🔍 Read the full analysis: 10 Moves Across AI: A Week With Anthropic, OpenAI And More 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 Nov. 2–8 roundup published by ThorstenMeyerAI.com tracks ten AI-related developments across seven labs, including work on agent memory, scientific research, model weights, video and usage rules. The source says it checked announcements against company posts, documentation or code, but many performance and cost figures remain company-reported, and three items from its initial list were excluded after verification.
The developments span products, research and access policies rather than one shared release. Anthropic’s Nov. 6 engineering post described a four-part system for carrying information between Claude sessions: conversation transcripts, model-written memory notes, linked structures such as wikis, and a compact context block loaded at the start of a session. The post also describes a review process called “dreaming,” which revisits older records to correct notes and build new connections. Anthropic reported that memory raised Opus 5.5’s mean reward from **0.655 to 0.838 on one database task**; the result is from the company’s own test.
Anthropic also described using Claude in a biology project to investigate the structure of a bacterial vault particle. According to its account, the team searched roughly **15 billion predicted proteins**, tested 44 candidates and found that seven of 15 soluble proteins formed vault-like particles. The work took less than four months, the company said, but its post also disclosed a protocol mistake that cost two weeks and a map that blended with its mirror image. The associated preprint had been submitted but was not yet public when the roundup was published.
Model access went in different directions. Cognition announced SWE-2 Flash, a coding model whose weights it has not made publicly available, citing potential security risks; the company said it plans to share them with trusted partners. Moonshot AI, by contrast, published Kimi K3.1 Pro’s weights 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 roundup notes that Moonshot’s license is its own, rather than a standard open-source license.
Other items included Google’s video model producing 15-second clips at a reported **$0.07 per second**, OpenAI’s image model generating animated GIFs, and a planned shared usage allowance for ChatGPT and Codex. The source says the shared allowance is due to start in December for Plus, Pro and Business plans. It also covers memory work from Meta and SpaceXSI, identified in the source as formerly xAI. These details come from the supplied roundup; it does not provide full underlying documentation for every item.
Ten AI moves in one busy week
A source-checked roundup tracks announcements across AI labs, from agent memory and research to model access, media tools and usage rules.
Memory moves into the workflow
Persistent memory could reduce repeated setup, while making review and correction important.
Carry context forward
Claude’s described system combines transcripts, memory notes, linked records and a compact session context.
Progress with caveats
Anthropic reported biology research progress and also disclosed a protocol mistake and a map that blended with its mirror image.
Model access takes different paths
Availability, security and license terms shape what organizations can do with model weights.
- Cognition is withholding SWE-2 Flash weights, citing security risks.
- It plans to share them with trusted partners; timing and eligibility are unspecified.
- Moonshot published Kimi K3.1 Pro weights on Hugging Face.
- Its model card lists a company-specific license, so check terms before use or redistribution.
Products and usage rules shift
The roundup also covers video, image generation and changes that may affect everyday workflows.
Google’s video model and OpenAI’s image model add video and animated GIF capabilities.
ChatGPT and Codex usage is planned to share an allowance starting in December for Plus, Pro and Business plans.
The roundup gives no exact allowance amounts, and a price per second alone does not establish total production cost.
Memory, access and usage rules shift
Taken together, the announcements point to several practical decisions for users and organizations: whether agents can retain information between sessions, who can inspect or run model weights, and how usage is metered. Persistent memory may reduce repeated setup, but it also creates a need to review what a system stores and correct errors. Anthropic’s account of the biology work illustrates both sides: it reports progress, while also describing mistakes that researchers had to catch.
The contrasting releases from Cognition and Moonshot show that labels such as “open-weight” do not settle questions of access. Cognition said it is restricting its model’s weights because of security concerns; Moonshot released weights under a company-specific license. For teams choosing a model, the terms governing use and redistribution can matter as much as parameter counts or benchmark results.
Changes to pricing and shared usage limits can affect operating costs and workflows, but the figures in this roundup need careful reading. Company tests are not independent comparisons, and a stated price per second does not by itself establish the full cost of a production workflow. The roundup offers a snapshot of activity, not evidence that one approach is best.
