01
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
02
OpenAI Begins GPT-5.5-Cyber Rollout — Then Quietly Adopts Anthropic's Restricted-Access Model
- Sam Altman: rollout 'to critical cyber defenders in the next few days' via the expanded Trusted Access for Cyber (TAC) program.
- TAC scaled to thousands of vetted defenders across government, critical infrastructure, security vendors, cloud platforms, and financial institutions.
- TechCrunch headline of the day: 'After dissing Anthropic for limiting Mythos, OpenAI restricts access to Cyber, too' — Altman called the same approach 'fear-based marketing' two weeks ago.
- No technical card or benchmarks against Mythos published; positioned as a fine-tuned GPT-5.5 with relaxed guardrails for security work.
models techcrunch.com
03
Mistral Ships Medium 3.5 and Vibe Remote Agents — 77.6% on SWE-bench Verified, Open Weights
- 128B dense multimodal model, 256K context, single set of weights for instruction-following, reasoning, and coding.
- 77.6% SWE-bench Verified and 91.4% on τ³-Telecom; beats Devstral 2 and Qwen3.5 397B-A17B on agentic benchmarks.
- Vibe Remote Agents run cloud sandboxes in parallel from the Vibe CLI or Le Chat, open PRs on GitHub, and 'teleport' a local session to the cloud.
- API at $1.5/M input, $7.5/M output; weights on Hugging Face under a modified MIT license; HN discussion at item 47949642.
models mistral.ai
04
Tencent Open-Sources a 440 MB On-Device Translator That Beats 72B Models in 33 Languages
- Hy-MT1.5-1.8B-1.25bit ships at 440 MB after Sherry ternary quantization (3:4 sparsity, 1.25 effective bits) — paper accepted to ACL 2026.
- Outperforms Tower-Plus-72B, Qwen3-32B, Microsoft Translator, and Doubao on standard zh/en↔X benchmarks; 1,056 translation directions across 33 languages.
- Custom STQ kernel hits SIMD alignment on mobile CPUs — runs offline on phones with limited RAM.
- Weights and GGUFs on Hugging Face (tencent/Hy-MT1.5-1.8B-1.25bit); release timed to the May Day holiday travel surge in China.
open-source huggingface.co
05
ElevenLabs Relaunches ElevenMusic as a Listen-Remix-Create Hybrid With ~4,000 Licensed Artists
- Pivot from a B2B production library to a consumer app that streams, remixes (genre/tempo), and generates from lyric/melody/mood prompts.
- Free tier caps at 7 songs/day; Pro at $9.99/mo unlocks 500 generations/mo; positioned as 'fully licensed, artist-first by design.'
- Roughly 4,000 mostly-emerging human artists onboarded with revenue share; direct shot at Suno and Udio amid mounting label lawsuits.
- Launch coincides with Taylor Swift's legal team filing on AI music likeness — TechCrunch frames the consumer pivot as the bet that licensing wins this round.
industry musically.com