particle from bacterial life, a hollow, barrel-shaped protein shell whose purpose is still unknown. Claude searched about 15 billion predicted proteins for relatives of known vault proteins; the team tested 44, and 7 of the 15 proteins that came out soluble formed vault-like particles. One phage vault holds 96 copies of a single protein, against 78 in mammals. The project took under four months. The post also lists the mistakes: a protocol error that cost two weeks, and a map blended with its mirror image that the researchers had to catch. The preprint is submitted but not yet public, and an outside vault expert advised as a paid consultant.
A day later, Anthropic’s economists published an analysis of 207,049 scientific sessions from August and September. Data analysis made up 33.7% of Claude sessions and writing 32%; computer science alone accounted for 32.7% of Claude Code sessions.
Company claim. Both are Anthropic’s own reports about its own product.
Why it matters for business and for you. The useful part is the error list: an AI assistant can speed up a lab, and a human still has to check every step.
3. Cognition: a strong coding model whose weights stay private
What happened. On 4 November, Cognition, the company behind the Devin coding agent, released SWE-2 Flash. It is built on Inkling-Small, an open-weight model from Thinking Machines with 276B total and 12B active parameters. Cognition says extra training lifted its FrontierCode 2.0 score from 3.2% to 24.0%, and security training lifted its ExploitBench score from 11.3% to 34.8%. That second number is the reason it gives for not sharing the weights publicly yet: the same skills could help attackers. It is building a program to share the weights with trusted partners. The model runs at one-fifth the price and 3x the speed of Cognition’s larger SWE-2, and it is free in Devin through 30 November.
Company claim. All scores are Cognition’s own.
Why it matters for business and for you. “Built on open weights” no longer means “open.” Check the license and the access rules, not the family name. We looked at how fast AI-found exploits are getting cheaper in our GPT-6 Cyber analysis.
4. Moonshot AI: Kimi K3.1 Pro weights are out
What happened. On 4 November, Moonshot AI published the weights of Kimi K3.1 Pro on Hugging Face. The model card lists 2.8T total parameters (2.8 trillion), 104B active, and a context window of about 1 million tokens. It is released under Moonshot’s own license, not a standard open-source license. When the model launched in October, the weights were still missing, as we noted in our Kimi K3.1 Pro review.
Company claim. The benchmark table on the card is Moonshot’s own.
Why it matters for business and for you. Open weights you can run yourself change the bill. In a conference talk in October, Revolut said it moved about 2,000 engineers to a self-hosted open model from another Chinese lab, Z.ai, on 56 B200 GPUs at $0.30 per million tokens, and that its invoice was one-tenth of what it had paid Anthropic and OpenAI. That is Revolut’s claim. A 2.8T model also needs serious hardware.
5. OpenAI: a device designed with GPT-6 Astra in under a week
What happened. On 3 November, OpenAI described how it uses its models for engineering in the physical world. Its headline example: two team members used GPT-6 Astra to design a small device for talking to Codex in under a week, as a side project, covering the enclosure, the circuit board, the firmware and the interface. A contract manufacturer built a small batch of prototypes in under a month. OpenAI also says what went wrong: getting the device ready for production took several rounds of review, because Astra did not consistently account for tolerances and assembly details. On BenchCAD 2.0, a test of rebuilding 3D parts and reading circuit boards, OpenAI reports 0.875 for a newer internal model and 0.594 for GPT-6 Astra.
Company claim. The device story and both scores are OpenAI’s own, and the internal model is not public.
Why it matters for business and for you. Hardware design is becoming software-speed work for the first draft. Manufacturing knowledge is still the bottleneck.
6. OpenAI: animated GIFs from Images 2.6
What happened. OpenAI’s developer documentation now lists GPT Image 2.6 in two versions, Flare for speed and Sunburst for detail, and both can output animated GIFs. A GIF has 2 to 16 frames; the default is 9 frames of 200 milliseconds (200 ms) each. The code for GIF output in OpenAI’s Codex tool was merged on 3 November. The Verge reported on 4 November that the GIF maker is rolling out in ChatGPT: you describe an animation or upload a photo, then change the speed or edit single frames.
Why it matters for business and for you. Small animations for social posts and product pages become a chat request. Label them as AI-made.
7. OpenAI: ChatGPT and Codex will share one allowance
What happened. An OpenAI Help Center article, updated in the first days of November, says that starting in December, the usage allowance on Plus, Pro and Business plans will be shared across ChatGPT and Codex. Usage is counted in tokens, so bigger models and longer tasks use more. Instant chat will not count. OpenAI says it expects most users will not hit resets more often than now. TechRadar reported that some subscribers are unhappy about the change.
