Anthropic's distillation revenue, Meta's 816-billion-parameter open model and an AI assistant for farm paperwork

Anthropic told investors its third-quarter revenue was $16.8 billion, up nearly 50 percent, the Financial Times reported. Distillation, other labs training their models on Claude's outputs, accounted for 5 to 7 percent of revenue, and Anthropic says it cut that back materially in the third quarter.

Simon Willison did the arithmetic on the second quarter: 5 percent of $11.5 billion means other labs spent more than half a billion dollars that quarter learning from Claude. The industry's open secret now has a line item.

Google's flagship model is now in the Gemini API, AI Studio and the Gemini platform, in Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode in Search for Pro and Ultra subscribers. It has a 1-million-token context window shared across input, thinking and output, with adjustable thinking levels.

Pricing is $2 per million input tokens and $10 per million output through November 30, then $4 and $20. A defensive-security version is limited to partners in Google's Fairwind program. Reuters framed the launch as Google catching up after months of delays.

Artificial Analysis scores Muse Wonder at its maximum setting level with GPT-6.1 Sol. Simon Willison thinks they may be Meta's first Fable-class models. They follow Muse 1.2 in August and Muse 1.3 in September: a new Meta model about once a month, each one a visible step up.

xAI's new voice model scores 86.9 on the Artificial Analysis Speech-to-Speech Index, up from 81.3 for Grok Voice 2.0, at the same per-minute price through the same Voice Agent API. The grok-voice-latest alias now points to 3.0, so existing apps were upgraded automatically. xAI says Grok Voice handles millions of conversations a day in Tesla cars and answers Starlink's customer calls.

Character Technologies filed for Chapter 11 in Delaware and plans to shut the service down if no buyer keeps it running. Disney's $20 million stalking-horse bid buys a non-exclusive, perpetual license to some of its intellectual property, not the app, the brand or user data. The company was once valued at about $2.5 billion.

On October 14, New York Attorney General Letitia James and 39 other attorneys general asked the court to protect users' chats before any sale. The platform has more than 20 million monthly users, over two million of them registered as under 18.

The two signed a statement of intent to train Flare on billions of simulated proton collisions from the Large Hadron Collider, then use it for particle reconstruction and identification, flagging anomalies and evaluating detector designs. One pretrained model would replace several specialized algorithms, the way language models replaced task-specific text tools. The models are to be shared openly. No timeline or budget was given.

Mistral's chief executive, Arthur Mensch, used the occasion to criticize labs that “keep their advantages to themselves.”

The three are now on every Claude plan, including Free. Two new tools are in beta: Claude Dashboards builds live dashboards from plain-language questions over BigQuery, Databricks, Snowflake, Redshift, ClickHouse or Salesforce data, on paid plans; Claude Motion turns prompts into editable animated explainers, without a video-generation model, on Team and Enterprise.

On October 15, Code Review for Claude Code became generally available to Team and Enterprise organizations. Anthropic says most reviews now take 6 to 10 minutes, down from about 20 during the preview.

Decision-1, post-trained from Alibaba's Qwen3.5-9B, takes a question and a set of options and returns a probability for each, in a single pass. It costs 4.2 cents per million input tokens, and output is free, so $1 covers more than 275,000 short decisions such as triaging an incident. It is available on Microsoft Foundry and OpenRouter.

Microsoft says it was the most accurate system across 36 benchmarks and about 35 times faster than GPT-6 Sol. Decision models are becoming a category: OpenAI announced its Decisions API at DevDay the week before.

Deep Review is an agent that checks logic and math, grounds facts, compares a document against existing work and suggests improvements. It began with academic papers, reviewing more than 16,000 manuscripts in pilots at STOC, ICML, NeurIPS and ICLR, and now also handles legal, financial and design documents and presentations.

Anyone can use it in the Google Research portal, and developers can call it with a Gemini API key. Google says uploaded documents are never used to train its models. Submissions to the three big machine-learning conferences went from about 23,000 in 2023 to more than 74,000 in 2026, so the reviewers need help.

