🔍 Read the full analysis: AI Lab Week: The Latest Moves From Anthropic, OpenAI And More on ThorstenMeyerAI.com
Get hardware and tech essentials 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 from ThorstenMeyerAI.com identifies ten announcements from seven AI labs, spanning agent memory, scientific research, coding models, model weights, video and subscription usage. The source says it checked the announcements against lab posts, documentation or code, while noting that many performance and cost figures are company-reported and that three items on its initial list did not pass verification.
Anthropic, OpenAI and other AI labs announced changes to agent memory, model access, scientific research tools, video and subscription usage during the week of Nov. 2–8, 2026, according to a roundup published Sunday by ThorstenMeyerAI.com. The report says it checked each of ten items against a lab post, documentation or code, but many performance and pricing figures remain company-reported measurements, not independent evaluations.
Anthropic’s announcements included a Nov. 6 engineering post describing four layers for Claude’s cross-session memory: transcripts, model-written notes, larger linked structures such as wikis, and a small block of information loaded at the start of a session. The post also described a review process the authors call “dreaming,” which revisits older records to correct notes and organize information. Anthropic reported that memory improved Opus 5.5’s mean reward from 0.655 to 0.838 on one database task. Those results come from the company’s own tests; the post also says Claude Code loads only the first 200 lines or 25 KB of its memory file.
In a separate science report dated Nov. 4, Anthropic described researchers using Claude in work on the three-dimensional structure of a bacterial vault particle, a protein shell with an unknown purpose. The team searched roughly 15 billion predicted proteins, tested 44, and reported that seven of 15 soluble proteins formed vault-like particles. The account also described a protocol mistake that cost two weeks and a map blended with its mirror image, which researchers had to identify. The source says the related preprint had been submitted but was not yet public. Anthropic’s economists also reported an analysis of 207,049 Claude sessions from August and September; those usage figures are the company’s own analysis.
Other moves concerned whether customers can inspect or use model weights. Cognition released SWE-2 Flash on Nov. 4 but kept its weights private, saying the model’s coding and security capabilities could pose cyber risks. The company said its ExploitBench score rose from 11.3% to 34.8% after security training and that it plans to share weights with trusted partners. Moonshot AI, by contrast, published the weights for Kimi K3.1 Pro on Hugging Face. Its model card lists 2.8 trillion total parameters, 104 billion active parameters and a context window of about one million tokens. The release uses Moonshot’s own license, rather than a standard open-source license.
AI Lab WeekThe latest moves
A source-checked roundup covers ten announcements from seven AI labs, with changes to memory, research, model access, media tools and subscription usage.
What AI systems remember
Agent memory can carry project details between sessions, while review steps help catch errors in stored information.
Anthropic described transcripts, notes, linked structures and a small session-start memory block.
A process called “dreaming” revisits logs to correct notes and organize information.
Claude Code loads only the first 200 lines or 25 KB of its memory file.
Model access takes different paths
For developers and organizations, access rules shape control, security and how a model can be used.
- Cognition kept SWE-2 Flash weights private, citing cyber risks.
- The company plans to share them with trusted partners.
- Moonshot published Kimi K3.1 Pro weights on Hugging Face.
- Its model card lists Moonshot’s own license.
Research gains need human checks
Anthropic described AI-assisted work on a bacterial vault particle, with errors and protocol issues still requiring researchers’ attention.
Explore predicted proteins
Researchers searched roughly 15 billion predicted proteins and tested 44.
Seven formed particles
Seven of 15 soluble proteins reportedly formed vault-like particles.
Errors took time to spot
A protocol mistake cost two weeks, and researchers identified a map blended with its mirror image.
Product changes affect planning
Media tools and subscription rules may change how readers and teams budget usage.
| Announcement | What the roundup says | Why it matters |
|---|---|---|
| Google video | 15-second clips; listed price of $0.07 per second | Teams can consider clip length and stated cost. |
| OpenAI image model | Can make animated GIFs | Offers another media format for creators. |
| ChatGPT and Codex | Shared usage allowance on Plus, Pro and Business from December | Subscribers may need to divide usage across products. |
How to read the claims
Source checks trace announcements to lab posts, documentation or code; they do not independently validate company results.
Memory, Access and Usage Rules
Taken together, the announcements point to practical questions beyond model capability: what an AI system remembers, who can run its underlying weights, and how access is metered. Persistent agent memory may help a system carry project details between sessions, but it also creates a need to review stored information for mistakes. Anthropic’s own engineering account describes a process for revisiting records, while its science report offers a concrete example of researchers correcting errors in an AI-supported workflow.
The competing approaches to model access matter to developers and organizations weighing control, security and cost. Cognition says it is restricting access to SWE-2 Flash because it views the model’s security capabilities as sensitive; Moonshot has released Kimi weights under its own license. The label “open-weight” by itself does not settle whether weights are publicly available or what uses a license permits. The roundup also reports that ChatGPT and Codex will share one usage allowance on Plus, Pro and Business plans starting in December, a change that could affect how subscribers divide their work across products.
