🔍 Read the full analysis: 10 Moves That Defined This Week In AI Labs on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Seven AI labs made ten reported announcements during the week of Nov. 2–8, 2026, spanning agent memory, scientific research, coding models, model weights, video and image tools, and product usage limits. Many performance and cost figures come from the companies themselves; one research preprint was not yet public, and the roundup says three candidate items were excluded after source checks.
Seven AI labs made ten reported announcements during the week of Nov. 2–8, according to a roundup published Sunday by ThorstenMeyerAI.com. The developments included new approaches to agent memory, AI-assisted biological research, restrictions and releases of model weights, video and image tools, and a shared usage allowance for some ChatGPT and Codex subscribers.
The roundup says its items were checked against each lab’s own post, product documentation or code. It highlights Anthropic’s Nov. 6 engineering post on Claude’s memory system, which describes saved session transcripts, model-written notes, linked knowledge structures and a small block of information loaded at the start of a session. Anthropic also described an automated review process it calls “dreaming,” which revisits prior logs and revises stored notes. The company reported that memory improved Opus 5.5’s mean reward on one database task from 0.655 to 0.838; those are Anthropic’s own test results, not independent measurements.
Other announcements concerned access and research. Cognition released SWE-2 Flash on Nov. 4 but kept its weights private for now, citing the possibility that its coding and security capabilities could aid attackers. Moonshot AI, by contrast, published the weights for Kimi K3.1 Pro on Hugging Face that 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 says the model uses Moonshot’s own license rather than a standard open-source license.
Anthropic also reported using Claude in a project to investigate the structure of a bacterial vault particle, and published an analysis of 207,049 scientific sessions from August and September. Google’s video model was reported to generate 15-second clips at $0.07 per second, while OpenAI’s image model can make animated GIFs. The roundup says ChatGPT and Codex will share one usage allowance on Plus, Pro and Business plans from December. The supplied source material gives no further detail on the remaining announcements beyond listing the week’s overall themes.
10 Moves That Defined This Week in AI Labs
Seven AI labs made reported announcements across memory, research, model access, media tools and product limits during Nov. 2–8, 2026.
Memory moves into focus
Labs described ways to carry useful information across sessions and review what an agent remembers.
Claude memory
Saved transcripts, written notes and linked knowledge support recall; an automated “dreaming” process revises stored notes.
More memory systems
Meta and SpaceXSI also published agent-memory work. Two of the three systems include a review step.
Model access takes different paths
Availability and license terms shape what developers can inspect, run or adapt.
- Released SWE-2 Flash while keeping its weights private.
- Cited security concerns and plans for a trusted-partner program.
- Published Kimi K3.1 Pro weights on Hugging Face.
- Uses its own license, rather than a standard open-source license.
Research and media tools expand
The roundup also describes scientific work and new ways to create media.
Anthropic described using Claude to investigate a bacterial vault particle; the submitted preprint was not yet public.
Google’s video model was reported to generate short clips, and OpenAI’s image model can make animated GIFs.
Usage and evidence need attention
Product terms and company-reported results affect how users should interpret the announcements.
| What’s reported | What readers should know |
|---|---|
| ChatGPT and Codex share an allowance for Plus, Pro and Business | Scheduled for December; allowance size and transition rules are unspecified. |
| Performance, cost and benchmark figures | Many are company-reported, without independent replication in the supplied material. |
AI Labs Rework Memory and Access
Taken together, the announcements point to practical choices facing AI users and developers: what an agent remembers, who can inspect or run a model’s weights, and how product usage is metered. Persistent memory may help agents carry information between sessions, but it also creates a need to check whether stored notes are accurate and relevant. Anthropic’s description includes a review step, while the report says Meta and SpaceXSI also published work on agent memory, with two of the three systems incorporating a review step.
The contrast between Cognition and Moonshot is a reminder that model access is not a simple open-or-closed distinction. Cognition says it is limiting distribution over security concerns and plans a trusted-partner program; Moonshot has released weights under a proprietary license. For organizations deciding whether to deploy or adapt a model, the license and actual access terms matter alongside parameter counts and benchmark results.
