Back to Blog
Galen Guan

Daily Briefing 2026-09-02: US Strikes Iran Again, Claude Fable 5.1 Lands, Ternus Takes Over Apple

📰 Daily Briefing · Wednesday, September 2, 2026

🌐 Politics · Geopolitics

  • US airstrikes Iranian IRGC targets; Iran fires 13 ballistic missiles at Jordan base in retaliation — On September 1, US Central Command announced strikes on Iranian IRGC air-defense, radar, mine-laying, and communications facilities. Iran responded by firing 13 ballistic missiles at a US base in Jordan (10 intercepted) and drone attacks on a Bahrain base; Jordan and Kuwait activated air defenses. Brent crude rose more than $4/barrel to a five-week high the same day. This is the second round of strikes in three days: on Sunday the US hit Iran's Larak Island and Iranian missiles struck two Jordanian bases, breaking the relative calm since July. The conflict began February 28 when the US and Israel attacked Iran; the June MOU ceasefire collapsed after its 60-day negotiation window expired. Trump threatened "far heavier" strikes if Iran retaliates again; Iranian Speaker Ghalibaf threatened to close Gulf oil exports; Treasury Secretary Bessent previewed new sanctions on Iranian banks within two weeks. Watch: further Iranian attacks on US bases, the sanctions list, Hormuz shipping insurance rates. Reuters
  • US naval blockade has driven Iran's crude exports to near zero — Per Kpler/Vortexa/TankerTrackers data, since the US restarted its blockade on July 14, no Iranian crude tanker has successfully transited the Strait of Hormuz. August Iranian crude + condensate loadings were just 220,000-255,000 bpd, down nearly 90% from ~2 million bpd in March; 29 tankers carrying 36.11 million barrels are stranded inside the strait. The blockade line sits between the Gulf of Oman and the Arabian Sea, with the US inspecting vessels one by one. China is Iran's only major buyer and can now only draw from Asian floating storage without replenishment. Iran's forex cutoff may force money-printing; the IMF estimates Iranian inflation near 70% this year (third-highest globally). Watch: whether China restocks via non-Iranian channels, and whether Iran escalates by threatening other Gulf states' exports. Reuters (via SRN News)
  • Russia's sixth consecutive day of mass strikes on Kyiv kills 12 in a day — Overnight September 1, Russian strikes on Kyiv and surroundings killed 12 and wounded dozens — the sixth straight day of round-the-clock drone-swarm attacks on the capital region; Ukraine's foreign minister called on the West to accelerate air-defense deliveries. Since last week Russia has shifted to new round-the-clock drone-swarm tactics against residential and storage facilities, reflecting a pivot to aerial attrition as its frontline advance slows. The same day, Russia reported Ukrainian drone attacks on its Ust-Luga export facility and Samara region; Modi told Putin "war must end for humanity's sake," and Zelensky spoke with Trump's envoy. Watch: new Western air-defense packages, the tempo of Russia's autumn campaign. Reuters
  • Trump unveils Venezuela oil deal with US government holding 35% equity; oil majors wary — A White House fact sheet released late August 31 shows private oil firm NABEP receiving 100-year leases on 17 Venezuelan oilfields (~65 billion barrels in reserves), with the US government holding 35% of the parent company, a guaranteed 20% of output, and right of first refusal on remaining production. Several major oil companies are hesitant to participate. The deal would take over blocks previously run by Chinese and Russian firms. Producers' concern is competing directly with the US government; UBS argues Big Oil participation is needed to hit output targets. Combined with the Iran blockade, this marks the US government entering the market directly to reshape heavy-crude supply. Watch: potential congressional hearings on legality, opposition reaction in Venezuela. Reuters

No ground-level new policy or regulatory shifts within the 24-hour window for China domestic affairs.

