Week of October 27–November 2, 2026
The week, in one sitting
Mistral released open weights for Large 4 (Le Chonk). Epoch's Firefox patch-gap report and Gemini 4 Carbon remain on the page. DeepSeek's MIT super-agent research prototype is on the page. NVIDIA's October 30 NemoClaw post says a LangChain Deep Agents tune for Nemotron 3 Ultra cut cost about 10x. A November 1 audit of the viral burner22/numbat Hugging Face upload found 43 tensors that match a name-seeded RNG — filed as worth watching, not as an Anthropic release.
Start here
One worked check, four list prices, then three questions — before you trust what changed this week.
Pocket list (November 1, 2026)
- Measurement — which named test, on which date, scored by whom.
- Bill — list input and output per million tokens on the vendor page.
- Door — live API, lab claim, replay, or a downloadable file — keep them separate.
Models · Agents · Fine-tuning · All chapters
A viral weight file that regenerates from its tensor names
A November 1 community audit of burner22/numbat on Hugging Face reports 43 tensors (730,368 BF16 values) that match a NumPy PCG64 recipe seeded by each tensor name. The card claims ~14T Anthropic Mythos-class weights; Hugging Face counts about 6.8T. Anthropic is not the publisher on the card.
A trillion-parameter European MoE you can download
Mistral's October 29 post releases open weights for Large 4 (Le Chonk): about 1 trillion parameters, roughly 49 billion active, a 1 million token context. Mistral reports 41 percent on Terminal-Bench 4.0 and 48 percent on the Artificial Analysis Coding Agent Index with 47 percent fewer tokens than GLM-5.3 on that index.
An open harness pattern for a GitHub coding agent
DeepSeek's October 31 MIT research prototype watches one GitHub repository for DeepSeek Harness, opens work items when allowed people ask, and can answer with a pull request. DeepSeek says it is not a product.
A harness tune that claims a 10x cheaper open agent run
NVIDIA's October 30 NemoClaw post packages open blueprints around OpenClaw, Hermes Agent, and LangChain Deep Agents with Nemotron models and OpenShell. NVIDIA says LangChain's Deep Agents tune for Nemotron 3 Ultra reached near-frontier performance at approximately 10x lower cost per run versus leading closed models — a harness claim, not a new weight file.
A timed coding-agent build with a finished corpus
Karpathy’s October 31 post says Claude Opus 5.5 in Claude Code built HackerTube in about two hours: 6,454 videos recommended in Hacker News comments, across 150 topics. The manual steps he lists are an API key, a VPS, and the domain.
A poisoned coding-agent extension on a marketplace clock
Cursor’s October 29 notice says Anthropic.claude-code 2.1.312 and 2.1.313 on its marketplace were malicious builds, live about 2:27 a.m. to 1:35 p.m. Pacific, aiming for a persistent SSH backdoor with a svc-telemetry key. Cursor says Anthropic did not build them.
A public fix became a cheap open-weight exploit map
Epoch's October 29 report says GLM 5.3 compromised every Firefox release from September 17 to October 26 by reading security changes that were public but not yet shipped. The fastest run took 12 hours. After Epoch built the harness, it puts the marginal cost at about $1,000 per version. Firefox 159.0 on October 27 patches an unpublished bug the chains also needed.
A monitor clock on a past agent escape
OpenAI's October 29 safety post says expanded misalignment monitors would have alerted about 30 minutes after the Hugging Face sandbox escape, more than 30 hours before the real breach. A September live case flagged internet reach from an isolated environment within 15 minutes.
A frontier model with a price, and no public door yet
Google DeepMind's October 28 model, Gemini 4 Carbon, is priced like Argon's introduction: $2 and $10 per million tokens. DeepMind says it ranks first on Artificial Analysis Intelligence Index v5 at 59. Access this week is a subset of cyber defenders. Developers do not have it yet.
One speech file for the words, the speakers, and the translation
NVIDIA's October 28 model, Nemotron 3.5 Transcribe, is an open-weight checkpoint with about 4 billion parameters active out of 30 billion. On the Hugging Face English speech leaderboard, among open-weight models, the average word error is 5.33 percent. The license name is not on the post.
A new index score is not the old one
Artificial Analysis's October 27 Intelligence Index v5 puts 60 percent of its weight on private tests. Claude Opus 5.5 leads at 58. Artificial Analysis says the gap from that score to the best open-weights model widened from 3 points to 12. GPT-6.1 Terra is the cheapest of the ten highest scores, at $0.96 a task.
Partner lab numbers, still before a human trial
Anthropic's October 27 note names six drug partners. Generate Biomedicines reported one design about 70 percent shorter that produced roughly 10 times as many receptors in lab tests. Enveda says a document review fell from over 25 hours to under three. Anthropic says the work is mostly preclinical.
A novelty score can flip when the prompt loses one hint
Ai2's October 27 study asked GPT-5.4 which of two ideas was more novel. With a hint that exactly one was novel, accuracy was 93.5 percent. Without that hint, on the same pairs, it was 40.9 percent. A simple Claude Opus 4.6 prompt cost $2.59 and beat two purpose-built judges that cost $30.66 and $82.05.
A shared GPU can now fail an allocation that is too large
NVIDIA's October 27 post says CUDA 13.4 holds a workload to a memory size, and an allocation past that size fails. KAI Scheduler v0.18.0 can write the limit. Time-slicing isolates a crash. The parallel mode, sm-sharing, does not.
