DeAI Daily Brief — 2 October 2026

Today in DeAI: zkAPI takes API payment identities off the table on Ethereum mainnet, Cursor puts Z.ai's GLM 5.3 family in its default model picker, Vercel's index logs open weights at 56% of August token volume, OpenAI claims a 15,000-account distillation campaign, and Cloudflare ships its own open-weight decision models.

zkAPI: pay for AI APIs without an account

The Ethereum Foundation and the Open Anonymity Project launched zkAPI on October 1, a protocol where a one-time deposit into an on-chain vault replaces the API key and its billing identity: per-session zero-knowledge proofs mint short-lived, dollar-capped keys, so the provider sees prompts but never who pays. The vault contract is live on mainnet holding USDC, with a public repo and a working browser client. The blog itself flags the unsolved parts: no network anonymity by default, and content-based re-linking through writing style or reused history. Why it matters: it removes the billing link between you and years of your prompts — the layer that policy-based "privacy" never touched. (Ethereum Foundation) — our full coverage

Cursor ships GLM 5.3 as a first-party model

Cursor's docs now list Z.ai's open-weight GLM 5.3 as a first-party "Other Models" pool entry — 1M context, $1.40 input and $4.40 output per million tokens, billed at GLM 5.2's rates — alongside a GLM 5.3 Flash companion page. Cursor's own CursorBench 4.0 board ranks GLM 5.3 Max 20th of 63 at 42.6% ($5.05/task), the top open-weight row, one place above GPT-5.6 Sol Max and far under Opus 5.5 Max's 57.8% at $13.43/task. Why it matters: that ranking is Cursor grading models on Cursor's own benchmark — checkable, but the tasks are Cursor's — and the distribution fact is the real story: open-weight models are now the default menu in mainstream coding tooling, not a bring-your-own-key afterthought.

Vercel index: open weights carried 56% of August tokens on 14% of spend

Vercel's September AI Gateway Production Index reports open-weight models rose from 7% of token volume in December to 56% in August — the first majority — while accounting for 14% of spend, with Anthropic still capturing roughly two-thirds of gateway dollars. Why it matters: one router's measurement, not the market, but the volume-versus-revenue inversion quantifies exactly which work builders route to cheap weights and which still pay frontier prices. (Vercel) — related: our July index coverage

OpenAI claims 15,000-account distillation campaign disrupted

Per The Decoder's October 1 report, OpenAI says it stopped a coordinated effort by 15,000+ accounts to extract its models' hidden reasoning, tying part of the activity to people connected to Moonshot AI — OpenAI's characterization of its own enforcement action. Researchers add that the same technique kept working on Microsoft Azure for weeks, including against GPT-6 Astra. Why it matters: enforcement boundaries apparently stop at the first-party endpoint; the identical model sold through a hyperscaler may not carry the same protections, which is a where-to-run variable that has nothing to do with price.

Cloudflare ships Clef: open-weight decision models

Cloudflare released Clef (27B) and Clef-flash (9B), Apache-2.0 decision models post-trained from Qwen backbones that return typed probabilities instead of text, hosted on Workers AI with weights on Hugging Face. Per The Register, Clef prices at $0.24/M tokens — about six times TypeSafe's Jev at $0.042/M — while Cloudflare's own index frames it as more accurate. Why it matters: the decision-model category now has a hyperscaler entrant whose open weights undercut its own hosting price, which makes the weights, not the platform, the interesting artifact.

Watching tomorrow

Whether OpenAI publishes a primary disclosure behind the distillation-disruption claim, and whether any inference provider announces native zk-proof settlement after the zkAPI launch.