🔍 Read the full analysis: The AI Lab Week Briefing: 10 Moves From 2–8 November on ThorstenMeyerAI.com
Get smart everyday buys 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 November 8 briefing from ThorstenMeyerAI.com counts 10 announcements by seven AI labs during Nov. 2–8, covering agent memory, scientific work, coding models, model weights and media tools. The briefing says it checked items against lab posts, documentation or code, but many performance and cost figures come from companies’ own tests; three items on its initial list were excluded after verification.
ThorstenMeyerAI.com published a roundup on Nov. 8 of 10 announcements from seven AI labs during the week of Nov. 2–8, describing developments in agent memory, scientific research, model access, video and image generation, and subscription usage limits. The briefing says each included item was checked against a lab post, documentation or code, while warning that many performance figures are the companies’ own measurements.
The briefing’s central theme is agents designed to retain information across sessions. It describes memory-related work from Anthropic, Meta and SpaceXSI, identified in the source as formerly xAI. Two of the approaches include a review step for deciding what information to retain. The source does not provide enough detail here to compare the three systems directly.
Anthropic’s Nov. 6 engineering post, “The components of agent memory,” describes four layers: conversation transcripts, notes written by the model, connected knowledge structures such as wikis, and a small context block loaded at the start of each session. The authors describe a process called “dreaming” that reviews older logs, corrects notes and creates new structures. Anthropic reports that memory raised Opus 5.5’s mean reward from 0.655 to 0.838 on one database task; the briefing identifies that result as the company’s own test, not an independent evaluation.
Anthropic also described a biology project in a Nov. 4 post. Its team searched about 15 billion predicted proteins for relatives of known vault proteins, then tested 44 candidates. The lab reported that seven of 15 proteins that became soluble formed vault-like particles. The project took under four months, according to Anthropic. The post also disclosed a protocol error that cost two weeks and a map blended with its mirror image, which researchers caught. The related preprint had been submitted but was not yet public when the briefing was published.
On coding models, Cognition announced SWE-2 Flash on Nov. 4. The company says it built the model on Inkling-Small, an open-weight model from Thinking Machines, and reported higher results on its FrontierCode 2.0 and ExploitBench evaluations after additional training. Cognition cited the model’s cyber capabilities as a reason it is not publishing the weights publicly yet; it said it is working on access for trusted partners. The model was offered free in Devin through Nov. 30, according to the briefing.
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. The briefing says the release uses Moonshot’s own license rather than a standard open-source license and notes that the weights had not been available when the model launched in October.
The remaining items in the briefing include longer, lower-cost video generation from Google, animated GIF creation with OpenAI’s image model, and a forthcoming change to ChatGPT and Codex usage allowances. The source says Plus, Pro and Business subscribers will share one allowance across ChatGPT and Codex starting in December. It does not give the full terms or precise launch date in the supplied material.
10 AI lab moves in one week
From agent memory to model access and media tools, seven labs made announcements during Nov. 2–8. The roundup checked items against primary sources and flags that many figures come from company tests.
What changed
The announcements span tools, research, access and subscription rules.
Memory across sessions
Anthropic, Meta and SpaceXSI described memory work. Anthropic outlined transcripts, model-written notes, knowledge structures and session context.
Protein research
Anthropic reported searching predicted proteins and testing candidates for vault-like particles.
Different access choices
Cognition is withholding SWE-2 Flash weights while Moonshot published Kimi K3.1 Pro weights under its own license.
More formats
The roundup also covers longer, lower-cost video generation from Google and animated GIF creation with OpenAI’s image model.
How access affects users
Availability depends on more than a model’s name or benchmark score.
- Cognition cites cyber capabilities as a reason for restricting public access.
- It says it is working on access for trusted partners.
- Moonshot released Kimi K3.1 Pro weights on Hugging Face.
- Its model uses Moonshot’s own license, so terms matter for deployment.
What the evidence can tell you
Reported results are useful signals, but the briefing does not establish how they generalize.
The roundup says it checked included items against lab posts, documentation or code.
Many performance, speed and cost figures are company-reported; independent replications were not included.
The December shared ChatGPT and Codex allowance is described without complete limits or a precise launch date.
Memory, Access and Usage Rules
The announcements point to practical questions for people choosing or deploying AI tools: what an agent remembers, who can use a model, and how usage is metered. Persistent memory could let an assistant carry project information between sessions, but it also creates a need to correct outdated or inaccurate notes. Anthropic’s description of a review process and its limit on how much memory Claude Code loads illustrate some of the engineering constraints; the briefing does not establish how reliably these methods work across tasks.
