GPT-5.6 Sol gets 20% cheaper while 290B models move to consumer GPUs

Anthropic's Thariq Shihipar clarifies a Claude Code test config caused "10" reasoning effort display: Users panicked when Claude reported "10" on high effort, assuming it was 10/100. The number is a display quirk from a server-side experiment, the actual reasoning effort matches what users select and model performance is unchanged.

Linus Torvalds Endures a Debug Session from Hell, "Enormously Helped" by AI: The Linux creator personally debugged and patched an Intel Xe graphics driver bug on a Battlemage G21 card that caused GDM to crash-loop. He credited an AI for doing "much of the grunt-work," marking a rare moment of Torvalds both writing graphics driver code and praising AI assistance. The fix is already merged into the kernel tree.

China Is Building AI Models of American Voters: An investigation by War Room co-host Natalie Winters documents how Chinese universities and state-affiliated think tanks are constructing simulated U.​S. voters, labeling over a million X accounts by political stance, running synthetic Pennsylvania swing-state experiments, and using LLM agents to model "America First" attitudes toward tariffs and alliances. The report links to published papers (ElectionSim, SocioVerse, and a Wuhan University study) that confirm the work is real and ongoing.

Charlie Marsh reworked Codex CLI's lifecycle to make startup ~25x faster and immediately responsive: Astral founder Charlie Marsh (now at OpenAI working on Codex) shipped a change that eliminates the startup delay when running the codex CLI, making the agent feel instant. If you bounced off Codex because of slow launch times, this is the update to try.

Anthropic introduces Agent Skills to equip Claude with specialized instructions and tools: Agent Skills allow Claude to load dynamic sets of instructions and resources for specific tasks. One such internal skill shared by Anthropic engineer Thariq Shihipar uses the /eli5 prompt to break down complex topics into simple HTML artifacts with minimal text and large visuals.

The new MCP Roadmap focuses on agentic messaging, HTTP-native transport, and enterprise security: The Model Context Protocol team published an updated roadmap organized around five priority areas: agentic messaging primitives (server-initiated events, loops), HTTP-native transport unification (building on the July 2026 HTTP transition), agent identity and enterprise-ready security (OAuth, DPoP, client assertion flows), improved primitives, and better SDK developer experience. Each area has dedicated Core Maintainers and Working Groups.

English ↔ Claudish Translator, an over-engineered tool that translates plain English into the verbose, corporate style Anthropic's Claude is known for: A playful demo of Program-as-Weights technology that converts everyday English into "Claudish", that unmistakably formal, jargon-laden tone Claude uses in every response ("The translation did not land cleanly. Please try again."). Fun as a party trick and surprisingly accurate as satire.

OpenAI cuts developer pricing for frontier GPT-5.6 Sol model by more than 20%: API and Codex credit pricing for OpenAI's most capable model drops to $5 input / $30 output per 1M tokens for at least three months (through November 21), as the frontier-model price war intensifies against Anthropic and others. The Terra and Luna models in the GPT-5.6 family also received cuts of 20% and 80% respectively. Pro, Plus, and Business subscription pricing is unchanged.

Browser Use runs Qwen 3.8-27B at 400 TPS on 2x B200s, outpacing human browsing speed: Browser Use is an open source library that lets AI agents drive web browsers. By pairing Qwen 3.8-27B with DFlash2 speculative decoding on two B200 GPUs, they hit 400 tokens per second, fast enough that website loading, not the model, becomes the bottleneck.

FreeToken runs 290B+ MoE models on consumer GPUs by streaming experts across GPU, CPU, and system RAM: A team from UC Berkeley's FlashML lab (including Kurt Keutzer, Song Han, Matei Zaharia, and Ion Stoica) open-sourced FreeToken, an inference engine that treats every byte of memory in a personal machine as a unified pool for Mixture-of-Experts models. By dynamically keeping only active experts in VRAM and leaving inactive ones in system RAM, it lets a laptop with an 8GB RTX 4060 run a 35B MoE model at 39 tok/s, or a gaming PC run the 284B DeepSeek model interactively. The catch is that this only exploits MoE sparsity, so dense models don't benefit, and initial model conversion takes hours. A 284B model still needs around 192GB of system RAM, making a powerful workstation necessary for the largest models. (open-sourced FreeToken)

Michael Malis shows how to JIT compile code in 5μs: The pgrust co-creator walks through building a fast JIT compiler with AI assistance that compiles every SQL query rather than a subset. The post includes a full example building a regular expression engine.

NanoGPT (124M) in 90 seconds: the modded-nanogpt speedrun repo: OpenAI MTS Keller Jordan's collaborative repo documents the journey from Andrej Karpathy's llm.​c GPT-2 baseline (45 minutes, 10B tokens) to under 90 seconds and under 400M tokens on the same 8×H100 hardware, using techniques like rotary embeddings, QK-Norm, ReLU², and the Muon optimizer. It is both an education in extreme systems optimization and a practical starting point for fast small-scale GPT training.

Frontend design skill pack for AI agents with machine-enforced quality gates: Instead of dumping all frontend rules into an agent's context, this pack uses a 2,099-token router that loads exactly one of 19 skills per request out of ~350k tokens of reference depth. Every output passes 11 release-blocking gates before shipping, including one that checks the pack's own docs against its own rules. Install it for Cursor, Copilot, or any agent that accepts a skill pack.

OpenCode's Dax Raad pranks the AI community with a fake "Ox Alpha" technical explainer ending with "your mom": A mysterious reasoning model called Ox Alpha appeared on OpenRouter on August 20 with a 1M context window, free for a week, and nobody has claimed it. OpenCode creator Dax Raad leaned into the speculation with a plausible-sounding but entirely fake technical description built from AI buzzwords that builds up to a punchline.

OpenConnector puts an open-source auth gateway between AI agents and your apps: Every agent that reads your email or updates Notion needs your credentials, and handing over raw passwords or API keys is a hard no for most people. OpenConnector, from OOMOL, replaces that with a managed OAuth layer: connect your SaaS accounts once, then let any agent access them through scoped tokens. Supports 1,000+ providers and works with Claude Code, Cursor, Cline, and other agents via MCP, CLI, or SDK. Self-host with Docker or Cloudflare Workers.

onlygains.​ai Co-Founder Chaitanya Giri Goswami open-sourced Munder Difflin, an agent harness to run an office of your clones: Munder Difflin wraps a dozen CLI agent tools (Claude Code, Codex, Grok, Qwen, Gemini CLI, OpenCode, and more) into a single desktop app with an office-themed simulation where clones appear as employees on a virtual floor. It runs locally, uses your existing API subscriptions, and is free and open source under MIT. Grab the latest release from the GitHub repository.

Faber, an open-source coding agent that uses a code graph to navigate repos efficiently: Most coding agents read hundreds of files per query, burning tokens on irrelevant code. Faber instead builds and incrementally maintains a lightweight call-and-import graph of your repo, answering questions with a ~50-token graph query before touching files. It runs as a terminal REPL with diff approval, undo/redo, session memory, and support for Claude, OpenAI, and local Ollama models.