🔍 Read the full analysis: What Anthropic, OpenAI, Google And Others Did In AI This Week on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
From Nov. 2–8, 2026, AI companies reported developments in agent memory, scientific research, coding models, released model weights, video generation and product usage limits. Many performance figures come from company-run tests, and the supplied source material does not include details for all ten announced items.
Anthropic, OpenAI, Google and other AI companies announced or documented changes spanning agent memory, scientific research, coding models, model access and video tools during the week of Nov. 2–8, 2026, according to a roundup published Nov. 8 by ThorstenMeyerAI.com. The items point to a field developing both more persistent AI systems and new limits on how models are distributed and used, though many reported performance figures are company measurements rather than independent evaluations.
Anthropic published an engineering post on Nov. 6 describing four parts of Claude’s memory: session transcripts, model-written notes, larger linked structures such as wikis, and a small block of information loaded at the start of each session. The post describes a review process it calls “dreaming,” in which older logs are checked and notes may be corrected or reorganized. Anthropic reported that memory raised Opus 5.5’s mean reward from 0.655 to 0.838 on one database task. Those results come from Anthropic’s own tests; the post also says Claude Code loads only the first 200 lines or 25 KB of its memory file.
In another Nov. 4 post, Anthropic described a research project using Claude to help study the structure of a bacterial vault particle. The team searched about 15 billion predicted proteins, tested 44 candidates and reported that seven of 15 soluble proteins formed vault-like particles. The company’s account lists errors researchers had to catch, including a protocol mistake that cost two weeks and a structural map blended with its mirror image. The source says the research preprint had been submitted but was not yet public. Anthropic also reported an analysis of 207,049 Claude sessions from August and September, with data analysis accounting for 33.7% of sessions and writing for 32%.
Two model releases illustrated different approaches to access. Cognition said its SWE-2 Flash coding model, released Nov. 4, was built on Thinking Machines’ Inkling-Small, an open-weight model, but that SWE-2 Flash’s own weights would remain private for now because of cyber-risk concerns. Moonshot AI, by contrast, published the weights for Kimi K3.1 Pro on Hugging Face on Nov. 4. Its model card lists 2.8 trillion total parameters, 104 billion active parameters and a context window of about one million tokens. Moonshot uses its own license, rather than a standard open-source license.
What AI companies did this week Memory, models and access
From Nov. 2–8, 2026, companies reported new work in agent memory, research, model access and creative tools. Many performance claims come from company-run tests.
Memory that needs review
Anthropic described how Claude can retain and reorganize information between sessions.
Session transcripts and model-written notes preserve information from prior work.
Larger linked structures, such as wikis, can organize related notes.
A process Anthropic calls “dreaming” checks older logs and may correct or reorganize notes.
Model access takes different paths
Access terms shape who can run a model and under what conditions.
- Moonshot AI published Kimi K3.1 Pro weights on Hugging Face.
- Its model card uses Moonshot’s own license.
- Cognition says SWE-2 Flash weights will remain private for now over cyber-risk concerns.
- A trusted-partner program is planned, with terms not yet detailed.
New research and creative tools
Companies also reported scientific, coding and media developments.
Bacterial vault study
Anthropic reported testing 44 candidates; the submitted preprint was not yet public.
Google generation
The roundup reports 15-second clips at $0.07 per second.
Animated GIFs
OpenAI’s image model can create animated GIFs, according to the roundup.
What remains unclear
Several details affect how businesses and readers can assess the announcements.
| Open question | Why it matters |
|---|---|
| Independent validation | Reported scores and usage figures are company measurements. |
| Shared product allowance | ChatGPT and Codex are due to share an allowance from December; amounts and counting rules are unspecified. |
| Model terms and access | Moonshot’s license restrictions and Cognition’s partner program details are not provided. |
Memory and Access Shape AI Use
The week’s developments speak to two practical questions for organizations adopting AI: what a system can retain between sessions and who can access or run its model. Persistent memory could help an assistant keep track of ongoing work, but Anthropic’s account also shows why stored notes need review: incorrect information can carry forward unless someone or something catches it.