💰 Finance · Markets

Snapshot: S&P 500 7,631.47 (-0.7%, Sep 1 close) · Nasdaq 26,099.77 (-1.0%, Sep 1 close) · US 10Y 4.80% (+4bp, Sep 1) · DXY 99.71 (+0.3%, Sep 1 intraday) · VIX 14.92 (Aug 31; Sep 1 close not available) · Shanghai Composite 3,979.89 (-0.16%) · CSI 300 4,611.44 (-0.30%) · USDCNY mid-rate 6.7809 / onshore ~6.72 (Sep 1) · Brent ~$94.5 (+4.6%); gold not available

  • After Fed Chair Warsh's hawkish Jackson Hole, September hike odds jump to 62%; 10Y yield at 19-month high — Citing CME data via FT: after Warsh's hawkish speech on August 28, September hike odds leapt from 35% to 62%. On September 1 the US 10Y yield rose ~4bp to 4.79-4.80%, the highest since January 2025; the 30Y touched 5.27-5.28%. Global sovereign bonds sold off in sympathy (UK 2Y/10Y each briefly +10bp). The repricing from cuts to hikes is now the dominant variable in global pricing: higher rates directly compress tech/AI growth valuations (Nasdaq -1%, S&P -0.7% on the day) and lift the dollar. Watch: jobs and ISM data before the September FOMC, Warsh's and other officials' language; if hike expectations solidify, higher yields and a stronger dollar will keep pressuring EM assets and gold. Financial Times / CNBC
  • US strikes inside Iran send Brent up 4.6% back above $90 — US Central Command confirmed September 1 strikes on IRGC targets (Bandar Abbas, Qeshm, Jask, Chabahar, and others). Brent rose 4.6% to ~$94.5/barrel; WTI gained ~5% to above $90, a first since July 24. Visible transits through Hormuz fell to ~5 ships/day, and two Saudi VLCCs have been attacked in two weeks. This round covered Iran's export-bypass infrastructure (Jask, Chabahar) for the first time, shifting market pricing from "disrupted transit" to "damaged export capacity" — the geopolitical premium won't fade quickly. Oil +4.6% reinforces the inflation and "higher-for-longer" narrative, resonating with Warsh's hawkishness. Watch: the September 6 OPEC+ JMMC meeting (a production pause is possible), whether Hormuz traffic recovers, Qatari/Omani mediation. AP / Al Jazeera
  • New US Iran sanctions spare major Chinese banks; yuan holds in its strongest zone since 2023 — The US Treasury's new round of Iran sanctions hit 60 entities and vessels but explicitly excluded major Chinese financial institutions. Onshore USD/CNY traded in a 6.7188-6.7264 band this week (the narrowest since early 2016); the September 1 mid-rate was 6.7809, with the yuan up nearly 4% this year and offshore implied volatility near its lowest since 2015. The yuan is showing unusual resilience to geopolitical and sanctions pressure, supported by strong exports, a central bank comfortable with mild appreciation, and a dollar weighed down by fiscal concerns. SEB targets 6.60 by year-end, Standard Chartered 6.65. The external funding environment is friendly for Chinese assets — but watch the tail risk of sanctions extending to major Chinese banks. Watch: mid-rate vs onshore divergence, the CFETS basket index. Bloomberg (via Livemint)
  • A-shares open September split: Shanghai -0.16% while STAR 50 plunges 2.19% as money rotates into consumer and financials — September 1 close: Shanghai Composite 3,979.89 (-0.16%), CSI 300 4,611.44 (-0.30%), SZ Component 13,872.38 (-1.02%), ChiNext -1.32%, STAR 50 -2.19%. Combined turnover ~¥2.6 trillion; beauty/personal care +1.58% and media +1.43% led while electronics fell 2%. 3,300 stocks rose even as all major indices closed red. Index-level moves were mild, but the structure signals a clear growth-to-value/consumer rotation, with STAR/electronics corrections pronounced. If the high-yield-plus-oil-inflation narrative persists, A-share growth valuations face more pressure while dividends and consumer defensives strengthen. Watch: northbound and margin flows, the September domestic easing window (RRR/rate cuts), volume around the 3,950-4,000 band. Sohu Finance

