🔍 Read the full analysis: The Week In AI Labs: 10 Moves From Anthropic To Cognition on ThorstenMeyerAI.com
Get audio and creator gear 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 Nov. 8 roundup from ThorstenMeyerAI.com describes 10 announcements by seven AI labs during the week of Nov. 2–8, covering agent memory, scientific research, coding models, model weights, media tools and usage limits. The source says it checked items against lab posts, documentation or code, while noting that most performance figures are companies’ own measurements and three initially considered items did not pass its checks.
ThorstenMeyerAI.com published a roundup on Nov. 8 of 10 announcements from seven AI labs during the week of Nov. 2–8, highlighting developments in agent memory, scientific work, coding models, model-weight access, media generation and usage limits. The publication says it checked each item against a lab post, documentation or code, while cautioning that many performance figures are companies’ own measurements.
The roundup identifies agent memory as a recurring theme. Anthropic described a four-layer approach for Claude that includes session transcripts, model-written memory notes, linked knowledge structures and a compact context block loaded at the start of a session. Its post also describes a review process called “dreaming” for revising older notes. Anthropic reported that memory raised Opus 5.5’s mean reward from 0.655 to 0.838 on one database task; those figures come from the company’s test, not an independent evaluation.
Anthropic also reported using Claude during work on the structure of a bacterial vault particle. The company said its team searched about 15 billion predicted proteins, tested 44 candidates and found that seven of 15 soluble proteins formed vault-like particles. Its account describes setbacks, including a protocol error and a map that blended the structure with its mirror image. The preprint was submitted but not yet public as of the roundup. A separate Anthropic analysis covered 207,049 scientific sessions from August and September, according to the source.
On coding, Cognition announced SWE-2 Flash, built on Thinking Machines’ Inkling-Small open-weight model. Cognition said additional training improved its FrontierCode 2.0 score from 3.2% to 24.0%, and its ExploitBench score from 11.3% to 34.8%. It said it would keep the model weights private for now because of potential cyber risks, while developing access for trusted partners. Moonshot AI, by contrast, published the weights for Kimi K3.1 Pro on Hugging Face. Its model card lists 2.8 trillion total parameters, 104 billion active parameters and a context window of about one million tokens; the source notes that Moonshot’s own license applies.
The Week in AI Labs 10 moves from Anthropic to Cognition
Seven AI labs made announcements across agent memory, scientific research, coding models, model access, media tools and usage limits during Nov. 2–8, 2026.
Memory and science
Anthropic described persistent memory for Claude and reported using it in scientific research.
Notes across sessions
A four-layer system includes transcripts, memory notes, linked knowledge and a compact session context. Anthropic reported a higher reward on one database task in its own test.
Vault particle work
Anthropic reported testing 44 candidates; seven of 15 soluble proteins formed vault-like particles. The submitted preprint was not yet public.
Model access takes different paths
The announcements leave developers weighing access, licensing and security.
- Cognition says cyber risks are why SWE-2 Flash weights remain private.
- The company says it is developing access for trusted partners.
- Moonshot AI published Kimi K3.1 Pro weights on Hugging Face.
- Its model card lists 2.8 trillion total parameters and a context window of about one million tokens; its own license applies.
What readers can watch
The roundup says its items were checked against lab posts, documentation or code, while several claims still need independent scrutiny.
- Look for independent assessment of company-reported benchmark results.
- Check the model license and access terms before planning a deployment.
- Review published terms for the shared ChatGPT and Codex allowance due to start in December 2026.
- Watch for the public release of Anthropic’s submitted preprint.
Memory and Model Access Take Different Paths
The week’s announcements point to different choices about how AI capabilities are developed and distributed. Persistent memory could help agents carry information across sessions, but it also makes the accuracy and maintenance of stored notes consequential. Anthropic’s account describes a review mechanism, while the company’s reported benchmark result remains limited to its own test.
