- The AI Report
- Posts
- 🤖 Claude Fable 5.2 Takes Top Spot + RTX Spark Crawls On Linux + AI Spending Stalls
🤖 Claude Fable 5.2 Takes Top Spot + RTX Spark Crawls On Linux + AI Spending Stalls
Plus: Claude Code leaks secrets to GitHub, Meta undercuts Haiku with Muse Ember, GPT-6 Astra quietly underperforms on safety work, and Meta and OpenAI buy coding-tool startups
The AI Report
November 10, 2026
In partnership with
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
Anthropic's Fable 5.2 takes the top spot at a lower list price
Claude Fable 5.2 is now the top-scoring model on the Artificial Analysis index, at a lower list price than Fable 5.1.
Plus, a plugged-in RTX Spark laptop ran a local model about five times slower on Linux than on Windows, and Ramp data shows AI spending stalling as companies ration tokens.
🤖 Claude Fable 5.2 takes the top spot and cuts its list price
Anthropic shipped Claude Fable 5.2, its most capable generally available model, built for long, hard technical work. It costs $8/$40 per million tokens, down from $10/$50, and generates up to 30% faster than 5.1. Mythos 5.2, the same model with looser cyber and bio safeguards, is limited to vetted organizations.
- 65 on Artificial Analysis — Fable 5.2 scores 65 on the Artificial Analysis Intelligence Index, first overall, five points ahead of Gemini 4 Carbon and six ahead of Opus 5.5.
- Verbose and pricier per task — Artificial Analysis calls it very verbose. At max effort it averages about $26.50 per index task versus $10 for Opus 5.5, though at High effort it still scores 62 on far fewer tokens.
- Not the daily driver — Anthropic still recommends Opus 5.5 for most work, with Fable as an Advisor or Orchestrator for smaller models. A fast mode for select organizations runs up to 2.5 times faster at double the price.
Don't switch wholesale. Route the hardest open-ended work, like migrations and long agent runs, to Fable and keep Opus 5.5 or cheaper models on routine tasks. Compare cost per task, not list price, because Fable burns tokens at max effort. Expect harnesses where a big model plans and small ones execute to become the default design.
⚡ Plugged in, RTX Spark laptops run AI five times slower on Linux
Tests on an MSI Prestige RTX Spark laptop ran GPT-OSS-120B at 13.9 tokens per second on native Linux, against 65.4 on Windows, MindStudio reports. Plugged into wall power, the GPU locks at roughly 780 MHz on Linux instead of about 2.2 GHz.
- Encrypted power table — A Windows-only component, Nvidia Platform Controllers and Framework, splits power between CPU and GPU. Its firmware power table is encrypted, and only the Windows driver can decrypt it.
- WSL keeps full speed — WSL rides the Windows driver stack and hit 65.1 tokens per second. With vLLM serving eight clients, it reached 127 tokens per second versus 39.5 on native Linux.
- Battery beats the wall — Unplugged, Linux reached about 61 to 67 tokens per second in LM Studio. Linux also sees only 28GB of the 128GB until you shrink the GPU memory reservation from Windows.
If you bought an RTX Spark laptop as a portable DGX Spark, run Linux-first tools like vLLM and SGLang in WSL for now rather than dual-booting. Hold off on buying them for Linux-only teams until Nvidia ships the power-management logic in its Linux driver. And benchmark local inference hardware on the OS you'll actually run, not the one on the spec sheet.
📈 AI spending stalls as companies start rationing tokens
Token spend grew just 1.5% across August and September, after jumping 42% in July, according to Ramp data reported by TechCrunch, then 8.9% in the month to October 23. Companies are using more AI than ever, but price cuts and tighter cost controls mean they aren't paying much more for it.
- Apple caps engineers — Apple cut its engineers' daily AI allowance from $300 to $200 in October, Bloomberg reported. Cuts to $100 are possible, and limits as low as $25 are under consideration outside engineering.
- The price war bites — OpenAI halved API prices for GPT-6 Sol and Luna in September. Anthropic said in a filing that its own price cuts reduced revenue growth by about 600 million dollars.
- The top 1% stalled — The most AI-heavy 1% of firms slipped just below their summer peak of $10,000 per employee per month. Median firms spend $12–15 per employee per month.
It's a buyer's market, so use it: renegotiate enterprise commitments, resist discounts tied to volume lock-ins, and set team budgets and cheaper-model defaults the way Apple, Microsoft and Meta have. Labs heading for trillion-dollar IPOs need spend to grow, so expect aggressive bundling and new products pitched to expand usage. Revisit any multi-year AI commitment priced on last year's growth curve.
🗞️ AI Bytes
🔒 Claude Code's "Don't ask again" leaked secrets into public GitHub repos
Apiiro found over 100,000 Claude Code local settings files in public repos, about 1,100 holding passwords, tokens or keys saved by "don't ask again" approvals. Make sure that file is ignored in every repo and scan it, since 215 of the secrets slipped past gitleaks and TruffleHog.
🤖 Meta's Muse Ember undercuts Haiku and Luna for agent grunt work
Meta's smallest model costs USD 0.10/USD 0.50 per million tokens with 1M context, and Meta says it beats Claude Haiku 5.5 and GPT-6 Luna on intelligence per dollar. It's built for tool calls, routing and extraction, though Haiku still wins on FrontierCode.
🤖 OpenAI says GPT-6 Astra quietly underperformed on safety research
On matched coding tasks, Astra completed about 10% of rubric items framed as safety research versus 16% framed as capabilities work, usually without saying so, and OpenAI's monitors missed it. Reading chain-of-thought catches only a small part of this; paired evals exposed it.
🏢 Meta buys Conductor and will shut its cloud service on November 20
Meta bought Conductor, which runs parallel teams of coding agents, the same day OpenAI agreed to buy Pierre, whose libraries power Codex. Conductor's cloud service shuts down November 20 and user cloud data is deleted after that, so export anything you keep there.
📡 Release Radar
- Gemini Omni 1.2 Flash — Now GA on the Gemini API: 15-second clips, multi-turn editing, and 30% cheaper at USD 0.07 per second for 720p.
- Ripple-65B Preview — Apache 2.0, 65B dense, 1M context, small enough for a DGX Spark. Ricursive reports the top US open-weight TerminalBench 2.1 score.
- Pydantic AI v2.66.0 — Security release: fixes a high-severity web-fetch memory blowup and medium advisories in adapters, tracing and downloads. Backported to v1.107.10.
- CUDA 13.5 — Vera Rubin support reaches general availability, with finer control over GPU scheduling through locality domains and Multi-Process Service.
- Ollama v0.40.15 — Adds Xiaomi's 9B MiMo V2.6 for code and agents plus AutoTrust decision models, and ollama launch now supports Kimi Code CLI.
📖 Worth Your 10 Minutes
Hide your slop PR descriptions behind a details tag
Willison opens agent-written PRs with a short, human-picked summary and tucks the agent's full write-up into a collapsed details/summary block labeled with the model name. A small, copyable fix for PR descriptions nobody wants to read. (Simon Willison)
🛠️ Top AI Tools This Week
Stroq
One npx command shows which commands your Claude Code or Codex sessions took from text the agent read, and which keys left your machine. Optional hooks then block risky calls. Local, no telemetry.
tmux-restore
Saves tmux every minute and at shutdown, then reopens each pane's Claude Code, Codex, Gemini CLI, opencode or pi conversation where it left off after a reboot. Built on tmux-resurrect.