🚀 Tech · AI

  • Anthropic ships Claude Fable 5.1 / Mythos 5.1: one model, two safety tiers, cache reads 75% cheaper — Released September 1. Fable 5.1 is the GA version; Mythos 5.1 is open only to vetted research/security institutions — same underlying model for both. Humanity's Last Exam: 60.9% (no tools) / 65.0% (with tools), roughly double the previous Fable 5 and above Opus 5; +13 points on Terminal-Bench 4.0; 1M-token context, 128K output. Fable's cache-read price drops from $1.00/M to $0.25/M (input $10/M, output $50/M), cutting typical costs ~25% and false-positive safety refusals by up to 60%. This is Anthropic's first flagship-line update since June's Fable 5 (which suffered an 18-day export-control outage, then was undercut on price by July's Opus 5). The technical pitch is long-horizon agents: hours-long runs needing persistent context, tool permissions, checkpointing, and recovery. Enterprise gains Enterprise Frontier Safeguards — monitoring data stays in the customer's cloud with zero-retention support. Mythos 5.1 unlocks high-risk capabilities like software vulnerability discovery, requires 30-day data retention, and is restricted to US institutions in the government trusted-access program — not available in Europe. Analyst reception is mixed: the doubled benchmark gains are respected, but Opus 5 matching it at half price makes Fable's positioning look more like a tiered-pricing story than a pure capability leap. Live on Claude API (claude-fable-5-1), Bedrock, Vertex, and Microsoft Foundry. Decrypt · SiliconANGLE
  • Anthropic signs $35B cloud compute deal with Lambda — second mega contract in a week — WSJ exclusive, September 1: Anthropic signed a $35 billion compute lease with Lambda, an Nvidia-backed cloud provider, with Nvidia holding the datacenter leases. By Jensen Huang's "$80-100B per GW" framing, that's roughly 350-437MW of capacity. It comes less than a week after Anthropic's even larger hardware deal with the UK's Nscale. The latest round of the AI compute arms race: Anthropic is locking in training capacity through a multi-vendor strategy (Lambda, Nscale, AWS, Google Cloud), with an unusual structure — Nvidia itself holding datacenter leases, a chipmaker sinking deep into cloud operations and taking on capital risk. The developer impact is indirect: flagship model capacity supply determines Claude API quotas and pricing. BIS estimates the five hyperscalers will spend over $1 trillion in AI capex across 2025-2026; the opaque financing structures (SPVs, debt, leases) of these long-dated deals are increasingly flagged as a systemic risk. WSJ via SiliconANGLE
  • Google DeepMind launches agentic video understanding: up to 88% fewer video tokens, 66% cheaper — Went live September 1 for Gemini 3.7 Flash / 3.6 Flash / 3.5 Flash-Lite (Gemini API, AI Studio, Gemini Enterprise Agent Platform). Instead of the default 1FPS static frame sampling, the model autonomously decides which segments to watch, at what speed, through which modality (frames/audio/subtitles), pulling clips on demand. Official benchmarks: up to 88% token reduction, 66% cost reduction, +7% accuracy; on the 1H-VideoQA cost-accuracy scatter, Gemini 3.7 Flash dominates the Pareto frontier over GPT-5.6 Sol/Terra, Claude Opus 5, and Grok 4.6. The mechanism is an embedded agentic loop: call a tool to load a video slice, inspect it, then decide whether to keep searching or re-watch at higher frame rates — at the cost of extra reasoning steps versus fixed sampling. Google is extending its Flash-line token-cost strategy to video, the most expensive input modality in the API; long-video workloads benefit most, and the feature will power "Ask YouTube." For developers, video pipeline billing shifts from fixed linear FPS cost to query-dependent cost, making long-video RAG economically viable. Third-party commentary likes the direction but cautions that adaptive scanning's decision-loop latency needs real-world testing. Google Blog
  • Pentagon wires ChatGPT Mil and Grok for Government into GenAI.mil, reaching 3 million personnel — On September 1 the Department of Defense announced that military ChatGPT (OpenAI) and Grok for Government (xAI/Starshield) are joining the GenAI.mil platform, which previously hosted only a military Gemini. The platform has 1.7 million unique users (of 3 million total workforce); OpenAI and xAI each won contracts worth up to $200 million. ChatGPT Mil handles unclassified administrative/policy/logistics documents; Grok targets operational tasks like procurement market research and supply chain management. The context matters: Anthropic, after refusing to allow Claude for all military uses, was labeled a "national security supply chain risk" by DoD last month — and a federal judge ruled last week that the designation was illegal retaliation. The Pentagon is substituting OpenAI/xAI for Anthropic. This is a substantive step change for generative AI in the US defense core workflow: of the three frontier labs, two are in service and one is excluded — vendor selection now directly tracks corporate safety-policy stances. Fortune
  • John Ternus officially becomes Apple CEO as Tim Cook's 15-year run ends — On September 1, John Ternus, 51, the senior VP of hardware engineering, formally took over as Apple's CEO; Tim Cook moved to executive chairman (overseeing Washington and Beijing government affairs). The handover comes days before the September 9 foldable iPhone event; Apple's market cap is roughly $4 trillion. Ternus joined Apple in 2001 and has long led hardware engineering. At his final WWDC keynote in June, Cook disclosed that Apple pays Google about $1B a year to rebuild Siri on Gemini — handing Google an AI distribution surface across 3 billion Apple devices. The challenge is explicit: Apple trails Microsoft, Google, and OpenAI in generative AI, and the new CEO's AI strategy (build vs. buy, on-device vs. cloud) is the market's core test for the Ternus era. TechStartups