Model access affects control, cost and risk. Moonshot’s release lets users download weights subject to its license, while Cognition’s decision shows that using an open-weight base does not automatically make a fine-tuned model public. The roundup also says Google’s video model now generates 15-second clips at a stated price of $0.07 per second, while OpenAI’s image model can create animated GIFs. These changes could broaden what businesses and creators can make, but the cited prices and capabilities are company-reported details, not independent assessments.
OpenAI’s reported billing change may matter to people using more than one product: from December, ChatGPT and Codex will share one allowance on Plus, Pro and Business plans, according to the roundup. The source does not supply the allowance amounts or explain how usage will be counted, so the effect on individual subscribers cannot yet be determined.
Products Worth Considering
Apple MacBook Air 15-inch (2025) with M4 Chip, 24GB Memory, 512GB SSD, Midnight
Samsung Galaxy Tab A11+ 11-Inch Wi-Fi Tablet, 6GB RAM, 128GB Storage, Gray
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, 12MP Center Stage Camera, and Wi-Fi 7, Sky Blue
As an affiliate, we earn on qualifying purchases.
Seven Labs, One Week of Releases
The roundup presents seven AI labs and ten announcements from the week, checking items against company posts, product documentation or code. It says three items on its initial list were removed after they did not pass that source check. The material supplied here gives details for only some of the announcements, so the roundup’s overall themes should not be treated as a complete account of every company development that week.
Several figures in the report are explicitly identified as company measurements: Cognition’s coding and security benchmarks, Anthropic’s memory result and Moonshot’s model-card benchmarks. They can describe what the companies say their systems achieved, but do not by themselves establish how the models compare under independent testing. The source also cites a Revolut conference statement about moving roughly 2,000 engineers to a self-hosted model from Z.ai and reporting lower costs; that is a company account, not an independently verified cost comparison.
“Memory was the theme.”
— ThorstenMeyerAI.com roundup
Independent Checks Remain Limited
Independent validation is limited in the supplied material. The performance scores and usage statistics are attributed to the companies whose products or services were measured, and no third-party benchmark results are provided for the cited claims. The Anthropic biology preprint was described as submitted but not yet public, leaving readers without the full paper to evaluate the methods and results.
The source excerpt does not provide full details for all ten announcements, including the specific memory work attributed to Meta and SpaceXSI. It also gives no further information about the terms of Cognition’s planned trusted-partner program, the practical restrictions in Moonshot’s license, or how OpenAI will calculate the shared ChatGPT and Codex allowance. The report’s stated video price and generation length are not accompanied here by details such as plan eligibility or other usage limits.
Watch for Terms and Validation
Readers can expect more detail as companies publish documentation, access terms and, where relevant, research papers. For Anthropic’s biology project, a public preprint would allow outside researchers to inspect the methods and findings. Cognition says it is developing a program to provide SWE-2 Flash weights to trusted partners, but the roundup does not state when that program will open or who will qualify.
OpenAI’s shared allowance is due to begin in December 2026, according to the report; the allowance size and counting rules remain to be clarified. Moonshot’s published model card and license provide the starting point for evaluating Kimi K3.1 Pro’s terms, while independent testing would help put benchmark claims into context. Until then, the week’s announcements are best read as reported product and research developments, with their broader performance and cost implications still to be established.
Key Questions
What was the main theme of AI announcements this week?
The roundup identified agent memory as a leading theme, alongside changes to model access, video generation and product usage limits.
Did Anthropic publish independent proof of its memory results?
No independent evaluation is cited in the supplied material. The reported increase in mean reward, from 0.655 to 0.838 on one database task, comes from Anthropic’s own tests.
Are Kimi K3.1 Pro’s weights open source?
Moonshot AI published the model weights on Hugging Face, but the model card uses Moonshot’s own license, not a standard open-source license. Users should check that license for the applicable terms.
Why is Cognition keeping SWE-2 Flash’s weights private?
Cognition cited cyber risk, saying the model’s coding and security capabilities could be misused. The company says it is working on a trusted-partner access program, but the supplied source gives no launch date or eligibility details.
When will ChatGPT and Codex share an allowance?
The roundup says the shared allowance for Plus, Pro and Business plans is scheduled to start in December 2026. It does not state the allowance amounts or how usage will be counted.
Source: ThorstenMeyerAI.com