The contrast between Cognition and Moonshot AI is also relevant to developers and organizations weighing whether to deploy models themselves. Cognition says cyber concerns justify restricting access to SWE-2 Flash’s weights; Moonshot has released Kimi K3.1 Pro weights under its own license. Neither “open-weight” label alone settles what users may do: access terms, infrastructure costs and security considerations differ. The roundup’s performance and pricing comparisons should be treated as company-reported claims unless independently verified.
Media tools and product limits affect a different group of users. The source says Google’s video model supports 15-second clips at a listed price of $0.07 per second, while OpenAI’s image model can produce animated GIFs. It also reports that ChatGPT and Codex will share an allowance on Plus, Pro and Business plans starting in December. These changes may affect creative workflows and usage planning, though the supplied material does not include full product terms.
Products Worth Considering
Apple iPad 11-inch with A16 Chip, 512GB, Wi-Fi 6, and Apple Pencil (USB-C) — Blue
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Black
As an affiliate, we earn on qualifying purchases.
Seven Labs, Ten Checked Announcements
The roundup covers the week of Nov. 2–8, 2026, and says its items were checked against primary materials such as company posts, documentation or code. It names Anthropic, Cognition and Moonshot AI among the companies covered, and describes activity from seven labs overall. The source also says three items on its initial list were removed after they did not survive its checks, although the supplied text does not identify those items.
Some of the figures are specific to company tests or claims. Cognition’s coding benchmarks, Anthropic’s memory result and Moonshot’s benchmark table are not presented in the source as independent evaluations. The roundup also cites a Revolut conference talk about a separate company’s self-hosted model deployment and costs, but that account is attributed to Revolut rather than independently verified here. These distinctions matter when comparing performance, cost or reliability across products.
“Memory was the theme.”
— ThorstenMeyerAI.com
Benchmarks and Access Still Need Scrutiny
Most reported performance figures in the roundup are self-reported by the labs; the supplied material does not provide independent replication or enough methodology to compare results directly. It is also not clear from the source when Cognition’s trusted-partner program will open, what eligibility rules it will use, or whether the model’s weights will later become broadly available.
Moonshot’s weights are available under its own license, but the source does not spell out the license’s full restrictions. Anthropic’s vault-particle preprint had been submitted but was not yet public, and the source does not report subsequent peer review or independent confirmation of the findings. The roundup says three initial items were dropped after verification, without naming them in the material provided. Further details about Google’s and OpenAI’s media offerings and the shared ChatGPT-Codex allowance are also not included here.
Watch for Papers, Terms and Limits
Readers can look for the public release of Anthropic’s submitted vault-particle preprint and for any independent assessment of the work. Cognition’s next stated step is a program for sharing SWE-2 Flash weights with trusted partners; timing and conditions remain unspecified in the source. Moonshot’s model card and license are the primary references for users considering a self-hosted Kimi deployment.
The roundup says ChatGPT and Codex will begin sharing one allowance on Plus, Pro and Business plans in December 2026. Users will need the product’s published terms to understand how the shared limit is counted. Across the announcements, further documentation and independent testing will determine how far company-reported benchmarks, cost claims and research results generalize.
Key Questions
What is the main development in the roundup?
ThorstenMeyerAI.com summarized 10 announcements from seven AI labs during Nov. 2–8, 2026, spanning agent memory, scientific research, coding models, model access, media generation and usage limits.
Did Anthropic independently prove its memory benchmark result?
No independent verification is identified in the supplied material. Anthropic reported that memory raised Opus 5.5’s mean reward from 0.655 to 0.838 on one database task, and the roundup labels the score as the company’s own test.
Why is Cognition keeping SWE-2 Flash’s weights private?
Cognition said the model’s coding and security capabilities could also help attackers. It said it was developing a program to share the weights with trusted partners, but the source gives no launch date or access criteria.
Are Kimi K3.1 Pro’s weights released under a standard open-source license?
The roundup says Moonshot AI published the weights on Hugging Face under Moonshot’s own license, not a standard open-source license. Users should consult that license for the applicable terms.
When will ChatGPT and Codex share an allowance?
The roundup says the shared allowance for Plus, Pro and Business plans is due to start in December 2026. It does not provide the detailed counting rules or plan-specific limits.
Source: ThorstenMeyerAI.com