The contrasting decisions by Cognition and Moonshot show that access labels alone do not tell users what they can do. Cognition says it is withholding weights because of security concerns, while Moonshot has released weights subject to its own license. For organizations weighing self-hosting against commercial services, license terms, access conditions and independently verified costs matter alongside benchmark scores. The briefing’s figures are useful as reports of what companies say, but they do not establish that another user will see the same speed, price or performance.
Longer video generation and animated image formats expand the types of media these tools can produce. The shared ChatGPT and Codex allowance, if implemented as described, could also change how subscribers plan usage across products. Readers will need the actual service terms to understand how the change affects their accounts.
Products Worth Considering
Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 256GB SSD, Liquid Retina Display, and 1080p Camera – Blush
Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight
Dell UltraSharp U3824DW 37.5-Inch Curved WQHD+ IPS Black USB-C Hub Monitor
As an affiliate, we earn on qualifying purchases.
How the Weekly List Was Checked
The briefing is a weekly recap, not a single product launch. It says it checked every included item against a primary source such as a lab’s own post, documentation or code. Its author also says three items on an initial list did not survive verification and were removed. The supplied material does not identify those three items or explain each exclusion, so they cannot be assessed from this account.
The source names Anthropic and OpenAI among the companies covered and says neither reviewed nor sponsored the briefing. It also links to a video explainer, described as a hand-drawn summary of the 10 developments. The article’s broad picture—memory, scientific assistance, model access, media generation and usage rules—comes from grouping separate announcements published over seven days, rather than from a joint statement or coordinated industry plan.
Several numerical claims are explicitly attributed to the labs themselves. Anthropic’s memory score and biology findings, Cognition’s benchmark and pricing comparisons, and Moonshot’s benchmark table are company-reported. The briefing cautions that such results should be treated as signals rather than proof of performance for other tasks or customers. Its account does not include independent replications or full test methods for those claims.
““Most numbers are self-reported.””
— Thorsten Meyer, ThorstenMeyerAI.com briefing
Limits of the Reported Evidence
The briefing does not provide independent testing of the performance, speed or cost claims. Anthropic’s memory score and biology findings, Cognition’s benchmark results, and Moonshot’s benchmark table are presented as company-reported evidence. The supplied material does not include enough methodology to determine how comparable the tests are or whether the results generalize to other users and tasks.
The biology preprint was submitted but not public at the time of publication, limiting outside review of the research. The source also says an outside vault expert advised the project as a paid consultant. It does not provide further detail about that relationship or the expert’s role.
The account mentions memory work from Meta and SpaceXSI but does not explain their systems in the same detail as Anthropic’s. It also does not identify the three initial-list items removed after checking, or provide the complete underlying documentation for all 10 announcements. The December shared allowance is described without full account limits or implementation details.
What to Watch in December
The clearest stated upcoming milestone is the December change to ChatGPT and Codex allowances for Plus, Pro and Business plans. Users will need to watch for OpenAI’s full terms, including the size of the shared allowance and how usage in each product will count against it.
Other developments to follow include whether Cognition opens its planned trusted-partner program for SWE-2 Flash weights, and whether the submitted vault-protein preprint becomes publicly available for outside scrutiny. Additional technical documentation or independent evaluations could help clarify how well the reported memory, coding and scientific results hold up beyond the labs’ own tests.
For buyers and developers, the immediate next step is to check the relevant model cards, licenses and product documentation before relying on a benchmark, access label or quoted price. The weekly briefing records what labs announced during Nov. 2–8; it does not establish that every product feature is broadly available or that reported results will persist in other settings.
Key Questions
What is the AI Lab Week Briefing?
It is a Nov. 8 recap from ThorstenMeyerAI.com covering 10 announcements from seven AI labs between Nov. 2 and Nov. 8, 2026. The briefing says it checked included items against lab posts, documentation or code.
Which development did the briefing identify as a major theme?
Agent memory was a leading theme. The source describes memory-related work from Anthropic, Meta and SpaceXSI, though it gives the most technical detail about Anthropic’s system.
Are the performance figures independently verified?
Not in the material provided. The briefing says many figures, including benchmark results and memory scores, are the labs’ own measurements. It does not report independent replication of those results.
Are Kimi K3.1 Pro’s weights released under a standard open-source license?
No, according to the briefing. Moonshot AI published the weights on Hugging Face, but the source says they use Moonshot’s own license, not a standard open-source license. Users should check the model’s license for applicable conditions.
When will ChatGPT and Codex share one usage allowance?
The briefing says the shared allowance is planned to begin in December for Plus, Pro and Business plans. It does not give a precise date or the full usage terms.
Source: ThorstenMeyerAI.com
