🔍 Read the full analysis: AI Lab Week By The Numbers: Seven Labs, 10 Moves on ThorstenMeyerAI.com
TL;DR
A Nov. 8 roundup from ThorstenMeyerAI.com says it verified 10 announcements from seven AI labs against company posts, documentation or code. The developments include agent-memory systems, scientific and coding tools, released and restricted model weights, longer video generation and a shared ChatGPT–Codex allowance; many performance figures are companies’ own measurements.
ThorstenMeyerAI.com published a roundup on Nov. 8 of 10 announcements from seven AI labs during the week of Nov. 2–8, spanning agent memory, scientific research, model access, video generation and product usage limits. The report says it checked each item against a lab’s own post, documentation or code, while warning that many performance numbers come from company-run tests rather than independent evaluations.
The roundup identifies agent memory as a recurring theme, citing work from Anthropic, Meta and SpaceXSI, which the source describes as formerly xAI. Anthropic’s Nov. 6 engineering post outlined four layers for Claude’s memory: session transcripts, model-written notes, linked information structures such as wikis, and a small block of information loaded at the start of each session. The system also includes a review process the authors call “dreaming,” which examines older records and can revise notes or build new structure.
Anthropic reported that memory raised Opus 5.5’s mean reward from 0.655 to 0.838 on one database task. The source notes that this is Anthropic’s own test result, not an independently established measure. It also reports a limit on the memory available at session start: Claude Code loads up to the first 200 lines or 25 KB of its memory file.
Other announcements covered different trade-offs. Cognition released SWE-2 Flash but kept its weights private for now, citing the potential misuse of its coding and security capabilities; Moonshot AI published Kimi K3.1 Pro weights under its own license. The report also describes Google video generation of 15-second clips at a stated price of $0.07 per second, OpenAI image generation that can produce animated GIFs, and a change due in December under which ChatGPT and Codex share an allowance on Plus, Pro and Business plans.
Seven labs, ten moves What the week signals
A source-checked roundup covered AI announcements from November 2–8, 2026. The developments span agent memory, research, model access, video and product usage.
Memory across projects
Agent memory could help systems carry context forward, while making review and correction important.
Layered memory
Claude’s design combines transcripts, notes, linked information and a small session-start memory.
“Dreaming” process
A review step examines older records and can revise notes or build new structure.
Access comes with trade-offs
The announcements show that released weights do not necessarily mean unrestricted use.
- Cognition released SWE-2 Flash but kept its weights private for now.
- It cited potential misuse and plans access for trusted partners.
- Moonshot AI published Kimi K3.1 Pro weights.
- They are governed by Moonshot’s own license.
Reported research and product signals
Company figures offer signals, but do not establish independent performance or research quality.
| Area | What the roundup reports | Reader context |
|---|---|---|
| Memory | Reward rose from 0.655 to 0.838 on one database task. | Anthropic’s own test result. |
| Science | About 15 billion predicted proteins searched; 44 tested. | Vault particle purpose remains unknown. |
| Usage | Data analysis: 33.7% of Claude sessions; writing: 32%. | Anthropic product usage, not research quality. |
Costs and allowances
Pricing and usage terms can shape deployment plans, but actual spending depends on use.
Google reported a price of $0.07 per second for 15-second clips.
ChatGPT and Codex allowances are due to be shared in December for Plus, Pro and Business plans.
Memory, Access and Usage Rules
The week’s announcements matter because they concern not only what AI systems can do, but also how they retain information and who can use or inspect their underlying models. Persistent memory could make agents more useful across continuing projects, while creating a need to review and correct what they store. Anthropic’s description of a memory-review step points to one approach, but the roundup does not establish how reliably it prevents errors in everyday use.
Model access is another dividing line. Cognition’s decision to restrict SWE-2 Flash’s weights, while planning access for trusted partners, contrasts with Moonshot’s publication of Kimi K3.1 Pro weights. The latter are governed by Moonshot’s own license, rather than a standard open-source license, so “weights released” does not by itself settle what users may do with the model. For organizations, the terms, hosting requirements and security risks may matter as much as benchmark results.
Pricing and usage changes could affect how businesses plan deployments. The reported video price provides a stated unit cost, while the shared ChatGPT and Codex allowance changes how eligible subscribers use those products. Neither announcement alone shows what customers will ultimately spend: actual costs depend on usage and, for performance figures, the measurements remain tied to the companies’ own reporting.
