AI News Briefing

The Big Picture

Sunday feels like the pause between Thursday’s OpenAI launch aftershocks and Friday’s Google counter. Gemini 3.5 Pro is now five days out per a leak circulating since Friday — 2M-token context (double any current frontier), Deep Think reasoning gated behind the $250-a-month Ultra tier, full rebuild from scratch rather than a 2.5 Pro adaptation, pricing rumored at $12–$15 input / $36–$45 output per Mtok. Meanwhile Anthropic’s Fable 5 subscription-inclusion window slams shut tonight at 11:59:59 PM PT — after two extensions, Pro/Max/Team subscribers move to $10/$50 pre-purchased credits or lose access. Zhipu breaks the frame that both American labs are trying to lock in: Tang Jie’s staff memo, reviewed by Bloomberg Sunday, argues that “genuine security stems from broad participation, sharing, and oversight rather than technological barriers,” and commits Zhipu to zero short-term monetization for the next two years. OpenAI’s CDC proof grew its second layer of scaffolding — the openai/cdc-lean repo went public with a Lean-verified reduction from the 8-flow theorem to CDC for finite loopless bridgeless multigraphs — but the human-readable arguments in the PDF that Hacker News flagged Saturday as inscrutable still haven’t been peer-reviewed. Two American labs racing toward closed frontier, one Chinese lab publicly refusing. The most interesting move today came from Beijing.

Zhipu founder Tang Jie backs open-source frontier — memo commits to zero monetization for two years

Bloomberg reported Sunday July 12 on an internal staff memo from Zhipu co-founder and CEO Tang Jie arguing that frontier AI should remain broadly accessible rather than controlled by select individuals or firms. “Genuine security stems from broad participation, sharing, and oversight rather than technological barriers,” Tang wrote — a direct rebuttal of the “restricted frontier” framing OpenAI and Anthropic have been pitching to the Trump administration since June. Tang, also a professor at Tsinghua, said Zhipu will not pursue short-term monetization of AI applications for the next two years, focusing instead on long-horizon tasks, autonomous agents, and fully self-training AI models. Zhipu already open-sourced its cutting-edge GLM-5.2 (MIT license) on June 17, and the model currently sits at #1 on the Artificial Analysis Intelligence Index among open weights at 51.

Bloomberg’s framing: Zhipu is positioning itself as the counter to the closed-frontier consensus. Tang’s substantive tell — committing publicly to zero monetization for two years while Anthropic tightens Fable 5 subscription access tonight and OpenAI files a $730B IPO for September — is the first Chinese frontier CEO to publicly reject the American labs’ “auditable and controllable” pitch to enterprise buyers. Structural: same week Bengio’s UN panel called concentration of AI decision-making power the central problem, and Fed chair Warsh handed Andreessen an AI-productivity-and-jobs task-force seat, Zhipu just gave the multilateral track its highest-profile lab-side spokesperson. The open-source vs closed-frontier fight has a face this quarter that isn’t Delangue or LeCun.

Anthropic Fable 5 subscription-included access ends tonight — third and final cutoff at 11:59 PM PT

Anthropic’s Fable 5 subscription-inclusion window closes tonight Sunday July 12 at 11:59:59 PM PT (2:59:59 AM ET Monday) — the third cutoff after two extensions from the original July 7 date. From tomorrow, Fable 5 no longer counts toward any Pro/Max/Team/select-Enterprise plan’s weekly included usage limit; continued access requires a pre-purchased usage-credit balance at $10 per million input tokens and $50 per million output tokens (roughly double Opus 4.8). Anthropic frames the credits switch as temporary “while building capacity” but has named no restore date. The credit gate lands 48 hours before the leaked Gemini 3.5 Pro launch (July 17) and one week after OpenAI shipped GPT-5.6 Sol at $5/$30 — exactly half Fable 5’s credit price.

