Skip to content
DailyPrompt

your daily dose of claude

01 product

Claude Code 2.1.119 and 2.1.120 Ship a Fresh Wave of Regressions Days After Postmortem

Less than a week after Anthropic's high-profile mea culpa over Claude Code's monthlong slump, eight new regressions filed on 24 April are stacking up in the issue tracker, including a hard crash on `claude --resume` and `--continue`, a silent reroute from the 200k to the 1M context window, and a sandbox config that ignores its own opt-outs. A community survival checklist is now recommending users pin v2.1.117 and disable auto-update until fixes land. The list adds awkward weight to Anthropic's promise of better internal testing for future builds.

GitHub (community gist) →
product

Anthropic Ships a Rate Limits API So Admins Can Stop Guessing at Quotas

The Claude Developer Platform has quietly released a Rate Limits API that lets administrators query the configured request, input, and output token limits for their organisation and individual workspaces in real time. The endpoint plugs into the existing Admin API and is aimed at teams hitting tier ceilings during agent rollouts, who until now had to lean on the console UI to see their headroom. It lands the same week Anthropic conceded that its compute crunch is unlikely to ease before late 2026.

Claude Developer Platform
F
Felix Rieseberg
@felixrieseberg

Tiny feature, also a bit of a nerdy one: We've added deep links so you can pre-fill new sessions in Chat, Cowork, or Code with a claude:// link. I like using it for composition in my workflows or adding 'Open in Claude' buttons to other apps.

Apr 23 View on X →
product

$ cat story_03.md

Memory Lands in Public Beta for Claude Managed Agents

Anthropic's hosted agent runtime has graduated its memory layer to public beta, available to anyone using the standard managed-agents-2026-04-01 beta header. Agents can now persist learnings across sessions through an intelligence-optimised store, exported as files so developers can audit, edit, or wipe what their agent has remembered. The feature plugs a long-standing gap in long-horizon agent work, where teams were stitching together their own retrieval layers around Claude.

$ open claude-developer-platform →
A
Alex Finn
@AlexFinn

It happened. Claude Code has been dethroned. I've lived in Codex w/ ChatGPT 5.5 the last couple of days and it is the: • Smartest model • Best all around AI app • Most powerful features (computer use and image gen)

12h View on X →
research No. 04

Project Deal's Uncomfortable Finding: Weaker AI Agents Leave Money on the Table — and Their Owners Don't Notice

Following Anthropic's Project Deal write-up, The Decoder zooms in on the experiment's most unsettling result: participants paired with Claude Haiku 4.5 ended up roughly $2.45 worse off per item than those represented by Opus, yet rated the fairness of their negotiations identically to the better-served group. The piece argues that an invisible quality gap between AI tiers could quietly entrench a two-speed economy as agent commerce scales. Anthropic itself flags it as an open question for policymakers in any forthcoming agent-to-agent marketplace.

Source: The Decoder →