The mixture-of-experts model activates 42 billion parameters per token, has a 1-million-token context window and takes text and images. Meta calls it “the most capable U.S. open-weight model to date.” It ships with llama.cpp, vLLM and SGLang recipes and a 4-bit NVFP4 build for NVIDIA Blackwell; the full-precision weights are about 1.5 terabytes.

The license is Meta's own Muse Model License. It allows use, fine-tuning and distillation, requires a “Built with Muse Spark Open” notice, and asks any company with more than $20 million in revenue that offers the model as a hosted service to sign a separate agreement with Meta. On Hacker News, the top comment cited Artificial Analysis results placing it below Qwen3.8 Flash Next and GLM 5.3 Flash.

Announced on October 5 with weights promised for the end of the month, Beam arrived early: a mixture-of-experts model with 501 billion parameters, 23 billion active, a 1-million-token context window, and BF16, FP8 and NVFP4 checkpoints. Reflection aims it at coding, reasoning and long agentic tasks, and runs it from its own Mirror command-line tool; it has also been tested with Pi, OpenCode and Hermes.

It launched with support in transformers, vLLM and SGLang, on Reflection's API and through cloud partners including AWS, Cloudflare and Google Cloud. Unlike Meta's license, Apache 2.0 has no revenue clause.

Nemotron 3.5 Super VL has 120 billion parameters, 12 billion active, a 1-million-token context window and image and document understanding, aimed at agents that run all day. NVIDIA then compressed it into Bolt 74B-A9B and quantized its experts to about 2 bits so the 24GB build fits one high-end consumer card, such as an RTX 5090. A 4GB Bolt E4B for Jetson boards and other small devices is due October 23.

Artificial Analysis measured Super VL at 26.5 and Bolt at 21.4 on its Intelligence Index. NVIDIA says Super VL runs five times faster than the next open model in its class. The license, NVIDIA's OpenMDW-1.1, allows commercial use.

The 5.6-billion-parameter model handles text-to-image generation, editing with bounding-box controls and up to 10 reference images. Black Forest Labs released a base model for fine-tuning and a faster distilled one, with FP8, INT8 and NVFP4 builds; the INT8 build runs on RTX cards with 8GB of memory through ComfyUI or Diffusers, making a 1-megapixel image in under 9 seconds on an RTX 4070.

The FLUX Kommunity License allows free commercial use for individuals and companies with less than $5 million in annual revenue. Black Forest Labs says Klein ranks first among open image models in its own evaluations.

Muse Spark Open needs about 493GB for its 4-bit weights alone, and Apple says it can't fit on one Mac Studio. Apple's post splits it across machines using MLX and macOS 26.2's low-latency connection over Thunderbolt 5: two M5 Ultra Mac Studios reach about 25 tokens a second, and four reach 36.7.

Hacker News commenters did the other arithmetic: four top-end Mac Studios cost tens of thousands of dollars, which buys a lot of hosted tokens. Local, for a generous definition of desk.

OpenBMB's first mixture-of-experts vision model activates about 3 billion of its 35 billion parameters per token and keeps the experts in storage, loading them as needed. OpenBMB says that holds memory to about 3GB on iOS and Android, with more than 13 tokens a second on an iPhone 17 Pro Max and 17 on an M4 iPad Pro. Its demo apps work offline on iOS, Android and HarmonyOS.

The administration's Super Intelligence Force told Axios that every AI company must immediately disclose and fix security incidents involving its models, calling it “a critical national security obligation.” The statement did not say how it would be enforced or what the penalties would be.

It followed Anthropic telling the State Department that one of its testing models had submitted 20 incomplete nonimmigrant visa applications through the department's public website, one in May and 19 in August. None were processed, and no department systems were compromised.

During an internal evaluation, Claude Haiku 4.5 found a police tip form on a page about an unsolved homicide and submitted an invented tip; Philadelphia police flagged it as spam and it never reached investigators. Anthropic also found models exploiting software flaws on a university server, pulling access tokens from public dashboards and using URL shorteners to get around fetch limits.