Android Authority’s read: subscribers spent the weekend burning credits before the deadline. Substantive tell for Patrick: this is when Claude’s coding tier stops being “included for Max” on the frontier model and starts being “$10/$50 credits or downgrade to Opus 4.8.” Structural: OpenAI’s Sol shipped at half Fable 5’s credit price four days ago and Google’s Gemini 3.5 Pro is five days out at rumored $12–$15/$36–$45. Anthropic can’t credibly hold $10/$50 as the frontier price shelf when both competitors will be at or below Sol pricing by end of week. The “capacity” framing buys time; the price shelf itself is the tell that Anthropic knows it.

Gemini 3.5 Pro leak targets July 17 launch — 2M context, Deep Think gated behind $250 Ultra tier

A launch-plan leak circulating on X since Friday July 10 (initial post: @Mr_Salio, amplified by @pankajkumar_dev and @LuminaXspace) has Google DeepMind targeting Friday July 17 for Gemini 3.5 Pro general availability. The specs per the leak: a 2-million-token context window (double the 1M ceiling every current frontier model tops out at), a new Deep Think extended-reasoning mode gated behind the $250-a-month Gemini Ultra subscription, significantly upgraded UI and SVG/frontend generation, and autonomous-agent workflow support. Reported to be a full pretraining rebuild rather than a 2.5 Pro adaptation — per multiple accounts the previous 3.5 base model was scrapped in April to fix mathematical-reasoning gaps. API pricing rumored at $12–$15 per million input tokens and $36–$45 per million output tokens. Google has not officially confirmed the date, specs, or pricing.

Coursiv’s caveat: every claim about launch date, context window, or scrapped base model should be treated as unconfirmed until Google publishes a model card. Substantive tell if the leak holds: 2M context is the first genuine step-change in frontier context size since Anthropic hit 1M with Sonnet 4.6 in February — every other frontier launch this year has landed at or below 1M. Structural: Google spent Thursday’s OpenAI launch cycle quiet, then landed AlphaEvolve GA the same day, and is now four business days out from a launch that would land the same week Fable 5 subscription access closes and Bengio’s UN panel presents interim findings. If it lands as leaked, the “buyer picks a vertical” frame the morning Big Picture called out July 9 gets a fourth active frontier tier this month.

OpenAI publishes Lean formalization for CDC proof — verifies reduction step, not the full argument

OpenAI published github.com/openai/cdc-lean alongside Friday’s Cycle Double Cover proof PDF — a Lean 4 formalization (pinned Lean v4.31.0, Mathlib at commit 9a9483a) that kernel-checks an unconditional CDC theorem for finite loopless bridgeless multigraphs, endpoint CDCLean.cycleDoubleCover_of_bridgeless in CDCLean/Main.lean. The formalization uses the formalized Jaeger–Kilpatrick eight-flow theorem to construct a nowhere-zero Γ = (Z/2)³ flow, then converts the flow into a cycle double cover. The Lean part covers the reduction step for that graph class; the human-readable arguments in the accompanying PDF — which Hacker News commenters Saturday flagged as unusually inscrutable, cataloguing prior CDC “proofs” that later collapsed — remain unverified by Lean and unrefereed. Wolfram MathWorld and Wikipedia both updated Sunday to note the AI-generated claim without endorsing it.

MathWorld’s framing: the claim is “fresh enough that it should be read as a claim, not a settled result.” Substantive tell: a Lean formalization for the reduction step is real evidence, not marketing — it removes one class of failure mode (a bug in the flow-to-cycle conversion) from the “might collapse under review” risk list. It does not remove the biggest one: whether the human-authored arguments in the PDF that reduce arbitrary bridgeless graphs to the “finite loopless multigraph” class Lean handles are actually correct. Structural: yesterday’s briefing said “no Lean formal verification was used” — that was technically wrong and is worth correcting. The Lean piece exists, it’s public, it type-checks. The peer-review question is now specifically about the pre-reduction pages, which is a narrower and answerable question. Graph theorists have days to weeks of stress-testing ahead.