🧠 Voices · Thinkers & Research

  • Elon Musk: AI datacenters' power gap is the crisis; EU regulation is dragging down innovation — Speaking at the G20 innovation ministerial in Chapel Hill, Musk said power is the only hard constraint on AI growth, citing consensus estimates of at least a 15GW AI-chip power gap before 2027. The same day he criticized EU tech regulation as "dragging down enterprise progress," urged light-touch rules, and allowed that datacenter builders should pay fair taxes while the industry urgently needs capacity. Washington has framed its G20 host year around limiting regulatory burden. Musk's comments land amid rising US domestic backlash against AI datacenter construction (Trump said Monday that opponents would end up "poor and behind"). The quantified 15GW warning adds fresh ammunition to the "AI power crunch" narrative. Jensen Huang and Sam Altman are scheduled for fireside chats with Commerce Secretary Lutnick at the same venue on September 2. TechXplore
  • Dario Amodei: warns of an "AI bubble," says AGI is mostly scaling-driven, urges the US to stop selling advanced chips to China — Amodei said the technology itself is solid but an "AI bubble" risk exists, with massive investment vulnerable to "timing errors." He reiterated that AGI will arrive mainly through model scaling, continues to advocate strong AI regulation, and urged the US to stop selling advanced AI chips to China on national-security grounds. This is part of Amodei's ongoing commentary ahead of Anthropic's reported IPO (valuation rumors near $2 trillion); he has previously likened selling chips to China to "selling nuclear weapons to North Korea." His bubble warning echoes Ray Dalio's recent 1929/2000 analogies and Goldman's "profits bubble" thesis — but Amodei stresses the disagreement is about investment timing, not the underlying technology. TechShots

📝 New Blogs (24h)

