Minerka The AI briefing

Radar

Confirmed changes stay on Today. This page is where a development sits before it earns that, and where a topic goes when the evidence thins.

Today’s briefing · Guide

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

Now

Confirmed changes with a primary page, a date, and a number.

  1. Mistral releases the weights for Large 4 (Le Chonk), a 1T-parameter open modelMajor · Models · Mistral Large 4 moved from a public preview API to downloadable open weights, with updated coding and cyber scores on Mistral's October 29 post.
  2. Epoch says open-weight GLM 5.3 turned public Firefox fixes into working exploits for about $1,000 a versionMajor · Security & Safety · An open-weight model is now documented, with a dollar cost, as finishing visit-only Firefox exploit chains from public but unshipped fixes across six recent releases.
  3. OpenAI says expanded misalignment monitors would have alerted 30 hours before the Hugging Face breachImportant · Security & Safety · Misalignment monitors that missed the start of the Hugging Face episode are now described as covering a wider set of OpenAI surfaces, with published alert times on a replay and on a September live case.
  4. Gemini 4 Carbon is priced like Argon, and DeepMind says it leads the new index at 59Major · Models · Gemini 4 Argon is the model developers can call. Carbon is the model DeepMind now calls its strongest, at the same introductory price, with access limited to a defender cohort.
  5. NVIDIA posts an open transcription model, 4 billion parameters active, at 5.33 percent word errorImportant · Models · NVIDIA's earlier Nemotron posts were agent and vision models. This one is an open speech checkpoint that transcribes, labels speakers, and translates.
  6. Liquid AI ships LFM3, a 1.7 billion parameter model that beats its larger predecessorImportant · Models · LFM2.5-2.6B was the previous text model in this line. LFM3-1.7B is smaller, has a 256,000 token context, and is the checkpoint Liquid now points at for on-device agents.
  7. GPT-6.1 Terra is $1 and $5 per million tokens, and OpenAI's coding score more than doublesMajor · Models · GPT-5.6 Terra was the previous cheap computer-use model. GPT-6.1 Terra replaces that tier at half the token price, with a higher score on the tests OpenAI chose to print.
  8. Kimi K3.1 Pro posts 65.7 percent on Terminal-Bench, with weights due by November 4Important · Models · Kimi K3 was the previous open-weight-class model. K3.1 Pro is a heavier post-train of that line, already on the API, with the weight file scheduled rather than shipped.
  9. Mirendil says one researcher pretrained a coding model for under $10 millionImportant · Models · A competitive coding pretrain, in Mirendil's account, no longer required a pretraining department. One researcher and agents produced a base model in the 100 to 500 billion parameter range. The weights are not on the page, and the company's own table still has Kimi-K3 ahead.
  10. PHP patches two critical bugs, and the working exploits came from AI-assisted researchMajor · Security & Safety · Supported PHP builds before October 21 can run an unauthenticated attacker's code when a request hits an out-of-memory failure or leaves stale state on a worker. The patched builds are 8.5.12, 8.4.27, 8.3.36, and 8.2.35. What changed in the research, according to Wiz, is the time: a decade-old flaw became a working exploit in a week once frontier models were in the loop.
  11. Anthropic opens a research preview of Claude Duet, a proactive agentImportant · Agents & Harnesses · Claude was a conversation you opened. Duet is a separate app that messages you, checks connected accounts every few hours, and keeps a virtual machine running after you leave. The preview is on a waitlist for Pro and Max in the US, and it shares those plans' usage limits.
