AI Coding Weekly

Fable 5.2 tops the arena and still burns your limit

Fable 5.2 took #1 in Agent Arena at about 1.6x the next model, but Theo's log of 1.1 million requests says a $200 plan buys roughly $15,000 of Opus and only $3,000 of Fable. Benchmark rank and what you can afford to run all week are now two different questions.

The big disagreement

"You don't need to learn how to write this" set off the week's fight

Peter Steinberger quoted a Vercel post with a line about not needing to learn to write code because the agents will, and the replies turned into the week's loudest argument about craft. Mario Zechner, pointing at the people who reverse engineered Apple Silicon GPUs pre-AI, said deep knowledge is still relevant in 2026 and that the most important skill is telling your agent that its output is crap.

Mario Zechner
@badlogicgames
X
the most import skill in 2026 is still being able to tell your agents "wtf is this crap, no, not like this
Nov 9, 2026 · View on X

Read the full story

Receipts

Fable 5.2 wins the arena, loses on what a $200 plan buys

Arena put Claude Fable 5.2 (Max) at #1 with +30.4% net improvement across 4.7k agentic sessions, about 1.6x the #2 Opus 5.5, and Theo called it the biggest lead any lab has had since GPT-3. Then he analyzed over 1.1 million logged requests on his $200 Ultra plan and found Fable burns faster per dollar than Opus, on top of being capped at 50% of the weekly limit.

Theo - t3.gg
@theo
X
The $200 Claude Ultra plan gets you $15,000 of Opus 5.5 in a month, but only around $3,000 of Fable 5.5.
Nov 10, 2026 · View on X

Read the full story

dax rewrote OpenCode Browse in Rust for $60 in tokens

dax says he rewrote OpenCode Browse in Rust, including reimplementing the parts of Playwright and Patchright his team needed, and that it only cost $60. Separately his team is porting OpenCode from Electron to Rune, which he says launches instantly with no cold start feel, and he notes a big chunk of Rune was built with Gemini Carbon.

Read the full story

How they actually work

Pocock published the Home Repo pattern with a setup prompt

Matt Pocock turned his answer to running agents across multiple repos into a written-up pattern with a copy-pasteable prompt you can run on your org's repos. One repo holds all skills, steering files, wiki and tooling, you launch every agent there, and AGENTS.md carries pointers like "clone mattpocock/sandcastle to ~/repos/sandcastle". He says he had used it for months and thought it was too boring and obvious to post.

Read the full story

Model and agent watch

Theo traces Gemini's slowness to 11 seconds before first token

After calling Gemini 4 Carbon unacceptably slow, Theo dug in and said almost all of it is server side: 11 seconds before it starts responding, 20x longer than Fable. His timings were 22m21s for Carbon on low against 9m33s for Opus 5.5 on high, and a T3 Code Windows installer fix that took fourteen hours.

Theo - t3.gg
@theo
X
Almost all of it is server side. Their inference is awful.
Nov 11, 2026 · View on X

Read the full story

Open tabs

Frequently asked questions

Do I need to learn to code if AI agents can write it for me?

No, but you need to learn to judge whether your agent's output is good or bad. The debate this week was not about whether to use agents, but about what skills remain essential to have.

Which AI model is fastest for coding tasks?

Claude Fable 5.2 ranks highest in arena testing for agentic sessions (multi-step agent work), but Gemini 4 Carbon has a significant latency problem with 11 seconds of delay before it starts responding, compared to much faster times for other models.

What is the Home Repo pattern?

It is a setup where one repository holds all agent skills, documentation, and tooling, and every agent launches from there. This creates a single place to make changes that affect all your repositories at once.

Why was Cursor's data leak a security problem?

Static credentials (long-lived access keys) were sitting inside stolen LLM transcripts, which made them exposed to attackers who obtained those transcripts.

Is it now cheaper to rewrite code than maintain dependencies?

AI-assisted rewrites have become fast and cheap enough that what used to require weeks of engineer time now costs a small amount of money, which changes decisions about what dependencies are worth keeping.

Get the next one by email

Coding with AI, read daily so you do not have to. The experiments, the receipts and the arguments from engineers shipping real software. Not a changelog.