Measurement, bill, and door for each claim below.
Security & Safety
A cyber capability claim needs a count, a cost, and who ran it. OpenAI’s 948 LPEs are OpenAI’s validated exploits in its controlled lab — not a public leaderboard row, and the model is not a shipping API product yet.
Models
A bounded extraction job is not a general chat score. Dharma’s number is field-level F1 on its own suite — ask who measured it, and keep the self-host cost separate from the API list price.
Research
A cyber score is not one number until you name the bench, pass@k, and who ran it. CWE-bench v2 is public and programmatic — ask pass@1 vs pass@4, and whether the run used the open set or the held-out set.
Research
“Long context” is not one test. Ask which StateBench level and length — finding the latest assignment is not the same as chasing reference swaps across a long chain.
Ask measurement, bill, and door of each claim above. The pocket list is a separate bill: input and output when a vendor publishes them.
Pocket list (November 4, 2026)
Models · Agents · Fine-tuning · All chapters
The week · Timeline · Radar · Library · Guide
Research
This is not a fine-tune recipe. The claim is that Astra reasons about raw weights and edits them directly — treat MATH-500 and IFEval as the authors’ measured before/after, not a public leaderboard rerun.
Research
LifeVal is OpenAI’s own everyday-task suite — not a public leaderboard you can rerun from the post alone. The useful habit is still: which tasks, scored by whom, on which date.
Research
A frontier agent score of 12% is SemiAnalysis’s figure on GDPx — not a claim that agents replace domain work. Safety classifiers also pull some models down on cyber, biology, and alignment.
Models
The preview was an API door. The weights are a file you can host, with Mistral's own coding and cyber numbers attached.
Security & Safety
A published fix can be a map. Epoch's number is that an open-weight model, not a closed frontier lab tool, finished that map cheaply enough to matter for browsers that still ship weekly.
Models
The number people will quote is 59, and almost no one can call the model yet. The introductory price matches Argon. This week’s access is a defender cohort, not the public API.
Models
A model this small can be fine-tuned on one GPU and run on a device. The 38.42 is Liquid's WebArena score. The announcement does not name a license; the card names LFM 1.0.
Models
The cheap computer-use model is the one a high-volume workflow can actually call. The 50.3 percent is OpenAI's number, and OpenAI still points harder coding at Sol and Astra.
Models
The 65.7 percent is Moonshot's row, run in Claude Code. The downloadable card now names 2.8 trillion parameters and 104 billion active. The license name is Kimi K3.1 Pro License, which is a separate fact from the score.
Models
The post does not offer weights or an API. On Mirendil's own bits-per-byte table the average is 0.222, behind Kimi-K3 at 0.193 and ahead of the other three base and post-trained models it lists.
Security & Safety
Update the PHP that serves the site, then restart the workers. A configuration change is not the fix. The same class of bug is now cheap enough for a model to turn into a working exploit.
Agents & Harnesses
It shares the plan's usage limits, and background checks count against them. Sending, paying, signing, and deleting still wait for approval.
Models
The bill is on the page. Mini transcription is $0.0035 a minute, realtime is $0.008, and a voice agent that reasons with Mistral Small 4 is $0.04 a minute.
Agents & Harnesses
On MiniMax's Terminal-Bench 4.0 row it scores 38.1. The same table lists DeepSeek V4.1-Flash-1017 at 42.3 and Opus 5 at 51.8.
Models
If you already call deepseek-flash, this checkpoint is what that name serves. DeepSeek's own table moves Terminal-Bench 4.0 from 31.2 to 42.3, and still has Opus 5 ahead on that row.
Models
A forecast can be a small model on a CPU, not a chat model. Mistral says the 209 million parameter size ranks first on three public forecasting benchmarks. That rank is Mistral's, and the technical report is still to come.
Agents & Harnesses
A day after Meta's open agent model, a new lab shipped one a company can use commercially without a custom license. The coding scores are Reflection's.
