The Big Picture
G7 closing day, and the timing is almost too on-the-nose. While Altman, Amodei, and Hassabis sit down with seven world leaders at Évian for a working lunch on “safe, rapid and effective deployment” of AI — six days after Lutnick yanked Anthropic’s flagship line off the global market — Z.ai drops the GLM-5.2 weights on Hugging Face under an MIT license. 753B parameters total, 40B active, 1M context, 74.4% on FrontierSWE Dominance (above GPT-5.5’s 72.6, neck-and-neck with Opus 4.8’s 75.1), at roughly one-sixth the cost. That’s the only AI story today, even if it isn’t the one happening in the room. The lunch will produce a voluntary-commitments package on youth safety and cyber/bio frontier risks — the kind of pledge labs sign at every summit. The G7 also locked a binding 60%-supply cap on critical-mineral imports from any single country by 2030 — the substantive trade-policy decision of the week. Trump arrived late and joked “I’m the boss” to an amused-looking Macron. Anthropic and Commerce are still on regular calls toward a Fable/Mythos restoration deal; no announcement today. The actual story of the week is a Chinese MoE that just made the trusted-partners debate look quaint.
Evening update: Macron put a calendar on it at his closing presser — a platform among “a handful of Western democracies” inside a month, leaders reconvening in September, framed so the G7 “doesn’t just produce the most capable AI, but also the second most capable.” That’s the actual deliverable from the week, and it’s a process, not an outcome — the Leaders’ Communiqué shipped no formalized “trusted partners” framework, so Lutnick’s Monday floor remains a floor. Prediction markets read the lunch the other way: Polymarket’s Fable-5-restored-by-July-1 contract slid sharply from the 71% it carried at midday. Fable 5 still dark; isfable5back.com still “No.” Separately, Anthropic shipped Claude Code v2.1.181 — own story below.
Z.ai open-sources GLM-5.2 on Hugging Face — 753B MoE beats GPT-5.5 on FrontierSWE at 1/6 cost
Today VentureBeat, Hugging Face, Artificial Analysis, SCMP, InfoWorld
Z.ai (formerly Zhipu) released the full weights of GLM-5.2 today under an MIT open-source license on Hugging Face (zai-org/GLM-5.2), including an FP8 variant. The 753B-parameter MoE runs 40B active per token (256 routed experts plus one shared, 8 active across 78 layers) with a 1M-token context window. Headline benchmarks shipped with the weights: 74.4% on FrontierSWE Dominance, above GPT-5.5’s 72.6% and within striking distance of Claude Opus 4.8’s 75.1% — at roughly one-sixth the per-token cost. Artificial Analysis lists it as the new top open-weights model on its Intelligence Index at 51, on the Pareto frontier of intelligence-vs-cost-per-task. Z.ai had announced GLM-5.2 on Saturday June 13 alongside its standalone Coding Plan; today is the weights drop and the benchmark publication promised at launch. Zhipu’s Shenzhen-listed stock rocketed on the news.
VentureBeat’s framing: first frontier-grade open-weights coding model since DeepSeek V4, and the first non-US release to plausibly close the gap with closed-frontier at a price point that breaks the “trusted partners” debate before Wednesday’s G7 lunch even sits down. The G7 is convened to discuss the safe deployment of US-flagship AI six days after that flagship was yanked off the global market — and an MIT-licensed alternative just dropped on Hugging Face during the convening.
Macron commits to Western-democracy AI platform inside a month, G7 leaders to reconvene in September
Today Updated 8:00 PM Washington Post, AP, CNBC, The Next Web, Euronews
At the Évian working lunch Wednesday, Anthropic’s Dario Amodei and Google DeepMind’s Demis Hassabis pitched a US-led AI coalition: structured access to frontier models among allies, plus chip-and-critical-component trade explicitly excluding China, plus joint work on cyber, bio and intelligence risks. Canadian PM Mark Carney made Canada the first G7 leader to publicly back the architecture. Sam Altman — first CEO to speak, seated between Trump and Egypt’s el-Sisi — went the other direction, calling for “an international forum… that establishes globally accepted standards for testing, provides expert and impartial analysis of capabilities and risks, and serves as a venue for cooperation among nations,” with the sharper note that AI safety “should not be left to tech companies.” At his closing presser Wednesday afternoon, Macron formalized what had been “progress in coming weeks” at midday: a platform among “a handful of Western democracies” to be established within the next month, with G7 leaders reconvening in September. His framing: democracies coordinate so the G7 “doesn’t just produce the most capable AI, but also the second most capable.” The Leaders’ Communiqué adopted at the summit’s close carried no formalized “trusted partners” framework — it commits G7 governments only to involve financial, regulatory, and cybersecurity bodies in assessing frontier AI’s effects on stability, productivity, and jobs. Around a dozen tech execs attended the lunch, including Mensch (Mistral), Rombach (Black Forest Labs), Benioff (Salesforce), Wang (Meta), Riparbelli (Synthesia), Kumar (Sarvam) and Ito (Sakana).
The architecture now has a calendar (platform inside a month, leaders meeting September) but no binding text in the communiqué — Lutnick’s Monday “trusted partners” floor remains a floor, with a soft Macron-chaired process attached. Macron’s “second most capable” line — the democracies as a coordinated bloc, not a third-place finish behind a US/China duopoly — is the take that’ll travel.