These developments do not establish that any one product is more capable or cost-effective for all users. Many figures in the roundup—including benchmark scores, prices and claimed savings—come from the companies involved. They are signals about what the labs are building and how they are positioning products, not independent proof of results.
Products Worth Considering
Apple MacBook Pro 14-inch (2021) – M1 Pro Chip, 10-Core CPU, 16-Core GPU, 16GB RAM, 1TB SSD, Space Gray
Apple MacBook Pro 14.2-inch with M5 Max, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Pro 16.2-inch with M5 Pro, 18-core CPU, 20-core GPU, 64GB Memory, 1TB SSD, Space Black
As an affiliate, we earn on qualifying purchases.
Announcements Across One Week
The source describes seven AI labs and ten announcements over the week, framing the items as part of a broader shift toward agents that retain information, AI-assisted scientific work, longer or cheaper media generation, and new access rules. Its account says Anthropic, Meta and SpaceXSI, identified as formerly xAI, published work on agents that remember. It says two of those efforts include a review step for deciding what information is retained.
The roundup also covers media and product changes. It reports that Google’s video model produces 15-second clips at a listed price of $0.07 per second, while OpenAI’s image model can make animated GIFs. It says OpenAI will combine ChatGPT and Codex usage allowances for Plus, Pro and Business subscribers from December. The supplied source does not provide the full technical details or independent testing for these products, so those descriptions should be read as a summary of the lab announcements, not a comparative assessment.
ThorstenMeyerAI.com says it removed three items from its initial list after they did not survive source checks. That editorial note is relevant to the roundup’s scope: the article presents items it says it could verify against primary material, while acknowledging that company-reported metrics have not thereby been independently validated.
““Seven AI labs, seven days, ten announcements.””
— ThorstenMeyerAI.com
Metrics and Access Still Unverified
The roundup says it checked announcements against primary sources, but that does not independently confirm the companies’ performance claims. Benchmark scores, memory results, speed and price comparisons cited in the report are attributed to the labs or their own documentation. The source does not provide independent replications or enough testing detail to compare the products on a common basis.
Several specific points remain unresolved in the supplied material. Anthropic’s vault-particle preprint had been submitted but was not yet public, and the biological structure’s purpose remains unknown. Cognition’s planned trusted-partner program is still in development, with no details here on eligibility or timing. Moonshot’s model card lists its own license, but the excerpt does not explain the license’s practical restrictions. The source also says three proposed roundup items were dropped after verification checks, without identifying them in the supplied text.
The material gives only limited information about OpenAI’s animated-GIF feature and Google’s video model, and it does not state when all users will receive access. The exact terms and rollout details for the December shared ChatGPT and Codex allowance are also not included.
Rollouts and Research to Watch
The clearest dated next step in the roundup is the planned December change to ChatGPT and Codex allowances for Plus, Pro and Business plans. Users will need the published plan terms to understand how the shared limit works, including how usage is counted and whether the allowance differs by plan.
Cognition says it is developing a process for sharing SWE-2 Flash with trusted partners, but the source gives no launch date. Anthropic’s submitted vault-particle preprint may provide further detail when it becomes publicly available. Additional documentation or independent tests would help clarify the reported memory, benchmark, price and speed figures across the announcements.
For now, the roundup’s main value is a verified-at-the-source snapshot of what the labs announced during one week, paired with a warning that company metrics are not independent evidence. Further product documentation, access terms and outside evaluations will determine how much of the announced capability is available in practice.
Key Questions
What is the main development in the AI lab roundup?
ThorstenMeyerAI.com reported ten announcements from seven AI labs during Nov. 2–8, 2026. The items cover agent memory, scientific research, coding models, model weights, video and subscription usage.
Did Anthropic independently prove its memory results?
No independent validation is described in the supplied source. Anthropic reported that memory raised Opus 5.5’s mean reward from 0.655 to 0.838 on one database task; those figures are from the company’s own tests.
Are Kimi K3.1 Pro’s weights openly licensed?
Moonshot AI published the model weights on Hugging Face, but the model card uses Moonshot’s own license, not a standard open-source license. The supplied material does not detail the license’s restrictions.
Why is Cognition keeping SWE-2 Flash’s weights private?
Cognition cited cyber risk, saying the model’s capabilities could be misused. It said it is building a program to share the weights with trusted partners, but no timing or access criteria are given in the source.
When will ChatGPT and Codex share a usage allowance?
The roundup says the shared allowance is planned to begin in December 2026 for Plus, Pro and Business plans. The supplied material does not give the detailed limits or counting rules.
Source: ThorstenMeyerAI.com
