AI Daily Dev — May 2026

Anthropic Refunds Claude Code Users After 'HERMES.md' Commit Strings Silently Drained Quotas

Substring match in Claude Code's harness-detection logic routed any session whose recent git history contained 'HERMES.md' to extra-usage billing instead of plan quota.
One Max-20x user lost $200 in overage credits while plan usage was still at 13%; charges occurred silently with no UI warning.
GitHub issue anthropics/claude-code#53262 traces the bug to the wider crackdown on third-party agent harnesses (Hermes, Codex CLI, Goose).
Hit HN front page at 828 points; refunds plus $200 in credits issued ~9 hours later. Discussion: HN item 47948012.

tools github.com

Altman tweet at 2:33am on May 2: 'sign in to openclaw with your chatgpt account now and use your subscription there.'
ChatGPT Plus ($23/mo) now backs GPT-5.4 inside OpenClaw — directly opposite Anthropic's April 4 ban on running Claude Pro/Max plans through third-party harnesses.
OpenClaw sits at ~346k GitHub stars and 3.2M users; making ChatGPT the auth/billing layer locks that distribution to OpenAI.
Strategic split: OpenAI is buying agent volume, Anthropic is protecting margins on autonomous-agent token spend.

industry thenextweb.com

Public beta launched April 30 to all Claude Enterprise customers via the Claude.ai sidebar and claude.ai/security.
Opus 4.7 traces data flows across files and modules instead of pattern-matching for known signatures, then writes verified patches.
Adds scheduled scans, Slack/Jira webhooks, CSV/Markdown exports, and triage tracking on top of the February research preview.
Launch partners include CrowdStrike, Microsoft Security, Palo Alto Networks, SentinelOne, TrendAI, and Wiz; Team and Max access 'coming soon.'

tools siliconangle.com

CTO Praveen Neppalli Naga told the company it had spent the full annual AI budget by April after rolling Claude Code to 5,000 engineers in December.
Claude Code adoption jumped 32% → 84% of engineers; per-developer API spend ran $500–$2,000/month.
70% of committed code now originates from AI; ~11% of live backend updates ship with no human in the loop.
HN front page (item 47976415) — debate centers on whether $100k/year per engineer in token spend pencils out vs hiring.

industry news.ycombinator.com

Anthropic JV with Blackstone, Hellman & Friedman, Goldman Sachs at $1.5B — $300M each from Anthropic, Blackstone, H&F; $150M from Goldman; co-investors include Apollo, GIC, Sequoia, General Atlantic, Leonard Green.
OpenAI's 'The Deployment Company' closes $4B from 19 investors at a $10B valuation, anchored by TPG, Brookfield, Advent, Bain — with a 17.5% guaranteed return over five years.
Both vehicles embed forward-deployed engineers inside customer companies (Palantir-style) instead of selling API credits — direct shot at McKinsey/Accenture-style AI consulting.
Initial pipeline: each consortium's PE-owned portfolio in healthcare, manufacturing, financial services, retail, and real estate; both labs are reportedly targeting fall 2026 IPOs.

industry techcrunch.com

Replaces GPT-5.3 Instant across ChatGPT and ships in the API as `chat-latest`; old model stays available for paid users for three months.
52.5% fewer hallucinated claims on high-stakes medical, legal, and financial prompts; 37.3% fewer incorrect claims on user-flagged conversations.
Responses are 30.2% shorter in words and 29.2% shorter in lines on internal evals.
AIME 2025 jumps to 81.2 from 65.4; MMMU-Pro to 76 from 69.2.
Personalization from past chats, files, and connected Gmail rolling out to Plus and Pro on web.

models openai.com

Anthropic takes all the compute capacity at SpaceX's Memphis Colossus 1 site: 300+ MW and 220,000+ Nvidia H100/H200/GB200 GPUs landing within the month.
Claude Code five-hour limits doubled for Pro, Max, Team, and Enterprise plans; peak-hour throttling on Pro and Max removed.
Claude Opus API tier-1 input rate limit jumps 1500%; output tokens-per-minute up 900%.
Notable plot twist: Musk has publicly called Anthropic 'misanthropic and evil' while suing OpenAI; Anthropic also expressed interest in developing multi-gigawatt orbital compute with SpaceX.
On HN front page; Bloomberg, CNBC, and Engadget all leading with the Memphis tie-up.

industry anthropic.com

GPT-Realtime-2: speech-in/speech-out model with GPT-5-class reasoning, parallel tool calls, spoken preambles like 'let me check that,' and a 128K context (up from 32K).
GPT-Realtime-Translate: live speech-to-speech translation across 70+ input languages and 13 output languages, designed to keep pace with the speaker.
GPT-Realtime-Whisper: streaming speech-to-text that emits text as the user is still talking — aimed at captions, meeting notes, and live agents.
GPT-Realtime-2 scores 15.2pp higher than GPT-Realtime-1.5 on Big Bench Audio; Zillow reports a 26pp jump in call success rates in early testing.
Pricing: $32/$64 per 1M audio input/output tokens for GPT-Realtime-2; $0.034/min for Translate; $0.017/min for Whisper.

models openai.com

Multi-agent workbench built on Gemini 3.1 — a project coordinator fans research questions out to sub-agents that write code, search literature, and attempt proofs.
Solved 23 of 48 problems (48%) on Epoch AI's FrontierMath Tier 4, more than double the 19% scored by the underlying Gemini 3.1 Pro and a new SOTA.
Oxford's Marc Lackenby cracked Problem 21.10 from the Kourovka Notebook after a reviewer agent flagged a flaw in the AI's first proof attempt and surfaced a strategy he could finish.
Stateful, asynchronous workspace that tracks failed hypotheses and outputs LaTeX with margin annotations — DeepMind frames it as 'Claude Code for mathematicians.'

research arxiv.org

Google Threat Intelligence Group says it disrupted the first AI-built zero-day used in a planned mass-exploitation campaign.
The exploit, written in Python with help from cybercrime model OpenClaw, bypasses 2FA on a widely deployed open-source web admin tool.
Attackers tipped Google off by leaving a hallucinated CVSS score inside the script; Google notified the vendor before any victims were hit.
GTIG analyst John Hultquist: "It's here. The era of AI-driven vulnerability and exploitation is already here."
Front-page coverage at CNBC, Bloomberg, NBC News, and Bleeping Computer; on HN with active debate about Anthropic's Mythos and Project Glasswing.

industry cloud.google.com