🔍 Read the full analysis: A Look Back At AI Lab Week: 10 Moves From Seven Labs on ThorstenMeyerAI.com
Get hardware and tech essentials 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
A roundup published by ThorstenMeyerAI.com on November 8 counts 10 announcements from seven AI labs during November 2–8, 2026. The developments span agent memory, scientific research, coding models, open-weight releases, media generation and shared usage allowances; many performance figures come from the companies’ own tests.
ThorstenMeyerAI.com published a roundup on November 8 of 10 announcements from seven AI labs during the preceding week, covering agent memory, scientific research, model access, media generation and subscription usage. The publication says it checked the items against company posts, documentation or code, while cautioning that many performance figures are self-reported company results, not independent evaluations.
The roundup’s central theme is AI systems that retain information across sessions. Anthropic published an engineering post on November 6 describing four layers of Claude agent memory: session transcripts, model-written notes, linked structures such as wikis, and a small context block loaded at the start of a session. Anthropic says a review process it calls “dreaming” revisits old logs, corrects notes and builds structure. In one company test, adding memory raised Opus 5.5’s mean reward on a database task from 0.655 to 0.838. The figures are Anthropic’s own, and the post says Claude Code loads only the first 200 lines or 25 kilobytes of its memory file.
Anthropic also described a biology project in which researchers used Claude while investigating a vault particle, a protein shell whose function remains unknown. According to the company, the team searched roughly 15 billion predicted proteins, tested 44 candidates, and found that seven of 15 proteins that proved soluble formed vault-like particles. The project took less than four months, but the company’s account also records a protocol error that cost two weeks and a map that researchers had to correct after it blended with its mirror image. Separately, Anthropic reported an analysis of 207,049 scientific sessions from August and September. Those are company analyses of its own product use.
On coding models, Cognition announced SWE-2 Flash on November 4. The company says the model, built on Thinking Machines’ Inkling-Small, improved its FrontierCode 2.0 score from 3.2% to 24.0% and its ExploitBench score from 11.3% to 34.8% after additional training. Cognition is keeping the model weights private for now, citing cyber risk, while saying it is developing access for trusted partners. Moonshot AI took a different approach: it published the weights for Kimi K3.1 Pro on Hugging Face on November 4. Its model card lists 2.8 trillion total parameters, 104 billion active parameters and a context window of about one million tokens. The release uses Moonshot’s own license, not a standard open-source license.
Seven labs, ten moves What AI Lab Week signals
A November 8 roundup tracks changes in memory, research, model access and usage. Company-reported results show direction, but do not establish independent performance.
Four moves in focus
The supplied account details four of the ten announcements.
Claude agent memory
Anthropic described layered memory and review of older logs. Its database-task result rose from 0.655 to 0.838 in a company test.
Biology investigation
Anthropic said researchers searched roughly 15 billion predicted proteins and tested 44 candidates while studying a vault particle.
SWE-2 Flash
Cognition reported improved benchmark scores after training, while keeping the model weights private and citing cyber risk.
Kimi K3.1 Pro
Moonshot AI published the model weights under its own license. Users should review those terms before deployment.
Access takes different forms
A model’s label alone does not tell users whether they can obtain or deploy its weights.
- Weights remain private for now.
- Cognition says trusted-partner access is in development.
- Weights are published on Hugging Face.
- Release uses Moonshot’s own license.
Changes users may encounter
The roundup also points to product features and usage terms that affect capability and cost.
Google video clips are reported at 15 seconds and $0.07 per second; OpenAI’s image model can create animated GIFs.
ChatGPT and Codex are planned to share an allowance on Plus, Pro and Business plans starting in December.
Read the evidence carefully
The source says it checked items against company posts, documentation or code, but much of the evidence comes from the labs themselves.
- Benchmark and performance figures lack independent replications in the material provided.
- Details for all ten announcements are not fully described.
- Usage limits and the exact December change are not specified.
Memory, Model Access and Usage
Taken together, the announcements point to several practical choices that will affect how organizations use AI: whether an assistant can retain information, whether a model’s weights can be obtained, and how customers pay for access. Memory may make systems more useful across ongoing work, but stored information also needs review and correction. Anthropic’s example of a lab protocol mistake underscores that human checks remain part of the process, even when a company reports faster research work.