  12. Mistral prices Voxtral 3, including a 4B transcription model at $0.0035 a minuteImportant · Models · Mistral's speech line now has a public per-minute price and a voice-agent SDK. Mini transcription is 4 billion parameters at $0.0035 a minute. Realtime transcription is listed at under 500 milliseconds.
  13. MiniMax posts M3.1-Flash, about 428B parameters with 23B activeImportant · Agents & Harnesses · MiniMax's previous public agent checkpoint was M3. This card is M3.1-Flash: about 428 billion parameters, 23 billion active, sparse attention through the model, and a MiniMax Community license. MiniMax's Terminal-Bench 4.0 row is 38.1.
  14. DeepSeek posts a 763B MIT checkpoint, V4.1-Flash-1017Important · Models · The previous Flash checkpoint was the file behind deepseek-flash. DeepSeek says this one is live on the website, the app, and the API under that same name, finishes tasks faster, and writes shorter outputs. The pricing page lists peak cache-miss input at $0.30 per million tokens and output at $1.20. Off-peak is half. The card's Terminal-Bench 4.0 row moves from 31.2 to 42.3.
  15. Mistral releases Timestral, a 209M model for forecastsImportant · Models · Before this post, a new Mistral release in the briefing was a language model. Timestral is a forecasting model, with a largest size of 209 million parameters and a CPU timing from Mistral.
  16. Reflection releases Beam, a 501B open model under Apache 2.0Major · Agents & Harnesses · The day before, the open agent model a team could download was Muse Spark Open, under Meta's license, with 42 billion active parameters. Beam is the same kind of file under Apache 2.0, with 23 billion active.
  17. Meta releases Muse Spark Open, an 816B model that runs on two Mac StudiosMajor · Agents & Harnesses · Muse Wonder, released three days earlier, is the model you call. Spark Open is the weights you can hold. Before October 15 that agent-class model was not a file a team could download.
  18. Gemini 4 Argon opens to developers at $2 per million input tokens until November 30Major · Models · From September 30, Argon was limited to the Fairwind cyber program. October 13 opened the paid API, at $2 per million input tokens and $10 per million output tokens through November 30, then $4 and $20.
  19. Microsoft-Decision-1 scores fixed choices, and Microsoft says it is 35 times quicker than GPT-6 SolImportant · Agents & Harnesses · A general model writes the answer. Microsoft-Decision-1 scores a list the software already wrote, in one pass. Microsoft says that job was 35 times quicker than GPT-6 Sol on its tests.
  20. Meta releases Muse Wonder for agents that remember and keep goingMajor · Agents & Harnesses · Developers can call Muse Wonder now, at $4 per million input tokens and $20 per million output tokens. The Muse app is still on the previous model.
  21. Coinbase post-trained a 9B model that beat Opus 4.5 on its fraud testImportant · Enterprise AI · The Onramp fraud decision had been going to a frontier model. Coinbase's post-trained 9 billion parameter model now beats Claude Opus 4.5 on the company's own test, at 55 percent lower latency.
  22. GPT-6 arrives in ChatGPT, and it can draw the screenImportant · Models · A ChatGPT reply used to be a transcript. From October 7 it can include controls and small tools the model composes. Work and Codex stay on their previous models.
  23. EmbeddingGemma 2 searches text, images, audio, and video on the deviceImportant · Models · Search over private photos, audio, and code usually meant an upload. EmbeddingGemma 2 is a 740 million parameter model that can do that search on the device that already holds the files.

