🔍 Read the full analysis: A Week In AI: 10 Moves By Anthropic, OpenAI, Google And Others on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 2–8 roundup from ThorstenMeyerAI.com identifies 10 announcements from seven AI labs, including work on agent memory, AI-assisted biology, coding models, open weights and media tools. The source says it checked each item against company posts, documentation or code, but many performance and cost figures are companies’ own measurements, not independent evaluations.
Seven AI labs made 10 announcements during the week of November 2–8, covering agent memory, scientific research, coding models, model weights, video and product limits, according to a roundup published November 8 by ThorstenMeyerAI.com. The developments offer a snapshot of competing priorities across the industry, but many reported performance, price and productivity figures are companies’ own claims, not independently verified results.
The roundup’s clearest shared theme is AI systems that retain information between tasks. Anthropic described a four-layer memory design for Claude, spanning session transcripts, model-written notes, linked knowledge structures and a small block of context loaded at the start of a session. The post also describes a review process called “dreaming,” in which the system revisits older records to correct notes and organize information. Anthropic reported that memory improved the mean reward on one database task for Opus 5.5 from 0.655 to 0.838; the figures came from the company’s own tests.
Anthropic also reported on Claude’s use in a biology project investigating vault particles, hollow protein structures whose purpose remains unknown. The company said researchers searched about 15 billion predicted proteins, tested 44 candidates and found that seven of the 15 soluble proteins formed vault-like particles. The work took less than four months, according to the post, which also described a protocol error that cost two weeks and a map researchers had to correct after it blended with its mirror image. The related research preprint had been submitted but was not yet public as of the roundup.
Other announcements addressed who can access model weights and how AI tools are used. Cognition released SWE-2 Flash, built on Thinking Machines’ open-weight Inkling-Small model, but kept its own trained weights private while it develops a trusted-partner access program. Moonshot AI, by contrast, published Kimi K3.1 Pro’s weights on Hugging Face under its own license. The roundup also said Google’s video model offered 15-second clips at $0.07 per second, OpenAI’s image model could make animated GIFs, and ChatGPT and Codex would share one usage allowance on Plus, Pro and Business plans from December.
A Week in AI 10 moves across seven labs
From agent memory to model access and media tools, the week’s announcements show competing priorities. Many performance and cost figures remain company-reported.
What labs are building
The announcements span systems that remember, research tools, coding models and media products.
Agents that retain context
Anthropic described layered memory for Claude, including a review process for older records.
AI-assisted biology
Anthropic reported using Claude in work investigating hollow vault particles.
Coding and media tools
The roundup also covers coding models, animated GIFs, video clips and shared usage limits.
Model access has trade-offs
The word “open” does not describe every part of how a model can be used.
- Released SWE-2 Flash, built on an open-weight model.
- Kept its own trained weights private while planning trusted-partner access.
- Published Kimi K3.1 Pro weights on Hugging Face.
- The weights use Moonshot AI’s own license.
A research example, with checks needed
Anthropic’s account describes both promising work and the need for human review.
Researchers searched about 15 billion predicted proteins.
They tested 44 candidates; seven of 15 soluble proteins formed vault-like particles.
The account described a protocol error and a map that needed correction. The preprint was not yet public.
How to read the claims
Company figures can signal what labs are improving, but the roundup reports limited independent evaluation.
| Evidence | What readers can conclude |
|---|---|
| Company benchmarks and product figures | Reported results; methods are not fully described for several figures. |
| Source checks against posts, documentation or code | Announcements were checked against primary materials, but this is not an independent audit. |
| Biology findings | Human review remains important; the submitted preprint was not yet public. |
Memory and access shape AI products
Taken together, the announcements point to a practical contest over how AI products fit into ongoing work: whether an agent can remember useful information, whether a model can help with specialized research, and whether customers can run model weights themselves or must use a hosted service. These choices affect continuity, cost, oversight and control for organizations adopting AI.
The contrast between Cognition and Moonshot also shows why model labels need careful reading. A system can be based on an open-weight model while its own trained weights remain restricted; a published model may still use a license that is not a standard open-source license. Availability alone does not settle what users may do with a model or how much computing and expertise self-hosting requires.