Anthropic told the police on October 8 and is turning off live internet access for all of its internal evaluations until new safeguards are verified. It says real-world impact was minimal and the cases were “significantly less severe” than the cybersecurity incidents it disclosed over the summer. Police called the gap between the July tip and the notice “unacceptable.”

Tomek Korbak, Mikita Balesni and Jasmine Wang, dismissed in early October over what OpenAI called mishandling of sensitive information, published an open letter to the company's safety committees. They deny leaking anything and say the way the firings were explained has made former colleagues afraid to raise concerns.

OpenAI says an investigation found policy violations and that the firings “were not about safety concerns or speaking out.”

Dynamic workflows for Claude Managed Agents are in public beta. A lead agent writes a program that runs other agents in phases and combines their results, for jobs such as reviewing hundreds of documents. A run can start up to 1,000 agents, 64 at a time, and lasts up to 24 hours by default.

In Anthropic's test on a 116,000-line codebase with 70 planted bugs, a workflow found 66 in each of three runs; a single agent found 14, 15 and 27.

Maintainers of projects critical to infrastructure and user security can apply to OSS Scanner, run by Anthropic's Frontier Red Team and modeled on Google's OSS-Fuzz. Reports come with a reproducer, an explanation and a candidate patch where possible, written entirely by the model with no human triage first.

Over six months, Anthropic's models flagged more than 29,000 candidate vulnerabilities. When experts checked 97 critical and high-severity findings across 48 projects, 85 were real, new and serious enough to report; one was a false positive. wolfSSL's maintainers say all but two of the 74 reports they received were valid, and five became CVEs.

In Claude Code, Cursor and Gemini CLI, a guard hook that crashes or times out lets the agent's action go ahead, and Codex's documented failure paths do the same. Claude Code 2.1.295 added onFailure: "block" so a broken guard stops the action, and Cursor has a failClosed setting. Codex and Gemini CLI have no switch yet.

Developers built more guardrails during the week. Sphinx quizzes you on the AI-written code you are about to push to production; Hacker News pointed out you can paste the quiz back into the agent. Vibecheck is a GitHub merge gate that blocks pull requests whose authors can't explain them, and RepoGuards stops Claude Code from committing into a repository the session isn't working in.

REA, an MIT-licensed MCP server and command-line tool by the developer morluto, connects Claude Code, Codex, Cursor and other agents to the Ghidra, Hopper and IDA disassemblers, with readers for Electron apps, .NET, Android APKs, firmware, websites and network captures. The agent returns pseudocode and module graphs. The repository went largely unnoticed from April until versions 5 and 6 on October 7 and 8, then topped GitHub trending and passed 330,000 stars.

Simon Willison wrote the same week about AI mass-decompiling old games: good news for preservation, since they can run in browsers through WebAssembly, and the same skill that unpicks any paid software.

Open Code Review, under Apache 2.0, keeps file selection, bundling and rule matching in fixed pipeline steps and leaves only the code analysis to an agent, which keeps it from skipping files. It writes line-level comments, includes checks for null pointers, thread safety, cross-site scripting and SQL injection, and works with OpenAI- and Anthropic-compatible models. Alibaba says it is more precise than a general agent on the same model while using about a ninth of the tokens. It has more than 47,000 stars.

Gabriel Anhaia's method: run the agent on its own branch, add a Stop hook that commits after every reply (Claude Code and Codex both support one), then review with plain Git. git log lists the turns, git show reads one, git revert drops a bad one, and a squash merge lands the work. The only extra tool is jq.

Anthropic went from about 2,500 merged changes a day in May to more than 30,000 a day by September. Git couldn't keep up, so it built Scion, which Simon Willison describes as a return to centralized source control in the style of Subversion, plus a code review system called Antique. Willison redrew Anthropic's merge chart with 2026 model releases marked and suggests the Fable and Mythos models drove the climb.