Agents & Harnesses
The open-weight model a team can inspect is now in the same class Meta is using for agents. Running it yourself still means two very large machines, or a host.
Models
A frontier model for long jobs is on the public API, cheaper than the flagship rates until the end of November. The Gemini app only has it for paying subscribers.
Agents & Harnesses
A lot of agent work is a choice among options the software already knows. Microsoft is selling a model for that job, and the speed claim is the reason to try it on a real queue.
Agents & Harnesses
A new frontier model is priced and available for agent work today. The Muse app does not have it yet.
Enterprise AI
A company can train a small open model on data only it has, and beat a frontier model on one job. Other companies can copy that path.
Models
A chat box that was a transcript can now be an interface the model writes, on free accounts as well as paid.
Models
Private photos, voice memos, and code can be searched without uploading them.
Infrastructure & Chips
The useful number is the flexible megawatts, not a model score: OpenAI is publishing a demand-response commitment tied to a named campus and utility, while the facility itself is not operating yet.
Enterprise AI
The useful number is the pilot time cut, not a model score: DeepMind says routine householder apps are up to 70% of officers’ workloads, and PlanAssist only drafts support — a human still decides.
Models
A weight file is not a lab claim until someone can show it was trained. An exact RNG fingerprint is a concrete reason to keep the download separate from the story.
Agents & Harnesses
A named builder timed an end-to-end coding-agent loop on Opus 5.5 with concrete output size — useful when asking which piece moved: model or harness.
Security & Safety
A coding-agent extension is part of the harness. A poisoned marketplace build shows the supply-chain check belongs next to the model claim.
Models
The file you can download does transcription, speaker labels, and translation in one checkpoint. The 5.33 percent is the open-weight English row, with paid speech APIs left out of that comparison.
Research
A score on the old index is not a score on this one. The useful change is that most of the weight now sits on tests the public cannot read.
Research
These are partner measurements from the lab bench, not a medicine a patient can receive. A shorter design and a faster paperwork pass are not a result in people.
Research
A novelty score is not a property of the idea until you know the prompt. Ai2's own numbers show the same pairs flipping when a few words change.
Infrastructure & Chips
Sharing a GPU used to mean hoping the jobs stayed small. The new boundary is a hard cap, and the choice of compute mode decides whether one crash takes the neighbors with it.
Agents & Harnesses
The demo is real and the product is not. Testers have it. Everyone else is on a waitlist, and there is no published speed or resolution to compare with the last Genie.
Models
The file size is the useful fact. The quality and token claims do not name the other models, so they are not a ranking.
Industry & Companies
The public step is the filing and the ticker. There is no share count and no price, so this is not a valuation.
Research
The score is one pass on a small set of Epoch's own jobs. Epoch says the numbers are noisy, and that doing well on the easy rubric items is not the same as being 65 percent of the way to replacing the person.
Infrastructure & Chips
The file you can clone is not the service DeepSeek describes. The note says the KV cache and the load balancer stay tied to DeepSeek's own infrastructure.
Models
Another open agent model, this one tied to NVIDIA hardware. The speed claim is NVIDIA's. A smaller Bolt model at 4 GB is dated October 23 and is not out yet.
Agents & Harnesses
An on-device reasoning model with only 2 billion parameters active. Liquid's speed and tool-use scores are Liquid's, and the license is LFM 1.0 rather than Apache.
Agents & Harnesses
The training loop can sit in front of a harness you already run, instead of a copy of that harness rebuilt for the trainer. The coding gain is Microsoft's.
Agents & Harnesses
The same model that opened on the API has a robotics writeup. The 87 percent is Google's, and the robot clips are demonstrations, not a robot you can deploy.
Agents & Harnesses
A learned world that can train a robot policy would matter. That claim is the lab’s, and this is not a robot you can deploy.
Research
The number is a pile of candidates, not confirmed genes.