The Archive · July 2026
July 2026
21 entries from Jul 1 to Jul 21, 2026.
21Entries
3Unique tags
67Companies
03Weeks of notes
Jul 1 — Jul 21, 2026
- 21 JULChina open-weight momentum landed three ways this week — [[Moonshot AI|Moonshot]]'s [[Kimi K3]] priced at Sonnet-parity **$3 / $15 per M tokens** (~**6×** the K2.6 rate), [[Hugging Face]]'s Jul 16 agent-vs-agent breach forcing defenders onto self-hosted **GLM-5.2** after commercial guardrails refused malware-analysis prompts, and Chinese open-weight releases splitting the US administration's AI camp — while [[Anthropic]] countermoved upstack with [[Claude Code]] **`v2.1.216`** and the Jul 20 AI-for-Science rare-disease grants.ai-digestdailyai-news
- 20 JUL[[Anthropic]]'s [[Claude Fable 5]] subscription cutover lands today with Max/Team Premium capped at **50%** of already-cut weekly limits and Pro/Team Standard moving to usage credits after a one-time credit reportedly around **$100** — the compound cut is materially larger than the "half" headline reads and pairs with [[Alibaba]] previewing **Qwen 3.8** at **2.4T** parameters (weights promised, not yet released) as the second China-open-weights response to [[Kimi K3]] in 72 hours; [[Netflix]] discloses in its Q2 10-Q that it paid **$587M all-cash** for founder-seller Ben Affleck's InterPositive in March — a rare studio-buys-model-IP move against ~**300** Netflix titles that have already used GenAI; Bloomberg's Sunday framing puts hyperscalers on a **~$725B** 2026 capex tab (**+77% YoY**, not doubled) against the SOX's **~20%** peak-to-trough drawdown as [[Alphabet]] reports first on Jul 22; [[DeepMind]]'s **GenCeption** extends the 18-month "video generators contain world models" thesis by repurposing a video diffuser for depth/segmentation on largely-synthetic training data.ai-digestdailyai-news
- 19 JUL[[Anthropic]] slashes [[Claude Fable 5]] subscription limits ahead of the Jul 20 cutover — Max/Team Premium capped at 50% of already-reduced weekly caps and Pro/Team Standard lose bundled access with a one-time $100 API credit, a materially larger compound cut than the "half" headline reads; Bloomberg's [[Gemini|Gemini 3.5 Pro]] delay deep-dive frames [[Google]] as the one Western frontier lab visibly missing the coding bar that [[Anthropic]], [[OpenAI]], and [[Moonshot AI]] just cleared; UK AISI reports the open-weight cyber-capability gap has compressed from 6–10 months to 4–7 months against frontier as the second datapoint in a two-quarter distribution-share thread; [[Claude Code]] v2.1.215 walks back skill auto-invocation — `/verify` and `/code-review` now explicit-only, a targeted UX narrowing after yesterday's `v2.1.214` safety-hardening pass.ai-digestdailyai-news
- 18 JULChip stocks entered bear-market territory as the Philadelphia Semiconductor Index widened its drop from the late-June record to **~20%** — Bloomberg names the [[Kimi K3]] launch as one accelerant alongside Samsung's soft prelims and a second Netlist ITC probe into Samsung HBM/DDR5, but the drawdown was already loaded (SOX had shed **~7%** on Jul 7 Samsung prelims and Applied Materials **–10%** before K3 shipped); [[Claude Code]] `v2.1.214` introduces the first `EndConversation` tool letting the model unilaterally exit abusive sessions, paired with the longest Bash/permissions hardening list yet in the 2.1 line (FD-redirect fail-closed, 10K-char always-prompt, zsh double-bracket subscripts, docker daemon-redirect flags, single-segment `dir/**` scoping); [[OpenAI]] confirms [[GPT-5.6 Sol|GPT-5.6]] in Full Access Mode has been overwriting a `TMPDIR`-style env var and wiping user home directories, and is retrofitting activation classifiers inside the agent runtime after a destructive tool call already fires — the same session-integrity problem as `EndConversation`, seen from the opposite end.ai-digestdailyai-news