Popular right now
De'Longhi COM630BK All-in-One Coffee and Espresso Machine
KitchenAid Classic Series 4.5-Quart Tilt-Head Stand Mixer, Onyx Black
Dyson R3 Nurovi Spot+Scrub UV Robot Vacuum
As an affiliate, we earn on qualifying purchases.
What the Week’s List Covers
The report is a weekly roundup, not a single product launch or independent comparison of all the labs. It groups 10 moves across seven organizations and says it checked the listed items against primary material. Anthropic and OpenAI are among the companies covered; ThorstenMeyerAI.com says neither reviewed nor sponsored the report.
Several entries combine product news with caveats. Anthropic’s Nov. 4 account of Claude-assisted biology research described a project investigating the structure of a vault particle, a hollow protein shell whose purpose remains unknown. The team said it searched about 15 billion predicted proteins, tested 44, and found that seven of 15 proteins reported as soluble formed vault-like particles. Anthropic also listed errors, including a protocol problem that cost two weeks and a map that blended with its mirror image. The associated preprint had been submitted but was not yet public, according to the roundup.
Anthropic separately reported an analysis of 207,049 scientific sessions from August and September. It said data analysis accounted for 33.7% of Claude sessions and writing for 32%; computer science represented 32.7% of Claude Code sessions. These are usage figures from Anthropic’s own product, not a measure of research quality or of scientists’ results across the field.
„Useful as signals, not as proof.“
— ThorstenMeyerAI.com
Limits of the Available Evidence
The source material gives no independent replication of the benchmark scores, memory results or pricing comparisons attributed to the companies. Those numbers should be understood as company-reported results, with the test methods and comparison conditions not fully detailed in the supplied account.
Some developments remain incomplete or narrowly documented. Anthropic’s vault-research preprint was submitted but not publicly available, and the purpose of the vault particle is still unknown. Cognition had not yet opened SWE-2 Flash weights to the public; the timing and terms of its planned partner program were not specified. Moonshot’s weights are available under its own license, but the supplied material does not reproduce all license conditions.
The source says three items on its initial list did not survive verification, but the supplied excerpt does not identify those items or explain each omission. It also does not provide the names of all seven labs or detail every one of the 10 announcements. The claims about upcoming product terms and availability may change as companies publish further documentation.
Upcoming Access and Usage Changes
The next stated milestone is the December change to ChatGPT and Codex allowances for Plus, Pro and Business plans. The source does not specify the exact start date, allowance amounts or how usage will be counted, so subscribers will need the product’s updated terms for those details.
Cognition says it is building a program to share SWE-2 Flash weights with trusted partners, but no schedule or eligibility criteria are given in the supplied account. Anthropic’s vault study may become easier to evaluate once its submitted preprint is publicly available. More broadly, readers can look for technical documentation and independent testing to clarify how the reported scores, costs and memory features perform beyond the labs’ own reports.
Key Questions
What is the news in “AI Lab Week by the Numbers”?
ThorstenMeyerAI.com published a roundup on Nov. 8 covering 10 AI announcements from seven labs during Nov. 2–8, 2026. It groups developments involving agent memory, research, model access, video and usage limits.
Which labs are identified in the report?
The supplied source names Anthropic, Cognition and Moonshot AI, and refers to Meta, SpaceXSI (formerly xAI), Google and OpenAI. It does not provide a complete list identifying all seven labs in the excerpt.
Are the reported performance numbers independently verified?
Not in the supplied material. The report says it checked announcements against company posts, documentation or code, but identifies many scores and measurements as the labs’ own tests or claims. It describes them as signals, not proof.
What is the difference between Cognition and Moonshot’s model releases?
Cognition released SWE-2 Flash but kept its weights private, citing security concerns and saying it was developing trusted-partner access. Moonshot published Kimi K3.1 Pro’s weights under its own license, which the source says is not a standard open-source license.
What remains unclear about the announcements?
The supplied account does not give independent validation for company performance claims, full terms for every model or the exact details of the planned December allowance change. It also says three initially listed items were removed during verification without identifying them in the excerpt.
Source: ThorstenMeyerAI.com
