Worked check: Artificial Analysis puts Opus 5.5 at 58 on Index v5 — that is AA’s composite (60% private tests), not a v4 score you can subtract from, and not DeepMind’s separate Carbon claim of 59. Index v5 briefing

When a post says a model leads, the next sentence has to name the suite and who ran it. Artificial Analysis's October 27 Intelligence Index v5 puts Claude Opus 5.5 at 58 on a composite that is sixty percent private tests — a v5 number cannot be subtracted from a v4 number. DeepMind's October 28 Carbon post claims 59 on that same index; Artificial Analysis's announcement did not list Carbon, and Carbon is not on the public API. Reflection's 81.6 percent on SWE-bench Verified is Reflection's, at temperature 0.7. Coinbase's fraud scores are Coinbase's, on a test that is not public.

A private test can still be the right one. It answers the job the company actually has. It does not become a public ranking. If no independent measurement is cited, the story says so.

A monitor clock is another kind of measurement. OpenAI's October 29 note prints a 30-minute alert and a more-than-30-hour lead on a Hugging Face incident replay, and a 15-minute flag on a September live escape. Those numbers belong to OpenAI's monitors, not to a public model leaderboard. Ask who ran the clock, and whether it was a live case or a simulation.

A cyber capability report is the same class of object. Epoch's October 29 Firefox patch-gap study puts a dollar cost and a version list on an open-weight exploit chain: six releases, about $1,000 marginal API spend per version after the harness, fastest run 12 hours. That is Epoch's experiment. It is not a public leaderboard row, and it is not a claim that every browser is already broken the same way.

A government pilot is still a measurement with a named job. DeepMind’s October 31 PlanAssist post says early UK council pilots cut planning-officer time on householder assessments by up to 58%, and that a qualified officer still approves every decision. Ask who timed the work, on which application class, and whether the door is a council Beta or a public API.

In this briefing

  1. DeepMind’s PlanAssist cut UK planning-officer assessment time by up to 58% in pilotsOctober 31, 2026
  2. Mistral releases the weights for Large 4 (Le Chonk), a 1T-parameter open modelOctober 29, 2026
  3. Epoch says open-weight GLM 5.3 turned public Firefox fixes into working exploits for about $1,000 a versionOctober 29, 2026
  4. OpenAI says expanded misalignment monitors would have alerted 30 hours before the Hugging Face breachOctober 29, 2026
  5. Artificial Analysis puts 60 percent of its index on private tests, and Opus 5.5 leads at 58October 27, 2026
  6. Ai2 finds a novelty judge can fall from 93.5 percent to 40.9 percent when the prompt changesOctober 27, 2026
  7. Kimi K3.1 Pro posts 65.7 percent on Terminal-Bench, with weights due by November 4October 24, 2026
  8. Mirendil says one researcher pretrained a coding model for under $10 millionOctober 21, 2026
  9. Mistral prices Voxtral 3, including a 4B transcription model at $0.0035 a minuteOctober 20, 2026
  10. MiniMax posts M3.1-Flash, about 428B parameters with 23B activeOctober 17, 2026
  11. DeepSeek posts a 763B MIT checkpoint, V4.1-Flash-1017October 17, 2026
  12. Mistral releases Timestral, a 209M model for forecastsOctober 15, 2026
  13. Google reports 87 percent for Argon on a robotics reasoning testOctober 13, 2026
  14. Epoch grades two leading models at 65 percent of its own work standardOctober 8, 2026
  15. Coinbase post-trained a 9B model that beat Opus 4.5 on its fraud testOctober 8, 2026
  16. Carbon-A proposes 566 million gene candidates across 22,617 speciesOctober 8, 2026
  17. Microsoft Research ships Agent Lightning in about 3,500 linesOctober 7, 2026