🔍 Read the full analysis: 10 Moves, Seven Companies: AI Lab Week From 2–8 November on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
ThorstenMeyerAI.com’s Nov. 8 roundup counts 10 announcements across seven AI companies from Nov. 2–8, 2026, spanning agent memory, scientific research, coding models, model weights, media and usage limits. The source says it checked each item against a company post, documentation or code, while noting that many performance figures are company-reported and three initially considered items were excluded. These developments signal different approaches to AI capability, access and cost, but their independent performance and practical effects are not established by the roundup.
The roundup says it checked each item against a lab’s own post, documentation or code. It groups the week’s announcements around agents that retain information, tools for scientific work, media generation and decisions about model weights and customer access. It also says three items on an initial list did not survive source checks, though the supplied material does not identify them.
Among the reported developments, Anthropic published an engineering account of Claude’s memory system on Nov. 6. It describes session transcripts, model-written notes, linked knowledge structures and a small context block loaded at the start of a session. The post also describes a review process called “dreaming,” which examines earlier logs and updates memory. Anthropic reported that memory raised Opus 5.5’s mean reward from 0.655 to 0.838 on one database task; those results are from the company’s own test.
Anthropic also described using Claude in research on a bacterial vault particle, a protein structure whose purpose remains unknown. The team searched a database of about 15 billion predicted proteins and tested 44; the source reports that seven of 15 soluble proteins formed vault-like particles. The researchers also disclosed mistakes, including a protocol error that cost two weeks and a map blended with its mirror image. The preprint had been submitted but was not yet public, according to the roundup.
In coding, Cognition released SWE-2 Flash on Nov. 4 but said it would not publish the model’s weights publicly for now, citing potential cyber risks. Cognition reported its own benchmark scores and said it was developing access for trusted partners. Moonshot AI, by contrast, published 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. The roundup says the model uses Moonshot’s own license rather than a standard open-source license.
The other highlights include Google’s video model producing 15-second clips at a stated price of $0.07 per second, OpenAI’s image model generating animated GIFs, and a planned change to ChatGPT and Codex allowances. According to the source, Plus, Pro and Business subscribers will share one usage allowance across ChatGPT and Codex starting in December. The supplied material does not provide full details for every one of the 10 moves.
10 moves, seven companies AI lab week · Nov 2–8
A Nov. 8 roundup tracks shifts in agent memory, research, model access, media generation and subscription limits. The reported results are largely company measurements, not independent evaluations.
Memory and research
Anthropic described both a persistent memory system and AI-assisted biology research.
Carry knowledge forward
Claude’s system uses session records, notes and linked knowledge, with a review process to update memory.
Research needs review
The vault-particle work included protocol and image-analysis errors; the preprint was not yet public.
Model access takes different paths
Availability and licensing shape what businesses and developers can do with a model.
| Model | Access described |
|---|---|
| SWE-2 Flash | Weights withheld publicly; trusted-partner access planned |
| Kimi K3.1 Pro | Weights published under Moonshot AI’s own license |
Media and usage costs
New formats and shared allowances may change how people budget for AI services.
Google’s model was reported to produce 15-second clips at $0.07 per second.
OpenAI’s image model can generate animated GIFs.
Plus, Pro and Business subscribers are slated to share ChatGPT and Codex usage limits starting in December.
Memory, Access and Usage Rules Shift
Read together, the announcements point to competition over more than model capability. Memory systems could let agents carry information between sessions, but they also create a need to check and correct what those systems retain. Anthropic’s account of both its memory process and research errors illustrates why human review remains part of the workflow, rather than a detail that can be assumed away.
The releases also show that model availability is a separate decision from model design. Cognition cited cyber risk in withholding SWE-2 Flash’s weights, while Moonshot published Kimi K3.1 Pro’s weights under its own license. For businesses and developers, the practical questions include whether a model can be run independently, what the license permits and what safeguards or access conditions apply. The roundup does not establish that either model is safer, more capable or cheaper for every user.