G7 agrees no single country should supply more than 60% of critical mineral imports by 2030 — binding quotas for defense
Today Bloomberg, The Northern Miner, Reuters, Yahoo Finance
At the summit’s closing session in Évian, G7 leaders agreed that no single country should supply more than 60% of their critical mineral imports by 2030 — an explicit move to reduce dependence on China. The package includes binding quotas for companies in some industrial sectors (defense manufacturers specifically named) and a new platform combining efforts from recycling and new mining projects. Treasury Secretary Scott Bessent said Australia and India have been invited into the G7 critical-minerals working group. China currently refines between 47% and 87% of copper, lithium, cobalt, graphite and rare earths — the most extreme dependence Western defense industrials carry on any single supplier.
Bloomberg’s framing: this is the substantive trade-policy decision of the summit week — a hard numeric cap with binding-quota teeth and a 2030 deadline. It contrasts pointedly with the voluntary-pledge architecture playing out one room over at the AI lunch. The political logic is the same as Carney’s Monday “build out and diversify” framing, but locked at G7 scale: concentration risk is no longer a topic, it’s a quota.
Anthropic-Commerce talks continue past G7 — no Fable 5 date, prediction markets reprice past July 1
Today Updated 8:00 PM Globe and Mail, Yahoo Finance, Polymarket, isfable5back.com
As of Wednesday evening — with the G7 summit closed and the AI working lunch concluded — there is no announced deal to restore Claude Fable 5 or Mythos 5. Per Tuesday Globe and Mail and Yahoo Finance reporting, Commerce Secretary Howard Lutnick continues holding regular calls with Anthropic, with the agenda split across safety protocols, international access frameworks, and federal-agency use. Anthropic’s technical bench (Frontier Red Team lead Logan Graham, Safeguards head Dave Orr, Nicholas Carlini) closed two consecutive days of Commerce meetings Tuesday. No Lutnick post-lunch statement, no Amodei post-lunch interview. Through Wednesday afternoon Polymarket’s “restored before July 1” contract slid sharply from the 71% it carried at midday, with the market re-anchoring on dates past July 1 as Macron’s closing presser pushed the “trusted partners” timeline into a multi-week platform process. The availability tracker isfable5back.com still reads “No.”
The asymmetric political fact of the day: Amodei walked into the G7 lunch with no resolution, walked out with no resolution, and the alternative-Chinese-MoE on Hugging Face stayed up the whole time. The administration’s leverage is now denominated in days, not directives — and prediction-market repricing through the afternoon says the day didn’t move the timeline forward, it moved it back. Every day Fable 5 stays dark while GLM-5.2 weights live on Hugging Face is a day developers acclimate to substitution.
Claude Code v2.1.179 ships reliability pass — mid-stream drops preserved, WSL2 scroll, subagent transcript fixes
Yesterday Anthropic changelog
Claude Code shipped v2.1.179 Tuesday — a cleanup release behind Monday’s v2.1.178 Tool(param:value) permissions push. No new features, but a stack of long-tail reliability fixes. Headline: mid-stream connection drops now preserve partial responses instead of showing a raw error, and the spinner no longer gets stuck at “running tool.” WSL2 mouse-wheel scrolling under Windows Terminal and VS Code is fixed (regression in 2.1.172). A sandbox denyRead/allowRead glob over a large directory tree no longer makes the Bash tool description enormous and the session unusable on Linux. Subagent transcript view (Ctrl+O) now actually shows the subagent’s transcript. Other fixes: the feedback-survey no longer captures a single-digit reply as a session rating; the welcome screen stops stacking multiple promo banners (max one per session); clicking the prompt input returns focus from the subagent/footer panel; remote-session background tasks no longer appear stuck as “still running” between turns; plugin loading is faster in remote sessions.
The pattern in the changelog notes — connection-flap protection, the long-standing WSL2/VS Code scroll regression, the monorepo sandbox glob blowup, the subagent transcript view that wasn’t showing anything — reads as Anthropic shipping into remote-session and long-running-agent operational pain. None of these would show up in a marketing slide; all of them disproportionately affect daily UX.
Claude Code v2.1.181 ships /config key=value inline, mobile-push suppression, mid-thinking auto-retry
Today Anthropic changelog
Claude Code shipped v2.1.181 Wednesday — the second cleanup release of the week behind Tuesday’s v2.1.179. Headlines: /config key=value sets any setting inline (works in interactive, -p, and Remote Control); CLAUDE_CLIENT_PRESENCE_FILE suppresses mobile push when a marker file says you’re at the machine; long paragraphs now stream line-by-line instead of waiting for the first line break; mid-thinking API drops auto-retry instead of dying with “Connection closed while thinking.” Subagent panel auto-hides idle agents after 30s and caps at 5 rows with scroll hints. macOS gets a sandbox.allowAppleEvents opt-in that fixes open, osascript, and browser-auth flows previously failing with error -600. Bundled Bun bumped to 1.4. Plus 25+ bug fixes: prompt caching on custom ANTHROPIC_BASE_URL/Foundry, Write/Edit producing 0-byte files on network drives, startup regressions, /recap and forks using the previous model immediately after a model switch, AWS credential refresh thrash, AskUserQuestion silently dropping typed “Other” answers.
The three-day v2.1.178 → v2.1.179 → v2.1.181 cadence (Mon → Tue → Wed) is Anthropic shipping into long-tail agentic-engineering pain — remote sessions, subagents, sandbox, model switching, network-drive edits. The CLAUDE_CLIENT_PRESENCE_FILE primitive is the kind of detail that only ships when someone’s actually feeling the pain of getting buzzed by their phone while sitting at the same desk Claude’s already running on.