Open models pick the right tool, send bad arguments, and still say "done."
Anvil fixes the tool interface and tests it: 20 → 43/45 on the same model.
Built in @tryAlanAI with open models, now an MCP connector in Alan.
#OpenFire@prasad_pilla
Today we're launching @tryAlanAI so developers can stop babysitting their coding agents.
We kept running into the same problem ourselves.
Even as agents got better at writing code, we were becoming glorified button-pushers for our own AI: starting every run, feeding them context, unblocking failures, and checking their work.
We built Alan to change that.
Astra didn't beat Fable on the leaderboard. It beat it on the bill.
Same 62 on @ArtificialAnlys:
Fable 5.2 (high): $9.08/task
GPT-6.2 Astra (xhigh): $2.52/task
Same $8/$40 per token, way fewer tokens burned.
3.6x cheaper. Who's switching?
https://t.co/3DQTjZgonD
Some of you have been asking very politely. Some of you have just been asking.
GPT-6.2 Astra and Sol are now available.
Super efficient and built to perform where it counts—the real world.
https://t.co/qr2iP0GpfX
Winners: @mabelle_escreve and @ixxrllg 🎉 Picked by public randomness (drand round 6650306), anyone can check at https://t.co/xQxnGYcaUk. If a winner doesn't reply to my DM within 24h, the pass goes to @22022622A80853.
https://t.co/9ACFcVaLHi
Giving away my last 2 Claude Pro guest passes 👀
Each one is a free week of Pro, Claude Code included. It's Anthropic's own referral pass (if you subscribe later I get credit, full disclosure).
How to enter:
- Follow me
- Like and reply "Claude"
2 random winners in 24h, DM'd.
Giving away my last 2 Claude Pro guest passes 👀
Each one is a free week of Pro, Claude Code included. It's Anthropic's own referral pass (if you subscribe later I get credit, full disclosure).
How to enter:
- Follow me
- Like and reply "Claude"
2 random winners in 24h, DM'd.
spent way too long tonight benchmarking something i was sure was slow
turns out half of it was slow and the other half was fine and i was just measuring the wrong thing
always bench the boring worst case, not just the case that makes your change look good
Vibe coding isn't a scam and it isn't the end of engineering.
It's fine right up until the thing gets a pager. Then the only question that matters is whether anyone on call can read what it does.
@mattpocockuk Nice setup. One thing I'd worry about: the pointers say clone, not which commit. A sweeping change across repos then runs against whatever main happened to be that day. Do you pin, or have the agent check each repo is clean first?
WorkflowArena's split says a lot: formal math 95.83% resolved, SWE 29.41%, law 9.09%.
Math comes with a checker. Code mostly comes with acceptance criteria someone half-wrote.
Agents aren't bad at software. We're bad at writing down what "done" means.
@rauchg The agent can write the purpose-built language fine. The catch is the diff. Every new language shrinks the set of people who can review it, so I think the bottleneck moves to review and verification tooling, not codegen.
Everyone's on Fable 5.2 in Claude Code 2.1.321 @claudeai. The lines I'd update for are lower down:
- MCP OAuth could send refresh tokens to another host on a redirect
- a read-only Bash command could run outside the sandbox without a prompt
Worth bumping even on Opus.
Claude Fable 5.2 is out. What matters if you build with it:
- $8/$40 per M tokens, down from $10/$50
- 49.1% on FrontierCode 2.0 vs 45.5% for Opus 5.5
- Medium is now the default effort in Claude Code
- Gemini 4 Carbon still leads AutomationBench
Not a clean sweep.
@claudeai Easy to miss: Medium is now the default effort for Fable 5.2 in Claude Code. If you compare it to 5.1 or Opus 5.5 on your own tasks, check which effort level each run used first, or the comparison means nothing.
@penberg@tursodatabase Curious how this works with connection pooling. Behind a transaction-mode pooler, RLS policies reading a setting like app.user_id need SET LOCAL, or it can leak into another client's transaction. Does Turso tie identity to the connection or each transaction?
The todo reminder bug is the one that stuck with me. A trigger keyed on top-level tool names works fine until the calls move inside a program, and then it just goes quiet with no error. Did you end up hooking the inner calls, or move the reminder onto something like progress state?
@garrytan The hard part with apps like TV or Calendar probably isn't writing the fix. It's fifteen years of sync edge cases, plus the people who quietly rely on the weird behavior. Agents make the rewrite cheap, but they don't tell you which bug someone depends on.
@rough__sea Writing got way faster. Reading didn't.
So the job quietly turned into deciding which parts you actually read line by line and which parts you let a test vouch for. Nobody really trained us for that part.
“Not checked” shouldn’t mean “safe”.
In AegisCode, if npm or PyPI can’t be reached, the package check reports “not checked” instead of quietly passing.
Small distinction, but important when an agent is deciding whether to add a dependency. Missing evidence isn’t a clean bill of health.
https://t.co/zPBuerwKgR
Small detail I liked in Render’s OIDC write-up: each workspace gets its own issuer URL.
Even if an AWS trust policy forgets to restrict the subject, it won’t trust every Render customer. Useful protection against an easy config mistake.
https://t.co/aR0KycR0I5
@kellabyte I’d love a small change log alongside this: what the optimizer changed, what improved, and a way to pin the previous behavior. That would make automatic tuning much easier to trust.