Thursday, July 30

The signal

1. Anthropic ships Claude Opus 5 and Sonnet 5 (July 24) What happened: Opus 5 launched with 1M-token context, 128K max output, $5/$25 per MTok standard (half Fable 5 price), and a 5-level effort toggle (low/medium/high/xhigh/max) spanning ~407 Elo points on GDPval-AA v2. It lands within 0.5% of Fable 5 on CursorBench 3.2, beats Fable 5 on OSWorld 2.0, and more than doubles Opus 4.8 on Frontier-Bench v0.1. Zero data retention vs. Fable 5’s 30-day minimum. Why it matters: Near-frontier agentic and coding performance at half the Fable 5 price changes the cost-per-task math immediately; any team paying Fable 5 rates now has a direct comparison to run. Confidence: Confirmed Sources: CNBC (independent) · Anthropic (primary) · Enterprise DNA (independent)

2. Anthropic publishes formal open-weights position paper (July 27) What happened: Anthropic posted a public position stating it will not open-source frontier-capable models, citing catastrophic-risk arguments. The post came the same week that 77 AI firms (Nvidia, Hugging Face, Mistral, and others – but not OpenAI, Google, or Anthropic) signed a pro-open-weights letter urging the EU and US to protect open release rights. Why it matters: Anthropic has drawn a formal line between safety-justified closed weights and the broader open-source AI movement; the timing against the 77-firm letter makes it a direct political statement, not just a research position. Confidence: Confirmed Sources: Anthropic (primary) · ppc.land (community)

3. Anthropic Mythos cracks HAWK post-quantum signature scheme in 60 hours (July 28) What happened: Anthropic published research showing Claude Mythos independently identified cryptographic weaknesses and broke HAWK, a NIST-evaluated post-quantum signature scheme, in 60 hours – a vulnerability that took human cryptographers approximately two years to find. Why it matters: First published case of a frontier AI autonomously defeating a post-quantum cryptographic primitive; implies that security assumptions built on NIST-evaluated post-quantum schemes may need to be revisited faster than the standards process moves. Confidence: Confirmed Sources: Anthropic Research (primary) · The Hacker News (independent)

Since last brief

• oss Kimi K3 weights shipped July 26 (one day early); custom commercial license requires MaaS businesses over $20M/12-month revenue to sign a separate deal with Moonshot before use. Digital Applied · natolambert analysis: Interconnects
• agentic-ai AI Kill Switch Act introduced July 23 (Reps. Lieu & Moran, bipartisan): DHS gets authority to order slowdowns or full shutdown of AI systems above $500M revenue or $100M compute threshold; $2M/day noncompliance, $20M/day for defying an emergency directive. Nextgov/FCW · CNBC
• sama Altman on Capitol Hill July 29 told senators “we may have to pace the rate of AI development to give ourselves enough time for society to harden”; expressed support for AI cybersecurity guardrail legislation. More than 1,100 employees across OpenAI, Anthropic, Google, and Meta signed an open letter the same week calling for a verifiable coordinated slowdown. Bloomberg · Semafor
• model-releases OpenAI retired GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini in ChatGPT this week (273 pts / 357 comments on HN). HN thread

Models & research

Claude Opus 5

Claim: 1M token context, 128K output, $5/$25 per MTok standard (fast mode $10/$50, batch $2.50/$12.50). 5-level effort toggle; billing spans ~8x token consumption between low and max. Zero data retention. Evidence: CNBC and third-party coverage confirm pricing. CursorBench 3.2 within 0.5% of Fable 5; OSWorld 2.0 beats Fable 5 at ~1/3 cost; more than doubles Opus 4.8 on Frontier-Bench v0.1. Community reports noted elevated API errors July 27 and a broader Claude outage July 29 – both appear resolved. Caveat: Benchmarks are Anthropic-published; independent aggregated evals not yet out. Fast mode erases most cost advantage. “Opus 5 is a really bad model” discourse ran July 27-28 on HN and Reddit – likely a mix of expectation mismatch and the simultaneous outage rather than a fundamental quality regression. Worth your time: Read. simonw wrote a first-look July 24 if you want a practitioner sanity check. Anthropic · simonwillison.net

