Whether it’s a quick question or a detailed brief — we’d love to hear about it.
AI Hub 036 · Newsletter
Everything Is Getting Reverse Engineered. Oh, And Fable 5.2 And Astra 6.2 Released.
The most useful thing people found for AI to do these past few weeks was to point it at a piece of software and have it cloned.
Overview
The practical takeaway upfront is that agents can now crack open software and rebuild it from the outside, so use that only where we hold the rights and assume anything we ship can be read. Fable 5.2 ties Opus 5.5 on coding at over three times the cost, so Opus stays the default. GPT-6.2 Astra ranks second on general boards, did not improve on the coding boards and drew a mixed first reaction, so nothing changes for our seats. Gemini 4 Carbon is level with Claude on coding until its price doubles on December 1. Watch how visual work draws on the weekly limit.
SOFTWARE IS GETTING CLONED AND TAKEN APART
The loudest rebuilds came from the ArtCraft project, which between September 30 and October 9, 2026 published Rust equivalents of Photoshop, Illustrator, Premiere, Lightroom, Acrobat and After Effects, then Word, Excel and PowerPoint, which Tom’s Hardware says Claude built. By their own account they are clean-room builds from public specifications. The author rates the Word equivalent at about 60% ready for real work, and I found no response from Adobe or Microsoft.
People also pointed Claude Code at smart TVs. A post quoted by TechCrunch on October 19 described giving Claude Code the TV’s model number, after which it found the set on Wi-Fi, removed 25 Xiaomi packages and wrote a new launcher. The developer behind the cleanyourtv tool says a few thousand people did the same within two days. A November 10 video and its repository show Claude Code and Codex reading a 1994 Jeep Cherokee’s engine computer through a $3 Arduino, building on an existing open-source scanner.
Tom’s Hardware reported on October 12 that Maurice Heumann used Claude and Codex agents, mostly Sonnet 5, to decompile Call of Duty: Modern Warfare 2 over three months, with 83% of functions matching the original byte for byte, and that his video was disabled after a report by the copyright owner. On November 2 a developer ported the Battlefield 1942 demo to the browser with Opus 5.5.
FABLE 5.2 IS THE MODEL WE ESCALATE TO
Anthropic released Fable 5.2 on November 9, 2026 at $8 and $40 per million input and output tokens. On the independent boards it is the top model overall, ahead of GPT-6.2 Astra, Gemini 4 Carbon and Opus 5.5, and it leads on the standard ARC-AGI-3 test as well.
On coding it only ties Opus 5.5. Terminal-Bench 4.1 and the Artificial Analysis Coding Agent Index both have the two level, and the Terminal-Bench run cost more than three times as much. Anthropic recommends users “start with Claude Opus 5.5, our strongest daily driver.” Zvi Mowshowitz wrote on November 16 that for most purposes Fable 5.2 “will not be appreciably better than Claude Opus 5.5,” though he would reach for it on a difficult, very complex or long task. In an earlier issue I treated Fable as the reference point, and it now sits one step above the default.
On plans that include Fable, Anthropic says it is included up to 50% of the weekly limit and draws that limit faster than other Claude models. A launch reset restores spent Fable weekly usage and, where the account has one, refills the five-hour window, but it leaves other usage untouched, raises no limit, and expires March 8, 2027. Anthropic’s Advisor and Orchestrator roles, where smaller models consult Fable or Fable delegates to them, stretch the allowance.
GPT-6.2 ASTRA GETS STRONG SCORES AND A MIXED WELCOME
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026. Astra costs $8 and $40 per million input and output tokens, 20% below GPT-6 Astra and the same as Fable 5.2, while Sol is unchanged at $2 and $10. Both are available on Plus, Pro, Business, Enterprise and Edu plans in ChatGPT Work and Codex.
On the independent boards Astra ranks second overall to Fable 5.2 at a fraction of the cost per task, and it is the first model to score 100% on ARC-AGI-3, though only on a harness where the other frontier models score nearly as high. On the coding boards it has barely moved from GPT-6 Astra and still sits below both Opus 5.5 and Fable 5.2.
The first-day reaction from the community has been mixed. People like the price, the review quality and what it does with 3D and game work, but they complain that its output and summaries are hard to read and that long tasks stall, and many of them measure it against Opus 5.5. The release is one day old, so this may change.
Honestly, this reads to me like release fatigue. It is the third release from a major lab in nine days, after Fable 5.2 on November 9 and Gemini 4 Hydrogen on November 12, and I suspect people are running out of excitement for yet another model at the top of a chart. The launch still got plenty of discussion, so take that as my impression.
GEMINI 4 IS CROWDING THE TOP
Google brought three Gemini 4 models to general availability between October 13 and November 12, 2026: Argon, Carbon and Hydrogen. Carbon is the one near the top of most boards. It sits third on the Artificial Analysis index, behind Fable 5.2 and GPT-6.2 Astra, level with Opus 5.5 and Fable 5.2 on Terminal-Bench 4.1, and narrowly ahead of Argon on a preliminary Text Arena snapshot. Argon and Hydrogen rank lower on the coding boards.
