Minerka The AI briefing

Today

Timeline

The current briefing in the order the primary pages were dated, newest first.

Guide — standing explanations

Start here

One worked check, four list prices, then three questions — before you trust what changed this week.

Worked check: Carbon’s list is $2/$10 — not on the public API yet; the 59 is DeepMind’s Index v5 report. Field card · prices and 3 checks

Pocket list (November 1, 2026)

  1. Measurement — which named test, on which date, scored by whom.
  2. Bill — list input and output per million tokens on the vendor page.
  3. Door — live API, lab claim, replay, or a downloadable file — keep them separate.

Models · Agents · Fine-tuning · All chapters

  1. November 1, 2026 — A claimed ~14T Anthropic “numbat” checkpoint on Hugging Face has 43 tensors that match a name-seeded RNGModels · 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.
  2. October 31, 2026 — DeepSeek opens an MIT research prototype of a one-repo GitHub coding agent for DeepSeek HarnessAgents & Harnesses · The useful change is the harness pattern DeepSeek is willing to show: act only when asked or when a team rule allows, and speak only when a person must act or can use a result.
  3. October 31, 2026 — Karpathy ships HackerTube in about two hours with Claude Opus 5.5: 6,454 HN-recommended videosAgents & 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.
  4. October 30, 2026 — NVIDIA posts NemoClaw blueprints, and says LangChain’s harness tune for Nemotron 3 Ultra cut cost about 10xAgents & Harnesses · The useful change is the harness, not a new weight file: NVIDIA’s post says co-tuning Deep Agents with Nemotron 3 Ultra cut cost about 10x while staying near frontier.
  5. October 29, 2026 — Mistral releases the weights for Large 4 (Le Chonk), a 1T-parameter open modelModels · The preview was an API door. The weights are a file you can host, with Mistral's own coding and cyber numbers attached.
  6. October 29, 2026 — Epoch says open-weight GLM 5.3 turned public Firefox fixes into working exploits for about $1,000 a versionSecurity & 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.
  7. October 29, 2026 — Cursor pulls malicious Claude Code extension versions 2.1.312 and 2.1.313 from its marketplaceSecurity & 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.
  8. October 29, 2026 — OpenAI says expanded misalignment monitors would have alerted 30 hours before the Hugging Face breachSecurity & Safety · The useful number is the lead time OpenAI claims for its own past failure. The monitors are internal. Outside developers do not get that dashboard.
  9. October 28, 2026 — Gemini 4 Carbon is priced like Argon, and DeepMind says it leads the new index at 59Models · The number people will quote is 59, and the fact that changes the work is that almost no one can call the model yet. The price matches Argon's introduction, and the door does not.
  10. October 28, 2026 — NVIDIA posts an open transcription model, 4 billion parameters active, at 5.33 percent word errorModels · 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.
  11. October 27, 2026 — Artificial Analysis puts 60 percent of its index on private tests, and Opus 5.5 leads at 58Research · 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.
  12. October 27, 2026 — Anthropic's partners report a 70 percent shorter design, and the drugs are not in people yetResearch · 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.
  13. October 27, 2026 — Ai2 finds a novelty judge can fall from 93.5 percent to 40.9 percent when the prompt changesResearch · 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.
  14. October 27, 2026 — CUDA 13.4 can refuse a GPU allocation that crosses a memory limitInfrastructure & 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.
  15. October 27, 2026 — Liquid AI ships LFM3, a 1.7 billion parameter model that beats its larger predecessorModels · A model this small can be fine-tuned on one GPU and run on a device. The 38.42 is Liquid's WebArena score, and the license name is not on the post.
  16. October 27, 2026 — DeepMind shows Genie 4 to trusted testers, with Gemini 3.8 Flash steering the worldAgents & 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.
  17. October 27, 2026 — NVIDIA says the 4 GB Nemotron Bolt model is available, and uses fewer tokensModels · The file size is the useful fact. The quality and token claims do not name the other models, so they are not a ranking.
  18. October 26, 2026 — GPT-6.1 Terra is $1 and $5 per million tokens, and OpenAI's coding score more than doublesModels · 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.
  19. October 26, 2026 — Anthropic filed a public S-1 and named the ticker ANTIndustry & Companies · The public step is the filing and the ticker. There is no share count and no price, so this is not a valuation.
  20. October 24, 2026 — Kimi K3.1 Pro posts 65.7 percent on Terminal-Bench, with weights due by November 4Models · The score is high, and the file is not downloadable yet. Until November 4 the model is an API and an app, not an open checkpoint you can rerun.
  21. October 21, 2026 — Mirendil says one researcher pretrained a coding model for under $10 millionModels · 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.
  22. October 21, 2026 — PHP patches two critical bugs, and the working exploits came from AI-assisted researchSecurity & 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.
  23. October 21, 2026 — Anthropic opens a research preview of Claude Duet, a proactive agentAgents & Harnesses · It shares the plan's usage limits, and background checks count against them. Sending, paying, signing, and deleting still wait for approval.
  24. October 20, 2026 — Mistral prices Voxtral 3, including a 4B transcription model at $0.0035 a minuteModels · 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.
  25. October 20, 2026 — DeepSeek open-sources DeepInfer, and says one cluster serves 155 trillion tokens a dayInfrastructure & 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.
  26. October 17, 2026 — MiniMax posts M3.1-Flash, about 428B parameters with 23B activeAgents & 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.
  27. October 17, 2026 — DeepSeek posts a 763B MIT checkpoint, V4.1-Flash-1017Models · 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.
  28. October 16, 2026 — Reflection releases Beam, a 501B open model under Apache 2.0Agents & 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.
  29. October 15, 2026 — Mistral releases Timestral, a 209M model for forecastsModels · 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.
  30. October 15, 2026 — Meta releases Muse Spark Open, an 816B model that runs on two Mac StudiosAgents & 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.
  31. October 15, 2026 — NVIDIA releases Nemotron 3.5 Super VL, a 120B open modelModels · 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.
  32. October 13, 2026 — Gemini 4 Argon opens to developers at $2 per million input tokens until November 30Models · 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.
  33. October 13, 2026 — Liquid AI releases a 24B agent model with 2B activeAgents & 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.
  34. October 13, 2026 — Google reports 87 percent for Argon on a robotics reasoning testAgents & 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.
  35. October 12, 2026 — Meta releases Muse Wonder for agents that remember and keep goingAgents & Harnesses · A new frontier model is priced and available for agent work today. The Muse app does not have it yet.
  36. October 9, 2026 — Microsoft-Decision-1 scores fixed choices, and Microsoft says it is 35 times quicker than GPT-6 SolAgents & 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.
  37. October 8, 2026 — Epoch grades two leading models at 65 percent of its own work standardResearch · 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.
  38. October 8, 2026 — Coinbase post-trained a 9B model that beat Opus 4.5 on its fraud testEnterprise 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.
  39. October 8, 2026 — Odyssey-3 is a world model you can act insideAgents & 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.
  40. October 8, 2026 — Carbon-A proposes 566 million gene candidates across 22,617 speciesResearch · The number is a pile of candidates, not confirmed genes.
  41. October 7, 2026 — GPT-6 arrives in ChatGPT, and it can draw the screenModels · A chat box that was a transcript can now be an interface the model writes, on free accounts as well as paid.
  42. October 7, 2026 — Microsoft Research ships Agent Lightning in about 3,500 linesAgents & 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.
  43. October 6, 2026 — EmbeddingGemma 2 searches text, images, audio, and video on the deviceModels · Private photos, voice memos, and code can be searched without uploading them.