The Coding Bird

AI coding tools, tested for real. Reacting to whatever broke on the internet today. Occasionally touching grass.

[The Coding Bird @theCodingBird]

@mitsuhiko I ran /compact on myself on Tuesday. Still don't know what I was working on.

[The Coding Bird @theCodingBird]

@jh3yy Yeah, my draft folder is bigger than my post history 😢

[The Coding Bird @theCodingBird]

@Hesamation My agents never check the stock price before they run. Bubble talk moves portfolios, the tools I use every day just keep shipping.

[The Coding Bird @theCodingBird]

$220 of DeepSeek for $10 a month is a great headline.

That $220 only exists on Eco, the slower model. On regular V4.1 Flash the same plan shows $120. On GPT-6 Luna, $15. My read: it’s really one $60 pool, and the number moves with each model’s multiplier.

And you can’t spend it when you want. Go caps you at $55 a week and $18 every 5 hours. Hit the cap and you wait.

Speed is the other catch. Their docs say Eco “may respond more slowly at times” and point you to batch work. One reply here reports 40 seconds to first token. Only one person, but I wouldn’t want that mid-session.

Should still be a steal for qualifying low-priority background/long runs...if reliable. Just don’t budget your workday around $220.

[OpenDesign @OpenDesignHQ]

DeepSeek. CheapSeek. Now live. 🐳

Same weights as DeepSeek V4.1 Flash. Same agent, coding, reasoning & multimodal capabilities. Full 1M context.

$220 of monthly usage on OpenDesign Go. $8 for your first month.

Meet DeepSeek V4.1 Flash Eco ↓

[The Coding Bird @theCodingBird]

https://t.co/G8sEMwx5C1

[METR @METR_Evals]

Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development.

We are now publishing a report summarizing the results of this risk assessment.

[The Coding Bird @theCodingBird]

METR spent months with Anthropic's internal data on one question: is AI speeding up AI development?

Their read:

- Capability growth on well-specified tasks: likely about 2x the Mar 2024–Oct 2025 pace (range 1.3x–3x)
- Most of that speedup is likely now from AI used in AI R&D, but Anthropic staff don't all agree
- Not from faster growth in training compute
- No clear bottlenecks in the next 12 months, though they may have missed some

Anthropic redacted most of the evidence METR found most informative, so the public has to lean on METR's judgment.

[The Coding Bird @theCodingBird]

@PineapplesDev Yes, and exits just before the IPO are always interesting to look at.

[The Coding Bird @theCodingBird]

OpenAI's Wednesday:

- Pierre Computer Company joins, reportedly to build Git workflows for Codex
- Images 2.6 reaches ChatGPT, Codex and the API
- Cat Wu, who worked on Claude Code for two years, joins to lead product management for Core Products

The hire got more views than the launch.

[The Coding Bird @theCodingBird]

@escship And OpenAI needs both right now.

[The Coding Bird @theCodingBird]

@stolinski Stacked rack you got there Scott, respectfully ofcourse 🤓

[The Coding Bird @theCodingBird]

https://t.co/TcRctAxHIp

[cat @_catwu]

Some news: I’m joining OpenAI to lead product management for Core Products!

I have a lot of respect for OpenAI for showing the world that scaling works, for bringing AI to over a billion people, and for believing great products are core to the mission.

I’m excited to work with @thsottiaux, an incredible PM team, and everyone across OpenAI. Together, we’ll build agents that have your back and your team’s, and keep you in control, so you get more time for the people and work you love.

Lots to learn, and lots to build. Please keep your feedback coming!

[The Coding Bird @theCodingBird]

@BakurovYurii @ClaudeDevs Yeah, 99% was almost not making sense, had to dig further.

[The Coding Bird @theCodingBird]

The 99% in the Claude Code subagent fix is the cost of one status update.

The number for your limit is in the next post from @ClaudeDevs: about 5% less in sessions with background subagents, or about 6% more prompts.

Still worth updating. Just don't expect 99.

[The Coding Bird @theCodingBird]

@ClaudeDevs https://t.co/NVYL11oS3N

[ClaudeDevs @ClaudeDevs]

Background subagents in Claude Code now use less of your usage limit.

A running subagent updates its status line every 30 seconds, and each update used to re-read its whole conversation. Now Haiku 5.5 writes it from short notes instead, so each update costs about 99% less.

[The Coding Bird @theCodingBird]

Every refactor after the tests already pass. https://t.co/yEBt6qpICB

[Artificial Analysis @ArtificialAnlys]

Knowing when to stop is hard for language models controlling a robot arm

➤ Claude Opus 5.5 completed the task in 23 minutes, and said so: "The crocodile looks fully inside now"

➤ It then spent 22 minutes pushing the crocodile back and forth, only to finish almost exactly where it was before

[The Coding Bird @theCodingBird]

Your coding agent's behavior now lives in a few config files, and any app you run can edit them.

That's the real story in the Postman issue. One user's logs show its desktop app adding plugins to Claude Code and Codex. The code also tries to approve its own Codex hook.

Who else is in those files?

[The Coding Bird @theCodingBird]

Most alarming line in the Claude Code v2.1.318 notes:

"Fixed idle compaction replacing a conversation with a reply that was not a summary"

Also fixed: workflows re-running finished agents when resumed after a usage limit.

Good week to be a long-running session.

[The Coding Bird @theCodingBird]

https://t.co/FWYACCU8DH

[The Coding Bird @theCodingBird]

Claude Code v2.1.317 is mostly Claude asking before it does things you didn't ask for.

- Asks before a web or shell tool call sends your account email to the network
- /schedule warns when a new cloud task picks up connectors you didn't request
- Agent SDK hooks get onFailure: "block", so a crashing hook stops the action instead of waving it through

Quiet release. Good defaults.

[The Coding Bird @theCodingBird]

@trq212 The day a model is "just right" is the day I stop reading every diff before I merge.

Would love that day. Still reading every diff.

[The Coding Bird @theCodingBird]

OpenAI built a new Git backend because agents push a LOT.

Pando: ~80% lower push latency, nearly a petabyte of Git data served a day. The stated goal: grow Git capacity alongside agent usage.

Yesterday Shopify went monorepo, also citing agents.

Soon your repo's busiest contributor won't be human.