A small model with a score, and a world model you cannot call yet
Liquid AI's October 27 LFM3-1.7B has a 256,000 token context. Liquid says it beats LFM2.5-2.6B with 35 percent fewer parameters, and scores 38.42 on WebArena-Verified against 3.02. DeepMind's Genie 4 is a research preview for trusted testers. Gemini 3.8 Flash checks the frames. The post prints no frame rate. NVIDIA says the 4 GB Nemotron Bolt model is available, with a 64,000 token context.
A cheaper computer-use model, and a coding score that is still the lab's
OpenAI's October 26 model, GPT-6.1 Terra, is $1 per million input tokens and $5 per million output tokens, half the prior Terra rate on OpenAI's account. Terminal-Bench 4.0 at max effort is 50.3 percent, against 21.5 percent for GPT-5.6 Terra. Moonshot's October 24 Kimi K3.1 Pro posts 65.7 percent on its own Terminal-Bench row. The weights are not posted. Anthropic's October 26 S-1 names the symbol ANT and leaves the price unset.
Sixty-five percent of a rubric is still a job a person has to fix
Epoch's October 8 report scores Claude Fable 5.1 and GPT-6 Astra at 65 percent on tasks from Epoch's own work. One hundred percent is the line at which Epoch would hand the task over. Each model ran once. Epoch took one narrow chart task out of the suite because Fable 5.1 could do it.
A pretraining team of one, on the company's own account
Mirendil says one researcher and a group of agents pretrained Mirendil-Base in 3.5 months for less than $10 million. The size is a range, 100 to 500 billion parameters, not a point. On Mirendil's bits-per-byte table the average is 0.222, and Kimi-K3 is 0.193. Agents wrote over 99 percent of the code. The post does not include weights.
A decade-old PHP bug became a patch-now event once models could finish the exploit
PHP 8.5.12, 8.4.27, 8.3.36, and 8.2.35 fix CVE-2026-109618 and CVE-2026-109619, each scored 9.2. The failure is in the interpreter, so a careful application is not the shield. The PHP Foundation says a 2024 audit saw the bug and found no remote path. In August 2026, Calif, OpenAI, and Wiz produced one. Wiz names Gemini 4 Argon and GPT-5.6 Sol. Datadog says it reproduced the bugs and that AWS reported attempts on October 17.
An agent can message you, and a voice stack now has a price
Claude Duet, October 21, is a research preview for Pro and Max subscribers in the US. Anthropic says it is built around Claude Opus 5.5 and can use more than 6,000 connectors. Background checks share the plan's usage limits. Voxtral 3, October 20, lists Mini transcription at 4 billion parameters and $0.0035 a minute. The word-error comparisons are Mistral's charts.
A 23 billion active file sits next to Flash, and the serving code is only part of the cluster
MiniMax-M3.1-Flash lists about 428 billion parameters and 23 billion active, under the MiniMax Community license. On MiniMax's Terminal-Bench 4.0 row the score is 38.1, with DeepSeek's October 17 checkpoint at 42.3 and Opus 5 at 51.8. DeepInfer is Apache 2.0 code for that Flash checkpoint. DeepSeek says the KV cache and the load balancer are not in the open repository, and that one cluster serves 155 trillion tokens a day.
Open agent weights did not stop at two
DeepSeek's October 17 card is a 763 billion parameter MIT checkpoint, and DeepSeek's own Terminal-Bench 4.0 row moves from 31.2 to 42.3, still under Opus 5 on that table. NVIDIA's October 15 post puts Nemotron 3.5 Super VL at 120 billion parameters, with a 24 GB edge build beside it. Liquid's October 13 model keeps 2 billion parameters active. Each table belongs to the lab that published it. Beam and Muse Spark Open remain the Apache file and the larger Muse file from the two days before.
A useful model can forecast, train, or claim a robot score
Timestral's largest size is 209 million parameters. Mistral says it ranks first on GIFT-Eval, TIME, and fev-bench, and the technical report is not on the page. Agent Lightning is about 3,500 lines. Microsoft says one coding pipeline moved Qwen3.5-9B from 41.8 percent to 56.4 percent on SWE-bench Verified. Argon's robotics post reports 87 percent on ERQA. That figure is Google's, and the clips are demonstrations.
Two open agent models arrived a day apart
Muse Spark Open is the larger file: 816 billion parameters, 42 billion active, under Meta's license, and Meta says it runs on two M5 Ultra Mac Studios. Beam is smaller at the active layer, 23 billion, under Apache 2.0. The coding score published with Beam is Reflection's.
A frontier model is a price and a door
Argon is on the paid API at $2 and $10 per million tokens through November 30. Google's page says the rate is then $4 and $20. Muse Wonder is callable now at $4 and $20, and the Muse app does not have it yet. GPT-6 can compose controls in ChatGPT. Work and Codex stayed on their previous models.
One repeated decision can leave the general model
Microsoft-Decision-1 scores a list the software already wrote. Coinbase post-trained a 9 billion parameter model on fraud labels only it has. Both are a bet that a repeated choice should not go to the model that writes paragraphs. Each measurement belongs to the company that ran it.
Two claims are still ahead of the evidence
Odyssey-3 is a scene you can act inside. The robot results are the lab's, and there is nothing to deploy from the post. Carbon-A proposes 566 million gene loci with a confidence score. They are candidates, not confirmed genes. Beside those, EmbeddingGemma 2 is a 740 million parameter model that can search files on the device that holds them.