🔍 Read the full analysis: AI Lab Week: 10 Moves And Updates From Seven Companies on ThorstenMeyerAI.com
Get the latest gadgets 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 November 2–8 roundup from ThorstenMeyerAI.com counts 10 updates from seven AI companies, spanning agent memory, scientific research, model access and media generation. The source says it checked items against company posts, documentation or code; many performance and cost figures remain companies’ own measurements.
Seven AI companies made 10 reported moves during the week of November 2–8, according to a roundup published by ThorstenMeyerAI.com on November 8. The updates covered agent memory, AI-assisted biology, coding-model access, model weights, video and image generation, and shared usage limits; many reported performance figures come from the companies themselves.
The roundup says it checked each item against a lab’s own post, documentation or code. Its central pattern was persistent agent memory: Anthropic, Meta and SpaceXSI, identified in the source as formerly xAI, published work on systems that retain information across sessions. Anthropic described a process that reviews past logs and revises memory notes, while the source says two of the three companies include a review step for deciding what to retain.
Anthropic also reported using Claude in research on the structure of a bacterial vault particle. Its team searched a database of about 15 billion predicted proteins and tested 44 candidates; the company said seven of 15 proteins that proved soluble formed vault-like particles. The source notes that the related preprint had been submitted but was not yet public, and that researchers reported catching errors during the work.
Other updates focused on who can use models and how. Cognition released SWE-2 Flash but kept its weights private, citing security concerns; it said access for trusted partners was being developed. Moonshot AI published Kimi K3.1 Pro’s weights on Hugging Face under its own license. The roundup also reported Google video generation at $0.07 per second for 15-second clips, OpenAI image generation of animated GIFs, and a planned shared usage allowance for ChatGPT and Codex on certain plans from December.
10 updates, seven companies What changed this week
A November 2–8 roundup spans agent memory, research, model access and media generation. Its claims offer signals for users and businesses, while many measurements remain company-reported.
Agents that remember
Persistent memory could help work carry across sessions, but retaining errors is a risk.
Review and revise
Claude’s memory process reviews past logs and revises notes to decide what to retain.
Memory research
Meta was among three companies whose work focused on information retained across sessions.
Persistent systems
The source identifies SpaceXSI as formerly xAI and includes it in the memory theme.
Access depends on the terms
Weight availability and licensing shape whether teams can host and reuse a model.
- SWE-2 Flash weights remain private, citing security concerns.
- A trusted-partner program is in development; criteria and timing were not given.
- Kimi K3.1 Pro weights were published on Hugging Face.
- The release uses Moonshot’s own license, so downloadability does not settle reuse rights.
Research and media updates
The roundup also covered scientific research and new generation features.
Anthropic reported testing 44 candidates from a database of about 15 billion predicted proteins; its submitted preprint was not yet public.
Google video generation was reported at $0.07 per second for 15-second clips; OpenAI added animated GIF generation.
A planned December change combines ChatGPT and Codex allowances on Plus, Pro and Business plans; the allowance size was not provided.
Read claims with care
The roundup checked items against company posts, documentation or code, but the supplied account lacks independent replication.
Memory, Access and AI Costs
Taken together, the updates point to practical questions facing AI users: whether an agent can retain useful information without carrying forward errors, whether a model’s weights can be downloaded and reused, and how much generation and inference will cost. These are not merely product-label differences. Memory systems may affect how work carries across sessions, while licensing and access limits shape whether organizations can host a model themselves or must use a provider’s service.
The contrast between Cognition and Moonshot is especially relevant to teams evaluating model access. Cognition said it was withholding SWE-2 Flash’s weights because of potential cyber risks, whereas Moonshot released Kimi K3.1 Pro’s weights subject to its own license. Neither “built on open weights” nor “weights released” alone settles what a user may do: the license and access conditions matter.