Kimi K3 (open weights, July 26)

Claim: 1.4 TB MXFP4 weights on Hugging Face; 2.8T-parameter MoE, 64-accelerator minimum for deployment. Custom license: free for most, but MaaS >$20M/12-month revenue requires a signed Moonshot commercial deal. Evidence: Weights are live and verifiable on Hugging Face; Together AI and Modal had day-0 hosted inference. natolambert framed it as the largest open-weight release ever. Independent inference benchmarks are just starting to appear. Caveat: 64-accelerator floor is a real barrier. MXFP4 needs patches in standard llama.cpp/vLLM. All published task-resolution benchmarks still come from Moonshot. License clause is easy to miss and non-trivial for API businesses. Worth your time: Read. Roo/Beehiiv · Interconnects

Discovering Cryptographic Weaknesses with Claude (Mythos)

Claim: Anthropic Mythos found and exploited weaknesses in HAWK (a NIST-evaluated post-quantum signature scheme) in 60 hours autonomously. Evidence: Anthropic research blog (primary source). Independent security press coverage confirms the claim and the 60-hour timeline vs. 2-year human comparison. Caveat: Paper details the methodology; external cryptographers have not yet independently replicated or disputed the finding. Scope is limited to HAWK; claims about other post-quantum schemes would need separate evaluation. Worth your time: Read – if you are building anything that relies on post-quantum assumptions. Anthropic Research · The Hacker News

Tools worth touching

Claude Code (Opus 5 default + background review)

• Does: Anthropic’s AI coding CLI; Opus 5 is now the default Opus model as of July 24
• Why now: Background /code-review shipped alongside Opus 5, running review asynchronously with richer screen-reader output; EnterWorktree now prompts before entering worktrees outside .claude/worktrees/
• Try if: You have been running Opus 4.x workflows and have not updated your default model config
• Skip if: You are pinned to Fable 5 for quality reasons; the cost differential matters less than consistency
• Links: Claude Code changelog · Releasebot

different-ai/openwork

• Does: Open-source desktop AI agent built on OpenCode; explicit alternative to Claude Cowork; users supply their own API key
• Why now: Trending on GitHub July 30 with 18k+ stars; no $200/month subscription, granular folder-level permission controls
• Try if: You want Cowork-style local agent execution with full key control and no subscription
• Skip if: You need Cowork’s collaboration and team features
• Links: github.com/different-ai/openwork

Moves on the board

Other

• GPT-6 “Spud”: no announcement through July 30; pre-training base reported abandoned in favor of a larger foundation. FelloAI
• 1,100+ employees (OpenAI, Anthropic, Google, Meta) signed an open letter calling on the US government to build infrastructure for a verifiable coordinated AI slowdown; tied to the White House August 1 safety framework deadline. Al Jazeera
• GitHub trending July 30: huggingface/speech-to-speech (8k stars) – local voice agents pipeline; mvanhorn/last30days-skill (55k stars) – AI research aggregator across Reddit/X/YouTube. GitHub trending

Watchlist

• OpenAI ExploitGym: AI Kill Switch Act in committee (introduced July 23); no NIST formal rule; watching for Senate companion bill, committee markup, and AISI cross-lab guidance (opened 2026-07-23)
• GPT-6 “Spud”: pre-training base reportedly replaced with larger foundation; “approximately one month” window from July 21 elapsed with no announcement; watching for OpenAI blog post or API changelog (opened 2026-07-22)
• Anthropic open-weights position: 77-firm coalition response and any EU/US regulatory action citing the paper are pending; watching for follow-up from the pro-open-weights letter signatories (opened 2026-07-30)