Popular right now
Ninja Flip Toaster Oven and Air Fryer, 8-in-1, Stainless Steel, SP151
De'Longhi COM530M All-in-One Coffee and Espresso Machine, 47 oz
Dyson Cool AM07 Bladeless Tower Fan, Iron/Blue
As an affiliate, we earn on qualifying purchases.
A week of releases, checked at source
The roundup is dated **Nov. 8, 2026** and covers announcements made during the preceding seven days. Its author says every included item was checked against a lab’s own post, documentation or code. The source also says three candidates from its initial list were removed because they did not pass that check, though the supplied material does not identify those three items.
Several reported figures are explicitly attributed to the companies behind the products. Anthropic’s memory score and research account, Cognition’s benchmark results and performance comparisons, and Moonshot’s benchmark table are not presented as independent testing. The source also summarizes an Anthropic analysis of 207,049 scientific sessions from August and September, reporting that data analysis accounted for 33.7% of Claude sessions and writing for 32%; those figures describe Anthropic’s own usage data.
The supplied text begins a comparison involving Revolut’s reported use of a self-hosted model but cuts off before completing it. That incomplete passage is not needed to establish the week’s central development and cannot support a full account of the comparison.
Independent tests and access remain open
The roundup does not independently verify the companies’ performance, cost or benchmark claims. It cautions that figures such as model scores and reported savings reflect company measurements, so readers cannot treat them as directly comparable without shared methods and baselines.
Several details remain unsettled or absent. Anthropic’s vault research preprint was submitted but not public at the time of the source report, and the source does not specify when it might appear. Cognition’s trusted-partner program for SWE-2 Flash weights is still being built; eligibility and timing are not given. Moonshot’s model card lists a company-specific license, but the supplied material does not detail all of its terms. The source also does not provide the full rules or exact allowance amounts for the planned ChatGPT and Codex usage change.
Finally, the source refers to ten announcements but the supplied text gives detailed accounts of only some of them. It names additional memory work by Meta and SpaceXSI and mentions OpenAI and Google media features, but offers limited supporting detail for those items here. The three rejected candidates are not identified.
December limits and research updates
The clearest dated next step in the roundup is the planned **December start** of a shared ChatGPT and Codex usage allowance for Plus, Pro and Business subscribers. The source does not state the allowance size or whether users will receive more operational detail before the change takes effect.
Further updates could clarify access to Cognition’s weights, the terms attached to Moonshot’s release and the publication status of Anthropic’s vault research. Those are possible areas for follow-up, not confirmed schedules. Readers should look to the companies’ documentation and subsequent announcements for any changes to availability, licensing or product limits.
For now, the roundup’s main value is as a record of a busy week, with an explicit distinction between source-checked announcements and company-reported results. It does not establish that the models’ reported scores, costs or research outcomes have been independently replicated.
Key Questions
What did the Nov. 2–8 AI roundup report?
ThorstenMeyerAI.com summarized **ten announcements from seven labs**, including work on agent memory, a biology research project, model-weight releases and restrictions, video features, and a planned shared ChatGPT and Codex usage allowance.
Are the performance and cost figures independently verified?
Not according to the supplied source. It says many figures, including benchmark scores and cost claims, are **company-reported measurements** and should not be treated as independent results.
Did Moonshot AI release Kimi K3.1 Pro’s weights?
The roundup says Moonshot published the weights on Hugging Face on **Nov. 4, 2026**. It also says the model is distributed under Moonshot’s own license, not a standard open-source license.
Why did Cognition keep SWE-2 Flash’s weights private?
Cognition cited potential security risks, saying the model’s capabilities could also help attackers. The company said it was developing a program to share the weights with trusted partners; the roundup gives no launch date or eligibility rules.
When does the shared ChatGPT and Codex allowance begin?
The source says the shared allowance is planned to begin in **December** for Plus, Pro and Business plans. It does not give the allowance size or further details about how the limit will work.
Source: ThorstenMeyerAI.com