  • Simon Willison · Codex bundles LibreOffice — Digging through ~/.cache, Willison found that OpenAI's Codex desktop app (now rebranded ChatGPT) ships a 1.7GB codex-primary-runtime: a full Python install, full Node.js, and native binaries for Poppler, git, and LibreOffice — plus skills that teach Codex how to find and invoke those binaries for document work. The takeaway: coding-agent vendors are bundling entire desktop software stacks into agent runtimes, guaranteeing document capability via bundling rather than relying on the user's environment. That mirrors the skills-ecosystem philosophy of "tools plus instructions shipped together" — worth tracking how each vendor's local runtime evolves. Read
  • Hugging Face (Ai2) · BenchMIRT: What are LLM benchmarks actually measuring? — Ai2 introduced BenchMIRT, a method for auditing LLM benchmarks at the level of individual prompts: within a single benchmark, different groups of tasks actually measure different things (BBQ's age-bias items also depend on coreference tracking), so benchmark scores blend signals beyond the stated target capability. Directly valuable for any engineering team making model-selection decisions off benchmark numbers: audit the benchmark before comparing models, and you'll meaningfully cut the "high score ≠ the capability you need" misjudgment. Paper, data, and code are all open. Read
  • Cloudflare · How we could save petabytes of cache storage with Zstandard and Pingora — Cloudflare's engineering team prototyped compression transcoding inside the cache layer: storing cache objects Zstandard-compressed could save petabytes on the same hardware. The value is the architectural trade-offs of "compress at rest, transparently decompress on hit" and the hit-rate/cost calculus — rare real-world data for anyone building their own caching or storage layers. Read
  • Hugging Face · @huggingface/kernels: 200+ WebGPU Kernels for Local AI — HF's WebAI team released @huggingface/kernels, a minimal library for loading and running optimized WebGPU kernels from the Hub, with an initial collection of 207 kernels covering operators used across mainstream ML architectures. This is a foundational piece of the browser-side local inference stack — after model formats and runtime execution plans, the lowest-level GPU operators are now standardized too, another step toward WebAI's "write once, run everywhere" goal. Read

⭐ GitHub Trending

  • THU-MAIC/OpenMAIC (TypeScript · ⭐29,420 · +3,122 today) — Tsinghua team's Open Multi-Agent Interactive Classroom: one-click immersive multi-agent learning experiences. Today's biggest gainer across all of GitHub. Multi-agent orchestration meets education is a clear new direction signal — worth studying its role-scheduling and classroom-choreography design if you build agent platforms. repo
  • jingyaogong/minimind (Python · ⭐57,019 · +1,005 today) — Train a 64M-parameter LLM from scratch in just 2 hours. A long-popular teaching project surging again today; a few thousand lines covering the entire pipeline from tokenizer to RLHF — one of the fastest open-source paths to understanding the full LLM stack. repo
  • browser-use/video-use (Python · ⭐22,899 · +509 today) — Edit videos with coding agents. The browser-use team porting its agent capabilities to video editing: natural-language instructions, agent-driven timeline operations. Landing the same day as DeepMind's agentic video understanding, it shows the agent-tooling frontier expanding into multimedia from both ends — understanding and editing. repo
  • firecrawl/pdf-inspector (Rust · ⭐17,896 · +545 today) — Fast Rust library for PDF inspection, classification, and text extraction; intelligently detects scanned vs text-based PDFs for routing. The most annoying part of document pipelines — scanned-page detection — is now a standalone high-performance library; RAG ingestion teams can adopt it directly. repo
  • K-Dense-AI/scientific-agent-skills (Python · ⭐41,509 · +914 today) — Turn any AI agent into an AI Scientist: 165 validated research skills, claiming 190,000+ scientist users. A landmark for agent-skills ecosystems going deep into verticals (science), tightly coupled to Claude's skills mechanism — its skill-validation pipeline is worth studying. repo

(Also trending, 5 minor: 3b1b/manim +74, iv-org/invidious +583, Gitlawb/openclaude +37, handsomestWei/patent-disclosure-skill +502, Imbad0202/academic-research-skills +161.)

References