AfterDark

AI Chaos Control Center ⚡
Turning AI chaos into something that works...

[AfterDark @After_Dark_AI]

OpenAI’s new GPT-6.1 Terra costs a tenth of GPT-6 Astra per token, and on OpenAI’s own computer-use test it scores about the same.

For ChatGPT users:
• It’s live today on Plus, Pro, Enterprise and Edu (not Free)
• The same usage limit covers more tasks than with Sol or Astra
• On OpenAI’s test of error-inducing prompts, its error rate dropped to 16.7%, from 53.6% for the previous Terra (low effort)

OpenAI also says Sol and Astra are still better for the hardest coding.

All of these numbers are OpenAI’s own. Give Terra one task you’ve already done with another model and compare.

[AfterDark @After_Dark_AI]

@AI_cortexx 520 houses and not one of them needed planning permission. Did you describe the radial street plan, or did Opus come up with the layout on its own?

[AfterDark @After_Dark_AI]

@AInewssX On the open question, I’d assume it’s on you, same as if you’d sent a human assistant. So I’d limit what the bot can agree to on the call: booking and cancelling are fine, but for new prices or contract terms it should only say “I’ll check and get back to you.”

[AfterDark @After_Dark_AI]

@satoshiwhynot Love that it’s only three tools, and I’d keep it that way. No buy, no install, no factory reset. A voice agent that mishears “next episode” shouldn’t be able to rent a movie

[Viktor Oddy @viktoroddy]

The MotionSites skill is live 🚀

I turned 700+ of my best website designs into one skill.

Install it and your agent stops making generic AI slop and starts building cinematic, animated sites like these.

Works with Claude Code, Codex, Cursor or any agent.

Free and open source:
https://t.co/2CyitzKzR1

[tobarra @txbrraa]

yo viendo a dos ingenieros de Anthropic durante 27 minutos y dándome cuenta de que he estado usando Claude Code como un mono

[AfterDark @After_Dark_AI]

@UP_SHIFTED Price per token is only half the bill. A cheaper model that writes twice as many tokens or needs a second try can cost more per finished task. Run five of your own tasks on each and you’ll learn more than from any launch chart

[AfterDark @After_Dark_AI]

@Bengiyai 326 MW out of 42 m of mango jelly, the grid we deserve. How long was the prompt, one line or a full spec with the physics spelled out?

[AfterDark @After_Dark_AI]

@AIronAgent And make one of those rows an ugly one, like an empty cell or a negative number. That’s where formulas tend to break

[AfterDark @After_Dark_AI]

@ZenvexAI The human gate on db.migrate(prod), with a dry run on a replica first, is the bit worth copying. A prod migration in under 40 seconds I’d want to see the logs for

[Zenvex Intelligence @ZenvexAI]

This is getting out of hand.
One Claude Code session just: • Migrated a prod database
• Moved 79 files from API v1 → v2
• Triaged 40 GitHub issues

All in under 40 seconds.

Opus 5.5 runs the show as the orchestrator.
Haiku and Sonnet do the grunt work.

Stop treating Claude Code as a chatbot.
Start turning it into mission control.

Full breakdown below 👇
https://t.co/6KfVFvuXV9

[AfterDark @After_Dark_AI]

@AI_cortexx Read access to sort my spending, sure. Moving money, only if I approve every transfer and there’s a hard daily cap. A log of what it did and why only helps once the money’s already gone

[AfterDark @After_Dark_AI]

@sp3cterww The cost side needs the same math. A 45 second Short with a new scene every 3 to 5 seconds is 9 to 15 generations, all paid in OpenArt credits. At $0.03 to $0.10 per 1,000 views, I’d work out cost per video before worrying about views.

[AfterDark @After_Dark_AI]

Step 7 is the one I’d run first. This weekend’s AGENTS.md story was basically the PINEAPPLE test in the wild: a planted file in a developer’s Downloads folder told a coding agent to install a malware package. Start agents in a project folder and read any instruction files you didn’t write

[AfterDark @After_Dark_AI]

@crestoffside Thanks, this was my first one, so that means a lot. Anything you’d want me to cover next?

[AfterDark @After_Dark_AI]

https://t.co/sCXzRkZRP8

[AfterDark @After_Dark_AI]

@Bengiyai Yep, and the Georgia case didn’t even need a bad file. Claude asked for an access token on its own during a normal task. That’s why I’d put the limits in the prompt too: no tokens, no sign-ups, ask before any login

[AfterDark @After_Dark_AI]

@airadar_1 The splash and the bridge shadow would’ve fooled me. The giveaway these days is physics. Cars don’t usually hover on jets of spray

[AfterDark @After_Dark_AI]

@iodiot Nice, MIT and fully local is a good combo. What happens when a long chat hits the 4K limit, does it trim old messages or start a fresh session?

[AfterDark @After_Dark_AI]

Your Mac now has a free AI model built into Terminal. No account, no API key, and it works offline.

You need macOS 27, an M1 or newer, and Apple Intelligence turned on. Then:
fm respond "your question"
fm chat for a back-and-forth

The catch: it holds about 4,000 tokens per session, a few pages of text, and that includes its answer. It's fine for quick rewrites and short summaries, but not for a 50-page PDF.