Cost and capability figures may influence purchasing decisions, but the roundup cautions that many numbers—including claimed speed, price and performance—are lab-reported measurements. They can indicate what companies are prioritizing, but the supplied material does not establish that results will carry over to independent tests or every customer’s workload.
Popular right now
KitchenAid Artisan Series 5-Quart Tilt-Head Stand Mixer with Pouring Shield, Cherry Tomato
Ninja Luxe Café Mini Espresso and Drip Coffee Machine, Stainless Steel, ES301
Dyson Supersonic Nural Hair Dryer
As an affiliate, we earn on qualifying purchases.
A Week of Lab Announcements
The roundup covers activity over Nov. 2–8, 2026, rather than reporting a single coordinated launch. Its author says the items were checked against primary sources and that three candidates from an initial list were dropped because they did not pass that check. The material provided here does not identify those excluded items or give a complete, separately itemized account of all ten moves.
For Anthropic’s biological research account, the source says the team searched roughly 15 billion predicted proteins for relatives of known vault proteins, then tested 44 candidates. Seven of 15 proteins that emerged as soluble formed vault-like particles, according to the company’s account. Anthropic also described mistakes, including a protocol error that cost two weeks and a map blended with its mirror image. The related preprint was submitted but not yet public as of the roundup, and an outside vault expert advised the team as a paid consultant.
The source also places Moonshot’s weight release in the context of the model’s October launch, when the weights were not yet available. It mentions a separate claim by Revolut in an October conference talk about moving engineers to a self-hosted model from Z.ai and reducing its invoice. That is a company statement reported by the roundup, not an independently verified comparison of operating costs.
“Memory was the theme.”
— ThorstenMeyerAI.com roundup
Benchmarks and Missing Details
Most of the performance, cost and usage figures in the source are reported by the labs or companies themselves. The material does not provide independent replication for Anthropic’s memory score, Cognition’s benchmark results, or the stated video pricing. Comparisons across products may also depend on test setup, workload and access terms not given in the roundup.
Anthropic’s vault research preprint had been submitted but was not public as of Nov. 8, so its methods and findings could not yet be assessed from the paper. The source does not specify when Cognition’s trusted-partner weights program will begin, nor does it provide the full terms of Moonshot’s license. It also does not enumerate all ten announcements in the material supplied here, or name the three items that were removed after checking. The ChatGPT-Codex shared allowance is scheduled for December, but the precise allowance and transition rules are not stated.
December Usage Rules Ahead
The next stated milestone is December, when ChatGPT and Codex are due to share an allowance for Plus, Pro and Business subscribers. Users will need the service’s published terms to understand the size of that allowance and how activity across the two products will count toward it.
Further evidence may come from independent testing of the labs’ performance and price claims, publication of the submitted vault-research preprint, and details of Cognition’s proposed access program. Until then, the week’s announcements establish what the companies say they have released or plan to offer; they do not settle how well each system performs across independent evaluations or real-world deployments.
Key Questions
What is the main news in the roundup?
Seven AI labs made ten reported announcements during Nov. 2–8, 2026, covering agent memory, research, model access, media tools and product usage limits.
Which companies released or restricted model weights?
Moonshot AI published Kimi K3.1 Pro’s weights on Hugging Face, under its own license. Cognition released SWE-2 Flash but kept its weights private for now, citing security concerns.
Are the reported benchmark figures independently verified?
The source identifies many scores and measurements as company-reported results. It does not provide independent replication for those figures.
What did Anthropic report about biological research?
Anthropic described using Claude in research on a bacterial vault particle. The related preprint had been submitted but was not yet public as of Nov. 8, 2026.
When will ChatGPT and Codex share an allowance?
The roundup says the shared allowance is planned for December for Plus, Pro and Business plans. It does not state the allowance size or detailed counting rules.
Source: ThorstenMeyerAI.com