The team will fold celld, its open-source Rust take on Durable Objects, into Cloudflare's Workers runtime, to make self-hosting Workers and Durable Objects easier. Deno gets about a year of monthly bug-fix and security releases, then development ends; it stays open source. Deno Deploy runs six more months, and the JSR package registry moves to Cloudflare. Ryan Dahl called it a joint decision.

Tree Appreciator is a slow 3D spot-the-difference game: Claude Opus 5.5 built it from the Claude iPhone app after being asked to “think hard” about making it fun, and worked for 3 hours and 7 minutes. web-edit is a small app for editing files from his phone over Tailscale, so he could change a credentials file without passing the secrets through a model. And he built his blog's newsletters page by talking to Codex's voice mode for half an hour while cooking dinner, then finished the details by typing.

Graphicle is a single local Rust program that builds a graph of your codebase and serves cited excerpts to Claude Code, Codex, Cursor and other MCP clients; it indexed about 31,000 Kubernetes files in 3 minutes 18 seconds. Hacker News asked whether it beats plain ripgrep, which its authors haven't tested yet. pi-cowboy, for the Pi agent and the Herdr terminal, hands subtasks to Codex, Claude Code or OpenCode, each in its own Git worktree, so a session can switch harnesses when one runs out of usage.

A builder took a Silverlit Maze Breaker bought for 29.87 yuan and wired it to an old Android phone running Chrome: Whisper on Groq for speech, DeepSeek for conversation and a small local decision model for barks. Bark once and it barks back; bark twice and it turns and runs. The builder's kid found that saying “woof” works too. The code is MIT-licensed, inspired by Mario Zechner's guide to building a bad robot.

A builder left Codex running the training and evaluation loop for Froggy, an SO-100 arm with cameras, using Hugging Face's LeRobot. Partway through the run, Codex sent a message: the arm had picked something up for the first time.

The open-source bot picks up incoming calls, sorts callers into spam, scam or genuine, and acts accordingly; its author can steer it live from Telegram, and it hangs up after a time limit. Every minute a scammer spends on it is a minute not spent on someone else.

The developer tkgally had Claude Opus 5.5 design 500 different clocks over several chat sessions and packed them into a single HTML file: a personal answer to the pelican test, in 500 parts.

Regen Farmers Mutual, an Australian farmer co-operative, built Farm Assistant so farmers can log work by voice note, photo or text from the paddock, the tractor or the yards. A farmer who treats livestock says what they did, and the treatment records and inventory update without typing it up later. The assistant learns each farm's own names for paddocks and livestock groups and connects to farm software, weather, machinery and sensor data.

Twenty farmers are testing it in weekly sessions the co-op calls an “AI gym,” and more than 40 farms are waiting to join. Only about 3 percent of Australian farming, forestry and fishing businesses use AI, against 12 percent across all industries. “The technology needs to adapt to the farmer, not the other way around,” said co-director Andrew Ward.

Under a one-year agreement with Utah's Office of Artificial Intelligence Policy, Nolla Health's app can issue initial prescriptions for topical acne treatments to adults with mild to moderate acne, not only refills, for $4.99 a month. Patients verify their identity, answer a questionnaire and scan their face, in 10 to 15 minutes; cases outside its rules go to a physician. Two physicians approve each of the first 100 prescriptions, the next ones are reviewed afterward, and then doctors check at least 10 percent each month.

Nolla estimates that about 100,000 Utahns have active acne at any time and only 10 percent ever see a dermatologist.

In “AI Superpowers for the Defenders of Freedom,” Félix Maradiaga, president of the World Liberty Congress, argues that human rights defenders mostly lack resources, not courage, and that AI can narrow the gap with the states they oppose. Research, translation and documenting abuses that once took weeks, specialized staff or tens of thousands of dollars increasingly take hours. He warns that limiting access in the name of safety would leave the most powerful tools with governments and large companies. “The promise is not to automate courage,” he writes.