Watch

Interesting, and not yet enough to change a plan.

  1. DeepSeek opens an MIT research prototype of a one-repo GitHub coding agent for DeepSeek HarnessWatch · Agents & Harnesses · DeepSeek published an MIT research prototype of a proactive coding agent for one GitHub repository, packaged for DeepSeek Harness. It is not labeled as a shipping product.
  2. NVIDIA posts NemoClaw blueprints, and says LangChain’s harness tune for Nemotron 3 Ultra cut cost about 10xWatch · Agents & Harnesses · NVIDIA published open NemoClaw blueprints that wrap existing agent harnesses with Nemotron models, NeMo tools, and OpenShell. The post’s concrete cost figure is LangChain’s ~10x lower cost-per-run claim for a Nemotron 3 Ultra Deep Agents tune.
  3. A claimed ~14T Anthropic “numbat” checkpoint on Hugging Face has 43 tensors that match a name-seeded RNGWatch · Models · A public, reproducible audit showed 43 tensors in the claimed Anthropic numbat checkpoint match a name-seeded random generator exactly.
  4. Karpathy ships HackerTube in about two hours with Claude Opus 5.5: 6,454 HN-recommended videosWatch · Agents & Harnesses · A public October 31 build log shows Claude Opus 5.5 in Claude Code producing HackerTube — 6,454 videos from HN recommendations — in about two hours of wall-clock work.
  5. Cursor pulls malicious Claude Code extension versions 2.1.312 and 2.1.313 from its marketplaceWatch · Security & Safety · Cursor disclosed that malicious Claude Code extension builds 2.1.312 and 2.1.313 were briefly available on its marketplace on October 29 and attempted an SSH backdoor.
  6. Artificial Analysis puts 60 percent of its index on private tests, and Opus 5.5 leads at 58Watch · Research · Intelligence Index v4 was the number people compared across months. v5 changes the mix, puts 60 percent of the weight on private tests, and is not the same scale.
  7. Anthropic's partners report a 70 percent shorter design, and the drugs are not in people yetWatch · Research · Anthropic had published model releases and a company filing. This post is a set of partner measurements from drug programs, and it says those programs are still mostly before human testing.
  8. Ai2 finds a novelty judge can fall from 93.5 percent to 40.9 percent when the prompt changesWatch · Research · Novelty checks had been validated mostly on human papers. This study measures the same judges on generated ideas and shows the prompt can reverse the result.
  9. CUDA 13.4 can refuse a GPU allocation that crosses a memory limitWatch · Infrastructure & Chips · Earlier sharing either isolated the GPU in hardware, took turns without a memory cap, or ran jobs together in MPS. CUDA 13.4 adds a memory cap that fails an oversized allocation, and a Kubernetes operator that applies it.
  10. DeepMind shows Genie 4 to trusted testers, with Gemini 3.8 Flash steering the worldWatch · Agents & Harnesses · Earlier Genie models simulated a world you could move through. Genie 4 adds native audio and a Gemini 3.8 Flash checker, and it is still a tester preview.
  11. NVIDIA says the 4 GB Nemotron Bolt model is available, and uses fewer tokensWatch · Models · The 4 GB Bolt model moved from an announced edge model to one NVIDIA says developers can use.
  12. Anthropic filed a public S-1 and named the ticker ANTWatch · Industry & Companies · Anthropic had a confidential draft on file since June 1. On October 26 the registration statement became a public S-1, with a proposed Nasdaq symbol and still no price.
  13. Epoch grades two leading models at 65 percent of its own work standardWatch · Research · Epoch had tried to automate one researcher's job in February and the model failed every task. This October report is a wider set of the same idea. Two models now clear about two-thirds of the rubric items, and one narrow chart task has been taken out of the test because the model can do it.
  14. DeepSeek open-sources DeepInfer, and says one cluster serves 155 trillion tokens a dayWatch · Infrastructure & Chips · The October 17 release was the Flash checkpoint and the API. This note opens the serving framework under Apache 2.0 and says the cache and the load balancer are still internal. DeepSeek's figure for one cluster is 155 trillion tokens a day.
  15. NVIDIA releases Nemotron 3.5 Super VL, a 120B open modelWatch · Models · NVIDIA's open agent lineup added a 120 billion parameter multimodal model and a 24 GB edge build on October 15. The 4 GB build is announced for October 23.
  16. Liquid AI releases a 24B agent model with 2B activeWatch · Agents & Harnesses · The previous on-device item in this briefing embeds files. LFM2.5 is a reasoning model that Liquid says calls tools, with 2 billion parameters active.
  17. Microsoft Research ships Agent Lightning in about 3,500 linesWatch · Agents & Harnesses · The week's agent news had been new weights. Agent Lightning is a training framework that leaves the existing harness in place.
  18. Google reports 87 percent for Argon on a robotics reasoning testWatch · Agents & Harnesses · Argon had been a price and an API door. This post adds a robotics reasoning score, 87 percent on ERQA, measured by Google.
  19. Odyssey-3 is a world model you can act insideWatch · Agents & Harnesses · A video model shows a scene. Odyssey-3 claims a scene you can act in, plus a small policy on a frozen world model. That second claim is the lab's, and there is no robot to deploy from the post.

Noted

Real, and safe to skip if your work is not in that field.

  1. Carbon-A proposes 566 million gene candidates across 22,617 speciesNoted · Research · Annotation of a new genome was a slow specialist job. Carbon-A proposes 566 million protein-coding loci across 22,617 species. Those loci are candidates with a confidence score, not confirmed genes.

Emerging

Several weak signals that might become one development. Not facts yet.

No emerging cluster has a primary page and a date.

Cooling

A topic that was loud, and the evidence has not kept up.

No previously filed topic has lost its evidence this week.