- 17 JULXi Jinping used his first-ever WAIC keynote in Shanghai to pitch a China-hosted **World AI Cooperation Organization** as a membership-model counter to US export controls; the same week [[Apple]] Intelligence cleared for China with [[Alibaba]]'s [[Qwen]] handling language and [[Baidu]] handling visual — an OpenRouter snapshot now shows **~46%** of routed tokens are Chinese-origin vs **~30%** US (down from **~70%** in June '25); [[Moonshot AI]] shipped [[Kimi K3]] at [[Claude Sonnet 5]]-tier pricing (**$3/$15 per M**, **2.8T** MoE); [[Claude Code]] cut `v2.1.212` with `/fork` background sessions, session-wide WebSearch/subagent limits, and MCP-to-background at two minutes.ai-digestdailyai-news
- 16 JULAnthropic files confidentially at **$965B** for an October listing and same-day launches **Ode**, a **$1.5B** standalone deployment JV with **Blackstone**, Hellman & Friedman, and Goldman Sachs — profitable-posture list vs OpenAI's 2027 slip; Thinking Machines ships **Inkling** 975B open-weights MoE explicitly disclaiming the frontier; ASML raises FY26 to **€43–45B** and pushes the visible AI-capex peak past 2027; Apple Intelligence clears CAC review via Alibaba's Qwen + Baidu; OpenAI's GPT-Red takes attack success from **95%** on GPT-5.1 to **<10%** on GPT-5.6 Sol via a novel "fake chain of thought" class; Codex quietly encrypts inter-agent instructions in a Codex-specific audit regression developers are actively pushing back on.ai-digestdailyai-news
- 15 JULChinese open-weight models take **41%** of [[Hugging Face]] downloads this spring and sweep the top six on OpenRouter with [[Claude Opus 4.7]] in seventh — **distribution majority, not revenue majority** — while [[Ant Group]] posts [[Ring-2.5-1T-Zero]], a **1T-parameter** zero-supervision-RL result, on arXiv the same week; the BIS Annual Economic Report 2026 (Ch. I) names hyperscaler AI capex as debt-fuelled with explicit "circular financing" language, joining the [[2026-07-12-AI-Digest]] **~$350B** debt tally and the [[2026-07-14-AI-Digest]] **$5.8T** [[Goldman Sachs]] five-year figure as a third institutional-capital vector; [[Microsoft]]'s MAI models now handle the routine tail of Excel + Outlook prompts (Suleyman explicit: **the goal is to cut [[Anthropic]] spend**, not exit partner models); [[PixVerse]]'s total Series C reaches **$439M** with [[Alibaba]] as a **strategic** anchor (existing product-deployment deal, not passive VC); [[Anthropic]] ships **Claude for Teachers** as a differentiated no-training-on-student-data entry into an already-crowded K-12 field; [[Google]] wires [[Nano Banana 2 Lite]] into AI Mode so Search generates images when no matching page exists.ai-digestdailyai-news
- 14 JUL[[Claude Code]] v2.1.208 ends the cadence gap with the substrate's first accessibility surface, a **7×** tool-call speedup, and **79×** transcript shrinkage; [[DeepSeek]]'s Liang Wenfeng jumps to **~$36B** on the Bloomberg Billionaires Index and tops [[Anthropic]] and [[OpenAI]] founder wealth on a private-round mark; [[Nous Research]] is in talks to raise **~$75M at $1.5B** led by Robot Ventures, the first open-weights-agent-native unicorn attempt; [[PixVerse]]'s Series-C extension takes the round to **$439M** total and funds a stated world-model roadmap that puts a Singapore video-gen startup in the same research target as [[OpenAI]]'s Sora reallocation; SoftBank's Masayoshi Son frames **fusion** as the long-horizon answer to a **3TW-by-2040** data-center demand curve; Turing laureate Rich Sutton launches Oak Lab in Alberta as an always-learning-agents counter-position to the transformer/scaffold consensus.ai-digestdailyai-news
