Claude Fable 5.2 and Mythos 5.2 launch; Fable builds a full game
Anthropic releases Claude Fable 5.2 and Mythos 5.2, and a developer reports a full beat 'em up built overnight. Also: Meta's Muse Ember, Claude Code secrets leaks, and a CPython formalization in Lean.
Top stories
Claude Fable 5.2 builds a 12-level beat 'em up from a child's drawings
- Developer Lex Toumbourou says Claude Fable 5.2 finished his nephew Aiden's beat 'em up game overnight from one prompt, rewriting most of his earlier code.
- The run used about 600 Claude requests, roughly 175 images and 140 audio clips, totalling around 340M tokens, and the agent tested all twelve stages with a bot in a headless browser.
- The result has 12 levels, four-player controller support and 18 characters after additions; he calls it a pretty wild milestone and says Claude Code with Fable 5.2 can make such games.
Anthropic releases Claude Fable 5.2 and Mythos 5.2 at $8 and $40 per million tokens
- Anthropic announced Claude Fable 5.2 and Mythos 5.2, the same underlying model, with Fable carrying cyber and biology safeguards and Mythos offered through verification programs.
- Fable 5.2 scores 49.1% on FrontierCode 2.0 at Xhigh effort and 69.9% on FrontierSWE v2, ahead of Opus 5.5's 45.5% and 62.3%.
- Pricing is $8 input and $40 output per million tokens, with a fast mode at $16 and $80; Anthropic says over 90% of merged changes in its core codebase are written by Claude.
Shane Melaugh says Opus 5.5 reasoning changed sharply on November 3
- Shane Melaugh asked on X whether Claude Opus 5.5 has been nerfed, after an analysis he ran of Claude Code sessions.
- He says Gemini 4 Argon analyzed 1,400 Claude Code sessions and found that on November 3 the model's reasoning took a sharp turn, though the model name did not change.
- The post offers no further detail or method beyond that claim.
In brief
- Shahar Glazner argues engineers still have leverage over AI agents. Shahar Glazner, co-founder and CTO of keepalerting, argues in a post titled "On Being an Engineer in 2026" that engineers still have leverage over AI agents, using a proof by contradiction: output gaps between great and average engineers persist despite shared models. He attributes that leverage to tacit knowledge, intuition, and lived experience that agents lack, while acknowledging future… (twitter.com)
- Simon Willison has Claude Code build Beep Locator to find a beeping device. Simon Willison describes Beep Locator, a system he had Claude Code build to triangulate an intermittent beep in a room using multiple listening devices such as iPads, a phone, and laptops. Based on the BeepBeep 2007 acoustic ranging paper, it locates sounds to within about 40cm, which led him to an old carbon monoxide detector needing new batteries. He also notes that Claude Code was given live… (simonwillison.net)
- CIQ report: Ginux kernel built in eleven days by a coding agent. Gregory Kurtzer of CIQ published a Zenodo report on Ginux, an independent Rust-first x86-64 kernel built over eleven days through Human-Directed Agentic Engineering (HDAE), where humans set architecture, constraints, and acceptance criteria while an autonomous coding agent implements and revises. The report says Ginux boots via UEFI, runs selected unmodified Alpine Linux programs, and boots a… (zenodo.org)
- Bruno Rocha says acting as a meat proxy for AI should be punishable. Bruno Rocha's blog post argues that developers who act as "meat proxies" for AI, prompting an agent and pasting back unreviewed output without validation or disclosure, should face punishment, possibly including firing. He cites wasted colleague time, harm to products, and AI-generated PR floods, and rejects fighting AI with AI as a remedy. (swiftrocks.com)
- Apiiro finds secrets in Claude Code settings files on public GitHub. Ars Technica reports that Claude Code's local .claude/settings.local.json files, which store commands approved with the "Yes, and don't ask again" option, have leaked thousands of secrets on public GitHub repositories. Apiiro found over 100,000 uploaded settings files, about 1,100 of which from 63 organizations contained secrets, 215 of them missed by gitleaks and TruffleHog. Claude Code's team… (arstechnica.com)
- Meta announces Muse Ember, a small model at $0.10 per million input tokens. Meta announces Muse Ember, its smallest and most efficient model, aimed at agent tasks like tool calls, coding subtasks, routing, and extraction. The post claims it beats GPT-6 Luna at every reasoning setting, matches Claude Haiku 5.5's best score at about a third of the cost, and costs $0.10 per million input and $0.50 per million output tokens, with a 1M-token context window. (research.meta.ai)
- Agents write 613,318 lines of Lean modeling CPython and surface 418 bug reports. The post describes a project where AI agents wrote 613,318 lines of Lean that copy CPython's internals line by line, test them against real Python, and prove properties checked by Lean's proof checker. So far 4,071 of 26,004 functions (15.7%) have passed independent review, and the work surfaced 418 bug reports, about 55 of which can crash the interpreter or corrupt memory from ordinary Python… (broyojo.com)
Chosen by software from the day's news, written by AI from the pages it links to, and checked against them by a second model. Published by Rapid Review, Inc.