Evening brief

Monday, June 29

4 people · 0 companies · 5 topics

TL;DR GPT-5.6 Sol previewed with top TerminalBench 2.1 score (88.8%), edging Mythos 5; Anthropic’s Mythos 5 gets partial US government clearance to ~100 companies; Z.ai’s GLM-5.2 open weights sit #1 on Artificial Analysis. Note: Reddit and several direct-fetch domains were blocked by egress policy this run; no morning brief existed for today, so this covers the full recent window.

Models & research

  • OpenAI previewed GPT-5.6 as three tiers on June 26: Sol (flagship, $5/$30 per 1M tokens), Terra (balanced, $2.50/$15), Luna (fast/cheap, $1/$6). Limited to ~20 US-vetted partners cleared by ONCD/OSTP; GA expected in weeks. Sol scored 88.8% on TerminalBench 2.1, just above Mythos 5 (88.0%). GPT-4.5 retired same day. VentureBeat · Axios · ExplainX benchmarks

  • US Commerce Secretary Lutnick authorized Anthropic to release Mythos 5 to ~100 vetted companies and federal agencies on June 26, primarily members of Project Glasswing. Fable 5 is still locked down. Mythos had been offline since June 12 due to export/cybersecurity concerns over foreign national employees. CNN · CNBC · Semafor

  • simonw wrote that GLM-5.2 from Z.ai is “probably the most powerful text-only open weights LLM” (June 17). The 753B MoE model (MIT licensed, ~39B active params, 1M token context) leads Artificial Analysis and beats GPT-5.5 on SWE-bench Pro (62.1 vs 58.6). Requires 256-512GB RAM to run locally. simonwillison.net · Artificial Analysis · VentureBeat

  • Gemini 3.5 Pro, promised by Sundar Pichai at Google I/O for “next month,” is reportedly slipping to July for additional tuning. Gemini 3.5 Flash is already live. Analytics Insight

  • On benchmarks as of early June: Claude Opus 4.8 leads overall (67.9 on Artificial Analysis), GPT-5.5 holds coding edge (59.1% SWE-bench). Chinese models Qwen 3.7 Max and MiniMax 3 are matching Western flagships at a fraction of the cost. llm-stats

People

  • karpathy joined Anthropic’s pre-training team May 19 to lead a new team using Claude to accelerate pretraining research. His announcement tweet: x.com/karpathy. Around June 24-26 he posted that X “has never been this toxic and Reddit-like”; Elon replied suggesting an algorithm overhaul. TechCrunch

  • simonw posted June 26: “What happened after 2,000 people tried to hack my AI assistant” — a write-up on red-teaming his own AI tool. simonwillison.net

  • emollick posted about a large China study showing AI homework use hurts test scores; his read: AI tutoring to support learning is good, replacing effort is bad. x.com/emollick

  • hwchase17 tweeted on building reliable agents and repeatable improvement lifecycles. x.com/hwchase17

Tools & dev

  • Claude Code wave of updates this month: hierarchical agent spawning up to 3 levels deep, session artifacts (shareable progress pages, Team/Enterprise beta), /rewind now works across /clear boundaries, 37% CPU reduction during streaming via 100ms text coalescing, claude mcp login shell auth, background subagent permission prompts now surface to main session. Releasebot · Sitepoint

  • Anthropic launched Claude Managed Agents in controlled sandboxes: agents can now run in self-hosted or provider-managed environments (Cloudflare, Daytona, Modal, Vercel) connected to private MCP servers. Also added Trusted Devices for Remote Control on Team/Enterprise. Anthropic newsroom

  • SpaceX agreed to acquire Anysphere/Cursor for $60B in stock on June 16, days after SpaceX IPO. Cursor had hit $4B ARR. Deal is pending regulatory review. CNBC · TechCrunch

OSS

  • Today’s GitHub trending standout: DeusData/codebase-memory-mcp (2,190 stars today) — an MCP server that builds a persistent knowledge graph from a codebase for LLM context. GitHub

  • browser-use/video-use trending today — edit videos via coding agents. GitHub

  • GLM-5.2 weights on Hugging Face under MIT license; already available on Z.ai API and 20+ third-party coding environments. interconnects.ai

Other

  • Polymarket pricing Anthropic at 99% for best AI model end of June (~$22M volume), resolving tomorrow June 30. Likely locked in given Claude Fable 5/Opus 4.8 dominance on leaderboards. (Polymarket: 99% yes, ~$22M volume) Polymarket

  • Colorado’s AI Act was substantially rewritten and its effective date pushed from June 30 to January 1, 2027. Governor Polis signed SB 189 on May 14 — the revised law strips the risk-based framework and duty of care, replacing them with narrow disclosure requirements. Troutman · Hunton

  • Both Anthropic and OpenAI have filed confidential S-1s: Anthropic at ~$965B valuation (June 1), OpenAI at ~$730B (June 8), both targeting potential fall 2026 offerings. CNBC