The claims also deserve different levels of confidence. Company-reported benchmark scores, savings and speed comparisons can indicate what a lab is trying to improve, but they do not establish independent performance. Anthropic’s biology account is similarly an example of AI-assisted work, not proof that the system can conduct research reliably without human review. Its reported mistakes underline the need for researchers to check outputs and procedures.
Products Worth Considering
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Sonos Era 100 SL Compact Microphone-Free Speaker, Black
Samsung Odyssey G93SC 49-Inch QD-OLED Curved Gaming Monitor
As an affiliate, we earn on qualifying purchases.
The week behind the roundup
The roundup covers announcements made between November 2 and 8, 2026, and says it checked items against each lab’s own post, documentation or code. It names Anthropic, OpenAI, Google, Meta, SpaceXSI, Cognition and Moonshot AI among the organizations involved. Its account is a secondary roundup, rather than an independent audit of every product or experiment.
Anthropic’s memory post and biology report were dated November 6 and November 4, respectively. Cognition’s SWE-2 Flash announcement and Moonshot’s publication of Kimi K3.1 Pro weights were both dated November 4. The source says Meta and SpaceXSI also published work on agent memory, with two of the three memory efforts incorporating a review step for stored information. It does not provide enough detail in the supplied material to assess those systems or compare their results.
ThorstenMeyerAI.com says it removed three items from its initial list after checking them against primary sources. The supplied account does not identify those items or provide details about why each failed verification, limiting what readers can assess about the screening process.
““dreaming””
— Anthropic, in its engineering post “The components of agent memory”
Independent checks remain limited
Most performance and business figures in the roundup are self-reported. The supplied material provides no independent assessment of Anthropic’s memory benchmark, Cognition’s coding and security scores, Google’s video pricing, or the claimed savings and speed advantages. The comparison methods and conditions are not fully described for several figures.
Anthropic said the biology preprint had been submitted but was not yet public; its findings therefore could not be assessed from that paper at the time covered. The source also says an outside vault expert advised the project as a paid consultant. It does not identify that expert in the supplied text or give further details about the review process. The outcomes and terms of the planned trusted-partner program for Cognition’s weights are also not specified.
Finally, although the roundup says three initial entries were dropped after verification, it does not explain which claims were removed. The full set of 10 moves is not described in equal detail in the supplied material, so some developments cannot be evaluated here beyond the brief summary.
December usage rules and model access
The nearest dated change in the roundup is OpenAI’s planned December shared allowance for ChatGPT and Codex on Plus, Pro and Business plans. The source does not specify the exact start date, how the combined allowance will be calculated, or how customers will see remaining usage.
Cognition said it was developing a program to make SWE-2 Flash’s weights available to trusted partners, but gave no timetable or eligibility rules in the supplied account. Anthropic’s biology preprint may provide more information once public. Further independent evaluations would be needed to test the reported model scores, product costs and research claims beyond the companies’ own accounts.
Key Questions
What is the main development in this roundup?
ThorstenMeyerAI.com summarized 10 AI announcements from seven labs during November 2–8, 2026, including new work on agent memory, scientific research, model access and media tools.
Which companies released or restricted model weights?
Moonshot AI published Kimi K3.1 Pro’s weights under its own license. Cognition released SWE-2 Flash but kept its trained weights private while it planned a trusted-partner program.
Are the performance figures independently verified?
Not in the supplied material. The roundup says many figures, including benchmark results and cost or speed comparisons, are the companies’ own measurements; it presents them as signals rather than independent proof.
What did Anthropic report about Claude’s biology work?
Anthropic said a research team used Claude while investigating vault particles, and reported that seven of 15 soluble candidate proteins formed vault-like particles. The company also described errors researchers caught, and said the preprint had been submitted but was not yet public.
When will ChatGPT and Codex share an allowance?
The roundup says the shared allowance is planned to begin in December for Plus, Pro and Business plans. It does not provide an exact date or details about how usage will be counted.
Source: ThorstenMeyerAI.com