Reception is also mixed. Several developers in the community say they use Argon for much of their work and like it. The recurring complaints, mostly about Argon and Carbon, are slow responses, tight plan limits and gated access. Hydrogen reached general availability on November 12 and is too new to judge.
In an earlier issue I said the introductory price rises on November 30. It rises on December 1, when Argon and Carbon go from $2 and $10 per million tokens to $4 and $20. I also called Argon the challenger to test first, and I would now test Carbon.
AGENTS ARE STARTING TO DRAW AND CLICK
A run of new repositories has agents drawing. Anatomy and Diorama have an agent build isometric illustrations and Three.js dioramas, Codixel turns reference images into pixel art through code, and AG-UI lets an agent render interface inside an application. I found more than a dozen other visual skills created on GitHub after October 1. The limits still show, since Codixel’s README says Claude Code “cannot draw pixel art from scratch,” and the most visible pixel art, Isometric SF, was image-model output that the community called “not actually pixel art.”
Repositories worth a look
- Anatomy
Interactive isometric SVG illustrations, drawn by the agent - Codixel
Reference images turned into pixel art through code - Diorama
Playbook and code kit for detailed Three.js dioramas - AG-UI
Open protocol that connects agents to application front ends
The vendors already ship the rails. Anthropic has offered computer use in Claude Code and Cowork since a March research preview, and Gemini API computer use became generally available on October 30.
The launch material points the same way. Anthropic lists vision among Fable 5.2’s gains, with chart reading well ahead of Opus 5.5 and computer use only slightly ahead. My read is that labs build where the demand shows, and what people post right now is visual, so I suspect the next round of gains comes there. No lab has said this, so treat it as my guess. I also suspect visual work will weigh on the weekly limit, since computer use works from screenshots, but I have not seen anyone measure it.
WHAT THIS MEANS FOR US
For us, reverse engineering cuts two ways. Legacy modernisation and integration with undocumented client systems both start with reading code we did not write. The developer behind the Modern Warfare 2 project estimates that three months of it cost him about €1,200 on subscriptions, where API prices would have run past $200,000, so this work fits seats like ours once licence and client permission are settled. The same holds for anyone looking at our work. Anything we ship as a client binary, app or firmware is cheaper for others to read, so secrets and trust checks belong on the server, and a client whose product is a closed desktop app should expect clean-room rivals.
Last newsletter I said model capability was levelling off, and for our work that has held. Fable 5.2 is the better model, but on the coding we do it only ties Opus 5.5, and GPT-6.2 Astra did not move the coding boards either. So far Opus 5.5 has solved almost everything we have given it, and we have not yet hit a problem that needed Fable 5.2, so for now Fable feels more like a novelty than a practical switch. Piling on tooling does not help either. When Surge AI tested Opus 5.5 earlier this month, plain Claude Code at a higher effort setting improved the score, while a setup loaded with dozens of skills and agents cost eight times as much and scored lower.
-
Keep Opus 5.5 as the default, and move up to Fable 5.2 only when Opus 5.5 cannot solve the problem.
-
Spend a Fable reset on a long, hard job, since it restores Fable usage and the five-hour window.
-
To try Carbon, do it before December 1 at the introductory price, through OpenRouter against internal repos only, or ask the AI Excellence team.
-
Use reverse engineering only where needed and necessary, and where outputs can be verified appropriately because code can lie, and expect refusals for any attempt that mimics known intellectual property because Anthropic has steered Claude Code away from recreating other people’s software.
Sources → ArtCraft apps · WordCraft · Tom’s Hardware: Adobe rebuild · Tom’s Hardware: Office rebuild · TechCrunch · xj-scanner (Jeep) · Tom’s Hardware: Modern Warfare 2 · levels.io · Anthropic Fable 5.2 · Artificial Analysis models · Artificial Analysis coding agents · Terminal-Bench · ARC Prize · Zvi Mowshowitz · Claude Fable on your plan · Fable 5.2 launch offers · OpenAI news feed · GPT-6.2 system card · Codex pricing · GPT-6.2 Astra API docs · Artificial Analysis on GPT-6.2 · ARC Prize GPT-6.2 results · Hacker News GPT-6.2 thread · Gemini changelog · Gemini pricing · Google: Gemini 4 Carbon · Text Arena (Internet Archive) · Hacker News: agentic setups · Hacker News: Gemini 4 Carbon · Hacker News: Gemini 4 Hydrogen · OpenRouter · Anatomy · Codixel · Diorama · AG-UI · Isometric SF · Hacker News: Isometric SF · Claude release notes · prod.codes · Tom’s Hardware: limits · Tom’s Hardware: loosened · Surge AI