DeAI Daily Brief — 10 October 2026

Today in DeAI: Qwen ships an 8-step Turbo checkpoint of its 7B image model, Anthropic launches a free open-source vulnerability scanner, Asana claims a 76x agent cost cut, OpenAI's revenue is re-reported lower, and Phala expands zero-retention serving on OpenRouter.

Qwen-Image-2.1-Turbo: the same 7B image model in 8 denoising steps

Alibaba's Qwen team released an accelerated checkpoint of Qwen-Image-2.1, its open-weight 7B image model, that finishes text-to-image and editing in 8 denoising steps instead of the base model's 40-step default, with the sampling schedule baked in and CFG=1 by default. Comfy-Org packaged it for ComfyUI within hours in full BF16, INT8, and LoRA forms, and Alibaba's hosted Pro and Turbo APIs went live the same day. Why it matters: a directly verifiable 5x cut in per-image compute on the same hardware, though the weights still carry the September research-only licence, so commercial use still needs the hosted route or a negotiated licence. (Hugging Face model card) — our full coverage

Anthropic launches Cyber Mission with a free OSS vulnerability scanner

Anthropic announced Cyber Mission, a security program pairing with CrowdStrike and Palo Alto Networks around critical-infrastructure defense, and shipped a free AI scanner that automatically checks open-source projects for vulnerabilities, with an expected accuracy above 90 percent that is Anthropic's own expectation rather than an audited result. The same day, OpenAI published a Sophos case study claiming its Daybreak system cut cyber-threat investigation time by 96 percent and automates 52 percent of managed-detection cases, both vendor self-reports. Why it matters: if the scanner is genuinely free to OSS maintainers, it is a new no-cost signal on the repositories your inference stack depends on, and an early test of whether AI-vendor scanners produce real findings or noise. (The Decoder)

Asana claims GPT-6.1 Sol made its browser agent 76x cheaper and 5x faster

OpenAI published a case study stating that Asana rebuilt its browser agent on GPT-6.1 Sol via Codex and saw it become 76x cheaper and 5x faster in internal tests. Both figures are a vendor-published customer self-report from a company that sells the model being praised, and the 76x folds together model choice, caching, and architecture changes with no published methodology. Why it matters: if anything like that ratio held at list prices on a like-for-like workload, it would reset cost assumptions for agentic browsing, which is exactly why the missing test methodology matters more than the number. (OpenAI)

OpenAI revenue re-reported at ~$50B as the company seeks $30B+

The Decoder reports OpenAI's annualized revenue rate now sits around $50 billion, well below an initially circulated ~$70 billion figure attributed to a different accounting method, a correction that reportedly sent chip stocks sliding while the company negotiates at least $30 billion in fresh capital at a reported $1.4 trillion valuation. All figures are press-reported and unverified. Why it matters: the gap between reported revenue and reported capital need is the pressure behind aggressive API pricing moves, so every list-price cut or enterprise discount in this cycle should be read against that balance sheet. (The Decoder)

Phala serves ~20 open models from TEE-backed zero-retention endpoints on OpenRouter

OpenRouter's zero-data-retention documentation now lists around 20 Phala endpoints covering models including DeepSeek V4, served from trusted-execution-environment-backed infrastructure where prompts are not retained. The TEE claim is Phala's architecture statement, not an independently audited attestation, and OpenRouter's per-provider logging documentation remains the verification path. Why it matters: zero-retention routing through a mainstream aggregator moves private inference from a specialist purchase to a routing option, but the privacy property is only as good as the attestation behind it. (OpenRouter ZDR docs)

Liquid AI open-sources d1 decision models with native llama.cpp support

Liquid AI released d1-3B and d1-omni-600M, small open-weight decision models that emit structured judgments rather than long text, with native llama.cpp support for edge deployment. Why it matters: routing decisions, filtering, and tool gating do not need a 100B-class model, and small open decision models running locally are the cheapest way to keep those steps off hosted APIs. (Liquid AI)

Watching tomorrow

Whether independent testing of the 8-step Qwen-Image-2.1-Turbo checkpoint surfaces quality comparisons against the 40-step base model, and whether the OpenAI safety-researcher firings pick up wire verification beyond The Decoder's report.