Cost and performance figures should be read with similar care. The roundup explicitly identifies many scores and measurements as company-reported, including model evaluations and claimed efficiency. They can indicate what a company is targeting, but they are not independent comparisons unless tested under comparable conditions by outside researchers.
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
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Eucalyptus Green
As an affiliate, we earn on qualifying purchases.
The Week’s Product and Research Updates
The source frames the week as a set of 10 developments rather than a single coordinated announcement. It says three items on an earlier list did not survive its source checks, although the supplied material does not give their names or explain each exclusion. Anthropic and OpenAI were among the companies covered, and the publisher said neither reviewed nor sponsored the roundup.
Some details have different evidentiary status. Anthropic’s biology work was described in a company post, but the source says its preprint was submitted and not yet public. The company’s report on scientific use of Claude drew on 207,049 sessions from August and September; those figures describe activity on its own product, not scientific work across the field. Likewise, Cognition’s benchmark changes and Moonshot’s model-card comparisons are company-provided results.
The roundup also links model availability to operating costs. It cites a conference statement from Revolut that about 2,000 engineers moved to a self-hosted model from Z.ai, running on 56 B200 GPUs, at a claimed cost of $0.30 per million tokens. Revolut said its invoice was one-tenth of what it had paid Anthropic and OpenAI. That is a company’s account of its own deployment, not a verified estimate of what other organizations would pay.
“Memory was the theme.”
— ThorstenMeyerAI.com roundup
Independent Evidence Is Limited
The supplied material does not provide independent replication of the companies’ benchmark scores, cost claims or product-performance measurements. It also does not specify the exact methodology behind each reported result, so readers cannot treat the figures as directly comparable across labs.
Anthropic’s vault research remains an early research report in this account: the preprint had been submitted but was not publicly available. The source also does not name the three items removed from its initial list, limiting what readers can assess about the scope of its corrections. Details of the trusted-partner program for Cognition’s weights, including eligibility and timing, are not given.
For the December ChatGPT and Codex allowance change, the roundup identifies the affected Plus, Pro and Business plans but does not give the allowance size or explain how usage will be counted. Licensing terms may also impose conditions beyond the basic fact that Moonshot’s weights are downloadable.
Access Rules and Research Details
The most immediate stated date is December 2026, when ChatGPT and Codex are expected to share one allowance on Plus, Pro and Business plans. The source does not provide a specific start date or the new limits, so plan details will need to be checked against OpenAI’s published terms.
Cognition said it was developing a trusted-partner program for SWE-2 Flash weights, but gave no timeline or access criteria in the supplied account. Anthropic’s submitted vault preprint may provide more detail if it becomes public. Further independent testing would also help clarify how the companies’ benchmark results and cost claims compare outside their own evaluations.
For now, the roundup’s clearest takeaway is a set of developments, not a settled ranking: labs are experimenting with persistent agents, applying models in research and changing how users access or pay for tools. The next evidence to watch is documentation that spells out limits and licensing, alongside independent evaluations of the claims.
Key Questions
What is the news in the AI lab roundup?
ThorstenMeyerAI.com reported 10 updates from seven AI companies during November 2–8, 2026, including developments in agent memory, biology research, model access and generative media.
Which companies were associated with agent-memory work?
The source names Anthropic, Meta and SpaceXSI, which it identifies as formerly xAI. It says Anthropic described memory notes, linked information structures and a process for reviewing past logs.
Did Moonshot AI release Kimi K3.1 Pro’s weights?
According to the roundup, Moonshot published the weights on Hugging Face on November 4. The model card lists 2.8 trillion total parameters, and the source says the release uses Moonshot’s own license rather than a standard open-source license.
Are the reported model scores independently verified?
Not in the supplied material. The roundup says many figures, including benchmark scores and performance measurements, are company-reported. It presents them as signals rather than independent proof.
When will ChatGPT and Codex share an allowance?
The source says the shared allowance is planned to begin in December 2026 for Plus, Pro and Business plans. It does not state the exact date or the allowance size.
Source: ThorstenMeyerAI.com