The two model releases also show why labels alone do not settle access questions. Cognition says its coding model is based on an open-weight model but is not itself being distributed publicly; Moonshot has released weights under its own license. Users must examine the actual terms and availability. The roundup also says Google’s video model produces 15-second clips priced at $0.07 per second, while OpenAI’s image model can create animated GIFs. It reports that ChatGPT and Codex will share one allowance on Plus, Pro and Business plans starting in December. These changes affect both capability and the limits or costs users encounter.
For readers comparing products, the evidence has limits: benchmark scores, speed claims, prices and usage analyses in the roundup are often supplied by the labs themselves. They can signal what companies are building, but do not by themselves establish independent performance or value.
Products Worth Considering
Apple MacBook Pro 16-inch with M1 Pro, 10-core CPU, 16-core GPU, 16GB RAM, 1TB SSD, Silver
Apple MacBook Pro 14-inch (2021) – M1 Pro Chip, 10-Core CPU, 16-Core GPU, 16GB RAM, 1TB SSD, Space Gray
Apple MacBook Pro 14.2-inch with M5 Max, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
As an affiliate, we earn on qualifying purchases.
What the Weekly Roundup Covers
The source article, written by Thorsten Meyer and dated November 8, 2026, covers the seven days from November 2 through November 8. It says each item was checked against the relevant lab’s own post, documentation or code. The roundup names Anthropic, Cognition, Moonshot AI, Google, OpenAI, Meta and SpaceXSI, identified in the source as formerly xAI. It says three items from an initial list were removed after they did not pass the source check.
The material available here gives detailed accounts of four of the ten moves: Anthropic’s memory design and biology work, Cognition’s restricted coding-model release, and Moonshot’s Kimi weights. It also summarizes announcements involving Google video generation, OpenAI animated GIFs and a planned shared allowance for ChatGPT and Codex. It does not provide full descriptions of every item or enough detail to independently assess all seven labs’ announcements. The article says Anthropic and OpenAI did not review or sponsor its coverage.
“Seven AI labs, seven days, ten announcements.”
— ThorstenMeyerAI.com
Limits of the Reported Evidence
Independent verification is limited in the material provided. The source identifies company-generated benchmarks and tests, but does not supply independent replications for the reported memory scores, coding benchmarks, pricing comparisons or product performance. The figures should be read with their stated test context, rather than as general guarantees of results.
The roundup’s supplied text does not fully describe all 10 announcements, including the specific Meta and SpaceXSI moves, and it does not identify the three items removed from the initial list. The biology preprint mentioned by Anthropic had been submitted but was not yet public at the time of the source article. Details of Cognition’s planned trusted-partner program are also not given. The exact terms and implementation of the shared ChatGPT and Codex allowance are not specified beyond the named plans and December start.
December Usage Rules and Access
The next dated change in the roundup is the planned December start for a shared ChatGPT and Codex allowance on Plus, Pro and Business plans. The source does not state the precise date, allowance amounts or how usage will be counted, so subscribers will need the forthcoming plan details to understand the effect.
Cognition says it is working on a process to share SWE-2 Flash weights with trusted partners, but gives no timeline in the supplied account. Moonshot’s weights are already available under its own license, whose terms users should review before deployment. For the scientific work, publication of the submitted preprint would allow readers to examine the research methods and results beyond Anthropic’s account. Further independent testing will be needed to establish how broadly the reported performance figures apply.
Key Questions
What does “AI Lab Week” refer to?
It is ThorstenMeyerAI.com’s roundup of 10 announcements from seven AI labs during November 2–8, 2026, published on November 8.
Which announcements involved model weights?
Cognition announced SWE-2 Flash but said its weights would not be public yet, citing cyber risk. Moonshot AI published Kimi K3.1 Pro’s weights under its own license, not a standard open-source license.
Are the reported benchmark results independently verified?
The source describes the cited benchmark and memory figures as company-reported results. The provided material does not include independent replications.
What does Anthropic say Claude’s memory system includes?
Anthropic describes session transcripts, model-written memory notes, larger linked structures such as wikis, and a small block of context loaded at the start of a session. Its account also describes a review process called “dreaming.”
When will ChatGPT and Codex share an allowance?
The roundup says the shared allowance is planned to start in December for Plus, Pro and Business plans. It does not provide the exact start date or allowance amounts.
Source: ThorstenMeyerAI.com