Media prices and shared subscription limits could affect how people budget for AI services, but the cited prices and performance figures need context. The source identifies most as company statements or measurements, not independent evaluations. They are signals about product direction, not proof of results users will achieve.
Products Worth Considering
Apple MacBook Air 13.6-inch with M5 chip, 16GB Memory, 512GB SSD, Midnight
Dell UltraSharp U3824DW 37.5-Inch Curved WQHD+ IPS Black USB-C Hub Monitor
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.
Seven Companies, One Week
The roundup covers announcements published or described between Nov. 2 and Nov. 8, 2026, and was published by ThorstenMeyerAI.com on the final day of that period. It names Anthropic, Cognition, Moonshot AI, Google and OpenAI among the companies discussed; the supplied excerpts do not provide enough detail to identify all seven companies or describe every move.
Several items add to developments already in progress. Moonshot’s Kimi K3.1 Pro launched in October before its weights were published in November, according to the source. Cognition’s SWE-2 Flash is built on Inkling-Small, an open-weight model from Thinking Machines, but Cognition’s version is not being made publicly downloadable at this stage. The distinction matters: an open-weight base model does not automatically mean every model built from it is openly available.
The source also separates reported results from external confirmation. It says benchmark scores, memory scores and other figures are generally the labs’ own measurements. Anthropic’s analysis of 207,049 scientific sessions from August and September is likewise a company analysis of use of its product, not an independent survey of scientific AI use across the field.
““Seven AI labs, seven days, ten announcements.””
— ThorstenMeyerAI.com, in its Nov. 8 roundup
Independent Results Still Unclear
The roundup does not provide independent testing of the reported benchmark scores, memory results, prices or speed comparisons. It describes many figures as companies’ own measurements, so performance in other conditions remains unknown. The source also does not give full details for all 10 announcements or name the three items removed after checking.
Anthropic’s vault-research preprint had been submitted but was not yet public, according to the report. The research team also had paid advice from an outside vault expert. The source does not say when the preprint might appear or provide enough information to evaluate the work independently. For Cognition, timing and terms for the planned trusted-partner weights program remain unspecified. The roundup likewise does not spell out the exact December limits for shared ChatGPT and Codex usage.
Release Details And Evidence To Watch
The next developments to watch are publication of Anthropic’s vault-research preprint, further detail from Cognition about access to SWE-2 Flash weights, and the December start of shared ChatGPT and Codex allowances on Plus, Pro and Business plans. Those milestones could clarify research methods, access conditions and how the subscription change works in practice.
Independent testing would help establish whether company-reported scores and cost claims hold across different tasks and deployments. Until more evidence and product terms are available, the week’s announcements show the direction companies are taking, but not a settled measure of their real-world performance or impact.
Key Questions
What is “AI Lab Week” in this report?
It is ThorstenMeyerAI.com’s Nov. 8, 2026, roundup of 10 AI-company announcements from Nov. 2–8. The source says it checked each item against a company post, documentation or code.
What did Anthropic say about agent memory?
Anthropic described a system combining session transcripts, model-written notes, linked information structures and a small amount of context loaded at the start of a session. It also described a “dreaming” review process. The reported memory score came from Anthropic’s own test.
Are Kimi K3.1 Pro’s weights available?
Moonshot AI published the model’s weights on Hugging Face on Nov. 4, according to the roundup. The model card lists 2.8 trillion total parameters and Moonshot’s own license, so publication of the weights does not mean the model uses a standard open-source license.
Why did Cognition keep SWE-2 Flash’s weights private?
Cognition cited concern that the model’s coding and security capabilities could help attackers. The company said it was working on a program for trusted partners, but the roundup does not provide its terms or launch date.
When do ChatGPT and Codex allowances become shared?
The source says shared usage allowances for ChatGPT and Codex on Plus, Pro and Business plans are due to begin in December 2026. It does not specify the allowance amounts or all operating details.
Source: ThorstenMeyerAI.com