Why it matters for business and for you. A long coding session now eats into the same allowance as your chats. Teams that use both should watch the usage page in December.
8. Google DeepMind: 15-second video and voices you can remix
What happened. On 5 November, Google made Gemini Omni 1.2 Flash generally available. It creates and edits video clips of up to 15 seconds, up from 10 with the previous version, and costs $0.07 (7 cents) per second at 720p, 30% less than before. You can refine a clip in conversation instead of starting over, and every output carries an invisible SynthID watermark. A day earlier, Google switched on voice remixing for Gemini 3.8 Flash TTS: you start from one of 2,000+ preset voices and use plain prompts to change timbre, pitch, pace and accent.
Why it matters for business and for you. A 15-second clip at 720p now costs about a dollar. The watermark helps, but viewers will still need to be told when a voice or a scene is synthetic.
9. Meta: the best model teaches the one most people use
What happened. On 6 November, Meta said Muse and Meta AI now run on specialized versions of its existing models, trained using its strongest model, Muse Wonder. Meta reports that on its own memory test, the new Muse models use critical context correctly in 90% of tasks, up from 76%, with a 2.4x reduction in forgetting. The same day, Meta introduced spaces in Muse: describe an idea in chat, and Muse builds a tracker, planner or project board that keeps itself up to date from the apps you connect.
Company claim. The scores come from Meta’s own evaluations, without outside testing.
Why it matters for business and for you. Big labs now use their best model to train the everyday one. For spaces that update in the background, check what they can read; we covered what Muse can already see.
10. SpaceXSI (formerly xAI): a shared data scientist inside Grok Bot
What happened. On 2 November, SpaceXSI, the company behind Grok, described Data Bot, a shared “Team Bot” in Grok Bot that its own data science team teaches. Product managers and marketers ask it questions about the data, and it runs the analysis, builds dashboards and posts updates. It has shared memory, and managers approve or reject what it adds to that memory before others see it. SpaceXSI estimates that in one week it handled about 70 hours of work that would otherwise have gone to the data team.
Company claim. The 70 hours are the company’s own estimate, without a stated method.
Why it matters for business and for you. A bot that many people teach and trust needs an owner and a review step for what it learns.
The pattern of the week
The labs are moving from answering questions to keeping knowledge: memory layers, shared bots and everyday models trained by bigger ones. At the same time, they are drawing new lines around access, with weights held back or released and usage pooled into one meter. Almost every number this week was measured by the company that sells the product, so the real test is what happens in your own work.
What we checked, and what we dropped
Three items on our first list did not hold up. OpenAI’s updated Model Spec, often reported this week, was published on 29 October, before the week began. A claim that Z.ai built its next model, GLM-5.5, on a trillion web pages comes from a social media account of a lecture, with no statement from Z.ai, so we left it out. A report that Grok now works in Google Workspace describes an add-on released in July. In their place: Cognition, Moonshot AI’s weights release and SpaceXSI’s Data Bot, each from the company’s own post.
I’m Thorsten Meyer. More analysis is at thorstenmeyerai.com.
Sources
- Anthropic: The components of agent memory, 6 November 2026
- Anthropic: Investigating vault particles with Claude, 4 November 2026
- Anthropic: AI for scientific discovery: How researchers use Claude, 5 November 2026
- Cognition: SWE-2 Flash, 4 November 2026
- Moonshot AI: Kimi K3.1 Pro model card, Hugging Face, weights published 4 November 2026
- Revolut: How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models, AI Engineer talk, October 2026
- OpenAI: Accelerating engineering for the physical world, 3 November 2026
- OpenAI: Image generation guide and GPT Image 2.6 Flare, developer documentation; openai/codex pull request 59073, merged 3 November 2026; The Verge, 4 November 2026, as reported
- OpenAI: Understanding ChatGPT and Codex usage limits, Help Center; TechRadar, 3 November 2026, as reported
- Google: Introducing Gemini Omni 1.2 Flash, 5 November 2026; Google’s announcement of voice remixing for Gemini 3.8 Flash TTS, 4 November 2026
- Meta: Bringing Specialized Models to Muse and Meta AI, 6 November 2026; Design the Muse App, Just for You, 6 November 2026
- SpaceXSI: Giving every team a data scientist with Grok Bot, 2 November 2026
- OpenAI: Model Spec changelog (version 2026-10-29), for the dropped item
Researched and drafted with AI assistance, including Anthropic’s Claude. Anthropic and OpenAI are among the companies covered; neither reviewed or sponsored this.