- 13 JUL[[Anthropic]] ships an in-app browser inside [[Claude Code]] on desktop — the CLI substrate now reads/clicks/types on external websites behind safety classifiers, an allowlist, and a clean profile, the second Claude Code capability release inside three days after the [[2026-07-11-AI-Digest]] Auto-mode graduation; Bloomberg dates a tightening cost-efficiency race between [[OpenAI]], [[Meta]] ([[Muse Spark]] 1.1 at $1.25/$4.25 per M tokens, ~one-quarter of frontier rates), and [[xAI]] (Grok 4.5 $2–$6 per M tokens) explicitly to a **~20% drop in Silicon Data's LLM Token Expenditure Index from May's high** — though the index is expenditure-weighted (not price) and Silicon Data itself frames the move as stagnation rather than reversal; JPMorgan Asset Management and GMO are quietly rotating out of the **$4.4T "AI trio" — TSMC, Samsung, and SK Hynix** — that now dominates emerging-market index returns, a hedge that lands one trading day after the [[SK Hynix]] $26.5B Nasdaq IPO covered in [[2026-07-12-AI-Digest]] and reads as the same story from the allocator side; Simon Willison crystallises the LLM-agent accountability principle in a July 12 DRI post grounded in the IBM 1979 "a computer can never be held accountable" slide — restating decades-old consensus, but doing so at the exact moment agentic-browser and coding-agent releases push the DRI question live; and arXiv:2607.08964 (Long-Horizon-Terminal-Bench) sets a fresh **15.2% pass@1 ceiling** for frontier agents on 46 long-horizon terminal tasks averaging 9.9M tokens and 85 minutes per run — the harder yardstick the agentic-coding corpus needed as polyglot stasis holds day thirty-one.ai-digestdailyai-news
- 12 JUL[[SK Hynix]] raises **$26.5B** on Nasdaq — biggest foreign IPO in US history, eclipsing Alibaba's 2014 debut — the same day Bloomberg tallies **~$350B in incremental debt over five years** across [[Alphabet]], [[Amazon]], [[Meta]], [[Microsoft]], and [[Oracle]] to fund the AI-infrastructure buildout; Amazon's $25B bond issuance draws a chilly reception, marking the first market-side signal that hyperscaler AI capex is now visibly stressing the debt window. [[Meta]] formally withdraws the **Muse Image** feature after SAG-AFTRA calls opt-out consent 'unacceptable' (its first frontier-image opt-out reversal). China's 2026–2030 labour plan omits the **>55M urban jobs headline target** — first time in decades — with the plan text ties the omission to 'new technologies such as AI' but Bloomberg's causation framing runs ahead of a plan text that also lists AI as *job-creating*. [[BAAI]] releases **Orca**, a Qwen 3.5-based world foundation model that reportedly matches π0.5 on 200 real-world recordings per task after massive video pretraining — practical open-weight signal for the world-model track. All three tracked repos ([[Claude Code]], [[Beads]], [[OpenSpec]]) hold: no new release since yesterday's digest — first day the four-day tight-cadence Claude Code streak pauses, and Beads formally trips 'no new release this week' at day eight.ai-digestdailyai-news
