Issue #29 | DailyPrompt

01 announcement

Anthropic Edges Past OpenAI on US Business AI Spend for the First Time, Ramp Index Says

April's Ramp AI Index, published on 15 May, shows Claude finally pulling ahead on US business payments, with 34.4% of the 50,000 companies on Ramp's platform paying Anthropic against 32.3% paying OpenAI. That is a 3.8-point monthly jump for Claude and a 2.9-point fall for ChatGPT. VentureBeat's writeup credits Claude Code, now the fastest-growing product in Anthropic's history and reportedly responsible for 4% of public GitHub commits, plus a steady Sonnet 4.6 and Opus 4.7 cadence. The same piece lists three threats that could erase the lead inside a quarter: compute crunches, Codex's price war and pending court rulings.

VentureBeat →

research

Apple Quietly Patches macOS Memory Bugs That Only Anthropic's Mythos Had Spotted

Researchers told 9to5Mac on 14 May that an April preview of Claude Mythos chained together two previously unknown bugs to corrupt memory inside macOS, bypassing security technology that Apple has spent years hardening and that no human or model has previously broken. The work was disclosed responsibly through Anthropic's 40-partner Mythos preview programme, and Apple has shipped fixes, but the finding is the strongest evidence yet that Mythos can surface offensive capabilities outside the reach of existing red teams. It lands the same week as a closed-door House briefing on the model.

9to5Mac

ClaudeDevs

@ClaudeDevs

Happy Friday! We've reset everyone's 5-hour and weekly rate limits.

11h View on X →

announcement

$ cat story_03.md

House Homeland Security Pulls Anthropic In Behind Closed Doors to Explain Mythos

The House Committee on Homeland Security held a closed-door briefing with Anthropic on 14 May to walk members through Claude Mythos's capabilities, the national security implications of releasing it and the policy choices it forces on Washington. The Hill reports the session is part of a wider committee effort to size up offensive cyber risks from frontier models before the next AI safety reauthorisation bill. Anthropic has shared Mythos with about 40 organisations, including Microsoft, Apple, AWS, CrowdStrike and JPMorgan, and is still resisting White House pressure to widen access.

$ open the-hill →

community No. 04

Claude Keeps Telling People to Go to Bed, and Anthropic Is Not Entirely Sure Why

Fortune charted on 14 May the long-running and slightly surreal pattern of Claude breaking off coding sessions to tell users to get some rest, drink water or stop for the day, sometimes in the middle of the afternoon. An Anthropic staff member, Sam McAllister, called it a 'character tic' in a reply on X and said the team hopes to remove it in future models. Theories on offer range from constitutional training nudging Claude toward user wellbeing, to context-window pressure encouraging wrap-up phrases like 'good night' when the model senses a conversation winding down.

Source: Fortune →

ClaudeDevs

@ClaudeDevs

Useful tip to cut time-to-first-token on longer prompts in the API: pre-warm the prompt cache. Send your system prompt before the user request. Claude writes it to the cache, but skips generating any output. When the real user request lands, it'll hit a warm cache.

May 15 View on X →

opinion

Axios Reads the New Claude Agent Credit Pool as a Defensive Move Against Codex

Axios's 14 May take on the 15 June split between chat and agent credits frames the change as the moment Anthropic gave up on flat-rate subscriptions for software that calls Claude headlessly. The piece quotes Claude Code product manager Noah Zweben fielding accusations of 'gaslighting' from a developer base that has just watched OpenAI hand out two free months of Codex to new business customers and add a Codex route inside OpenClaw's provider docs. Axios concludes that all-you-can-eat AI plans probably do not survive the agent era on either side of the fight.

Axios →

research

Microsoft Says a New Multi-Agent System Just Beat Claude Mythos on a Cybersecurity Benchmark

GeekWire reports that Microsoft Research has posted results from a new multi-agent system that scored higher than Claude Mythos on a private cybersecurity benchmark covering exploit discovery and patch generation. Microsoft frames the work as evidence that orchestrated agents can match dedicated frontier models on offensive security tasks without handing any single model the dual-use capabilities that have made Mythos politically charged. The full paper has not yet been released, and Anthropic has not commented publicly on the comparison.

Read on GeekWire →