🔍 Read the full analysis: 10 AI Updates From Anthropic, OpenAI, Google, Meta And More on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 8 roundup from ThorstenMeyerAI.com describes 10 announcements from seven AI labs during the week of November 2–8, 2026. The reported developments span agent memory, scientific research, coding models, open-weight releases, video and image tools, and changes to ChatGPT and Codex usage limits; many performance figures are company-reported.
ThorstenMeyerAI.com published a roundup on November 8 covering 10 announcements from seven AI labs during the preceding week, including new work on agent memory, scientific research, model access, and generative media. The report says the developments point to a shift toward agents that retain information across sessions, while also highlighting disputes over model weights and changes to how some users access AI services.
Among the reported updates, Anthropic described a four-layer memory system for Claude agents in a November 6 engineering post. The layers include conversation records, model-written memory notes, linked information structures such as wikis, and a small context block loaded at the start of a session. Anthropic also described a process it calls “dreaming” that reviews prior logs and updates memory. In a company-run database task, Anthropic reported that memory increased Opus 5.5’s mean reward from 0.655 to 0.838; the roundup cautions that this result comes from Anthropic’s own testing.
Anthropic also reported a biology project in which researchers used Claude while studying vault particles, hollow protein structures whose purpose remains unknown. The team searched about 15 billion predicted proteins for relatives of known vault proteins, then tested 44 candidates. According to Anthropic, seven of 15 proteins that were soluble formed vault-like particles. The project took less than four months, but the company’s account also describes a protocol error that cost two weeks and a map that blended with its mirror image. The related preprint had been submitted but was not yet public, according to the roundup.
Other model releases drew attention to competing approaches to access. Cognition released SWE-2 Flash on November 4, built on Thinking Machines’ Inkling-Small open-weight model, but said it would not publish the resulting model weights yet because of security concerns. Moonshot AI, by contrast, published the weights for Kimi K3.1 Pro on Hugging Face the same day. 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 uses its own license rather than a standard open-source license.
10 AI updates shaping access and agents
A November roundup tracks developments in agent memory, research, model access, generative media and product limits. Many performance figures come from the companies themselves.
Agents that remember
Anthropic described a layered memory system for Claude agents, including notes and linked information that can carry context between sessions.
Conversation records preserve prior interactions.
A process called “dreaming” reviews logs and revises memory.
A small context block is loaded when a session starts.
Open weights, different terms
Model availability affects who can run, modify and assess a release.
- Released SWE-2 Flash, built on an open-weight model.
- Withheld its weights, citing security concerns.
- Published Kimi K3.1 Pro weights on Hugging Face.
- Uses a company-specific license.
Research and media
The announcements also span scientific work, video generation and image formats.
Vault particles
Anthropic described a biology project using Claude; the particles’ purpose remains unknown.
Google generation
The roundup reports 15-second clips at $0.07 per second.
Animated GIFs
OpenAI image generation was reported to support animated GIFs.
Usage rules are changing
Subscribers may need to divide usage across products under a shared allowance.
Memory, Access and Usage Rules
The week’s announcements matter because they show that competition is about more than model capability. Persistent agent memory could let AI systems carry project information between sessions, but stored notes can also be wrong or outdated. Anthropic’s account of errors in its biology work underscores the need for people to verify outputs, even when a system accelerates research.
Access rules are another dividing line. Cognition says it is withholding SWE-2 Flash’s weights because its security performance could also help attackers, while Moonshot has released Kimi K3.1 Pro’s weights under a company-specific license. That difference affects who can run a model, modify it, and assess its terms. A model described as built on open weights is not necessarily itself available for unrestricted use.
The roundup also reports that ChatGPT and Codex will share one usage allowance on Plus, Pro and Business plans starting in December. If implemented as described, that would change how subscribers divide usage across the two products. Other reported media updates include Google video generation of 15-second clips at $0.07 per second and OpenAI image generation that produces animated GIFs. These product and pricing details are attributed to the source report, not independently established here.
Popular right now
Bose QuietComfort Ultra Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Black
Roborock Q7 M5+ Robot Vacuum and Mop with Self-Emptying Dock
Samsung 75-Inch QLED Q7F Series Smart TV (2025)
As an affiliate, we earn on qualifying purchases.
What the Week’s Releases Cover
The report groups announcements made from November 2 through November 8, 2026, rather than documenting one coordinated industry event. Its opening summary says the 10 items came from seven labs and were checked against company posts, documentation or code. It also says three items on an earlier draft list were removed after they did not pass that source check.
Alongside the releases described above, Anthropic published an analysis of 207,049 scientific sessions from August and September. The company said data analysis accounted for 33.7% of Claude sessions and writing for 32%; computer science accounted for 32.7% of Claude Code sessions. These figures describe Anthropic’s own product activity and should not be read as a measure of scientific work across the industry.
The roundup frames the announcements as signals of broader product choices: longer-lived agents, tools for research and coding, more media formats, and different limits on model access. It also stresses that many reported performance measures, including benchmark results and memory scores, are company-reported figures, not independent evaluations.
“Memory was the theme.”
— ThorstenMeyerAI.com
Claims Still Need Independent Checks
The roundup does not provide independent verification of the companies’ benchmarks, cost comparisons or performance claims. Cognition’s reported benchmark gains and Moonshot’s model-card results are company figures. The account also cites a company-reported claim that SWE-2 Flash runs at one-fifth the price and three times the speed of Cognition’s larger SWE-2; the test conditions and comparison methodology are not detailed in the supplied material.
Several details remain unresolved. Anthropic’s vault research preprint was not public when the roundup was written, and the purpose of the vault particles is still unknown. Cognition said it was developing a trusted-partner program for sharing SWE-2 Flash weights, but the report does not specify when it will open or who will qualify. Moonshot’s custom license also means the practical permissions depend on its terms, not simply on the availability of downloadable weights.
The supplied source material gives only a partial account of the 10 announcements: it names Google’s video and OpenAI’s GIF updates, and references changes to ChatGPT and Codex allowances, but does not provide complete supporting details for every item or identify all seven labs. The full scope and terms of those updates therefore cannot be established from the material provided.
Upcoming Access and Product Changes
The nearest stated milestone is December 2026, when ChatGPT and Codex are due to share an allowance on Plus, Pro and Business plans, according to the roundup. Cognition’s SWE-2 Flash was reported as free within Devin through November 30; the source does not say what access or pricing will follow that date.
Readers can expect further clarification through company documentation, release notes and research publication. The Anthropic biology work may be easier to evaluate once its submitted preprint becomes public. Cognition’s proposed partner access would also clarify how the company plans to balance testing and availability against its stated security concerns. Until then, the reported scores and product terms should be treated as attributed company claims, not independent confirmation.
Key Questions
What is the main news in this roundup?
ThorstenMeyerAI.com summarized 10 AI-lab announcements published during November 2–8, 2026, spanning agent memory, research, coding models, media tools and product access.
Which companies are described as changing agent memory?
The report says Anthropic, Meta and SpaceXSI published work on agents that remember. The supplied details explain Anthropic’s system; they do not describe the other companies’ approaches in comparable depth.
Are the model performance figures independently verified?
No independent verification is provided in the source material. The roundup identifies the cited scores and benchmarks as company-reported, so their testing methods and results should be read with that qualification.
Are Kimi K3.1 Pro’s weights open source?
Moonshot AI published the weights on Hugging Face, but the roundup says the model uses Moonshot’s own license, not a standard open-source license. Users should check 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 scheduled to begin in December 2026. It does not give further details about how the shared limit will be calculated.
Source: ThorstenMeyerAI.com