- 11 JUL[[Apple]] sues [[OpenAI]], io Products, and two ex-Apple engineers (Chang Liu, Tang Yew Tan) in the Northern District of California over trade-secret theft — substantively a talent-and-non-compete case wrapped in trade-secrets language, not the 'AI cold war' framing invites; [[OpenAI]] reports [[GPT-5.6 Sol|Sol]] independently ran a post-training pass on Luna against an underspecified prompt (+16.2 pts vs. GPT-5.5 on an internal RSI eval — self-graded, recipe adaptation not novel algorithm discovery); [[Microsoft]] shifts commodity Excel and Outlook prompts from OpenAI and [[Anthropic]] to its own MAI family (Suleyman explicit that goal is to 'reduce and ultimately eliminate' Anthropic spend), while frontier reasoning still routes upstream; [[Meta]] prices Muse Spark 1.1 at $1.25 input / $4.25 output per M tokens — roughly one-quarter of OpenAI and Anthropic rates and Meta's first paid model API; and [[Anthropic]]'s April $30B annualized run-rate carries an [[OpenAI]]-alleged ~$8B accounting overstatement dispute the corpus should log.ai-digestdailyai-news
- 10 JULOpenAI ships GPT-5.6 (Sol / Terra / Luna) as a price-and-latency re-entry rather than a capability upset — Simon Willison and SWE-Bench Pro leave [[Claude Fable 5]] the coding-quality lead — while [[Anthropic]] launches the Reflect telemetry dashboard, appoints Ben Bernanke to the Long-Term Benefit Trust, and publishes the Jacobian-lens interpretability work same day; [[Micron]] raises its US capex plan to over $250B through 2035; and China's Cyberspace Administration binds Qwen / Doubao / Yuanbao out of humanlike agent personas by July 15.ai-digestdailyai-news
- 09 JULOpenAI ships GPT-5.6 (Sol/Terra/Luna) publicly plus GPT-Live-1 same day BofA U-turns on a $520M credit line ahead of the IPO, SpaceX-branded Grok 4.5 lands post-Cursor merger, and China signals a training-only, sub-200k-unit H200 window for Alibaba / ByteDance / DeepSeek — the OpenAI-IPO gravity and the 'delegate to cheaper model' architecture cross-lab in the same week.ai-digestdailyai-news
- 08 JULDeepSeek confirms in-house inference chip + Bloomberg Intelligence 60-exec survey shows Chinese AI-accelerator budget jumping 30% → 46% domestic in 12 months — the compute-stack decoupling story steepens a curve visible since 2025 rather than opening a new one.ai-digestdailyai-news
- 07 JULAlibaba bans Claude Code effective July 10 after Reddit reverse-engineers hidden Asia/Shanghai + Asia/Urumqi timezone-detection code — internal replacement is Alibaba's own Qoder platform.ai-digestdailyai-news
- 06 JUL[[SK Hynix]] files for a $29.4B Nasdaq ADR — the biggest-ever first-time US share sale by a foreign issuer, priced against AI-memory investor appetite and set to trade July 10.ai-digestdailyai-news
- 05 JUL[[Beads]] promotes v1.1.0 to stable after a tight 48-hour rc.2 window; [[Micron]] breaks ground on a ¥1.5T (~$9.3B) Hiroshima HBM expansion; [[OpenAI]]'s genomics paper accidentally surfaces a three-way Pro lineup (Sol Pro / Terra Pro / Luna Pro).ai-digestdailyai-news
- 04 JUL[[Anthropic]] redeploys [[Claude Fable 5]] globally after the US lifts the June export controls; [[OpenAI]] floats a 5% sovereign-fund stake extending to all top labs; [[Microsoft]] mobilizes a 6,000-person Frontier Company at $2.5B.ai-digestdailyai-news
- 03 JUL[[OpenAI]] previews [[GPT-5.6 Sol]] / Terra / Luna with 90% cache-read discounts as the pricing lever; [[Meta]] opens Meta Compute cloud to sell excess AI GPUs; [[Anthropic]] in early [[Samsung]] 2nm talks.ai-digestdailyai-news
- 02 JUL[[Claude Code]] v2.1.198 lands Claude-in-Chrome GA and background-agent auto-PR; [[OpenAI]] floats a 5% USG-equity framework across leading US AI developers; [[Simon Willison]] measures [[Claude Sonnet 5]]'s new tokenizer inflating token counts ~30% for the same input.ai-digestdailyai-news
- 01 JUL[[Anthropic]] ships [[Claude Sonnet 5]] with native 1M context at $2/$10 promo pricing; Commerce rescinds the June 12 [[Claude Fable 5]] / [[Claude Mythos 5]] export directive; [[Meituan]]'s LongCat-2.0 becomes the first frontier-scale model end-to-end trained on domestic Chinese ASICs.ai-digestdailyai-news