AI Weekly — 2026-09-25 to 2026-10-02 | Frontier models, China silicon, and a $20.40 exploit

Two frontier releases, a chip-stack realignment, and a security disclosure that puts an exploit chain at $20.40 — the week OpenAI disclosed a model-distillation takedown and Anthropic published a $20.40 cyber-exploit claim, and the supply-chain story and the threat-model story stopped being separate.

Frontier model releases

Google unveiled Gemini 4 Argon this week, described as the company's "next era of frontier intelligence"Gemini 4 Argon: our next era of frontier intelligence - blog.google and confirmed by a separate report as a new frontier modelGoogle DeepMind unveils Gemini 4 Argon, its new frontier model - Dealroom. OpenAI countered with a preview of GPT-5.6 Sol, framed as a "next-generation model"Previewing GPT-5.6 Sol: a next-generation model - OpenAI. The vendor positioning — "frontier," "next-generation" — is the language of press releases, not the language of evaluation. Treat both as announced until independent benchmarks land; the gap between announced and commercially usable is exactly where most integration cost hides.

China silicon and the DeepSeek–Huawei stack

The week's most concrete infrastructure shift came from DeepSeek. The vendor unveiled software intended to run on Huawei AI chips, positioned as a challenge to NvidiaDeepSeek unveils Huawei AI chip software to challenge Nvidia - Seeking Alpha. Coverage describes a stronger software integration with Huawei GPUsDeepSeek Announces a Stronger Software Integration with Huawei GPUs - analyticsindiamag.com and characterizes the move as "open-source software for Huawei's Ascend chips"China's AI industry closes ranks as Deepseek ships open-source software for Huawei's Ascend chips - the-decoder.comDeepSeek opens tools to help Huawei chips supplant Nvidia in AI - South China Morning Post. Chip stocks were already pricing the pre-Nvidia window before any single shipmentChip stocks set for $229 billion payday before Nvidia ships a GPU - thestreet.com; the Reuters reality check on Nvidia's financing bet sits uneasily next to thatNvidia's bet that its chips can finance the AI boom gets a Wall Street reality check - Reuters.

Cyber capabilities and the security boundary

Anthropic published a report on GLM-5.3 and the spread of advanced cyber capabilitiesGLM-5.3 and the spread of advanced cyber capabilities - Anthropic. A follow-up headline reports the vendor's claim that GLM-5.3-Flash built a working exploit chain for $20.40Anthropic says GLM-5.3-Flash built a working exploit chain for $20.40 - mixed-news.com. That dollar figure is a vendor claim until reproduced in an independent lab — the value of the disclosure is the existence of the capability, not the price tag. On the defensive side, OpenAI disclosed disrupting a coordinated model-distillation campaignDisrupting a coordinated model-distillation campaign - OpenAI, with reporting tying the attempts to Moonshot AI's Kimi K3 via roughly 16,000 requestsMoonshot AI Of Kimi K3 Fame Tried To Crack OpenAI’s Encrypted Reasoning Through 16,000 Requests, Bolstering Trump Administration’s Distillation Claims - Wccftech. Kimi K3 itself was released as a 1.56TB open-weights drop earlier in the windowMoonshot AI’s Kimi K3 Hits Bedrock: 1.56TB Release [2026] - tech-insider.org. The read: distillation is now an industry-grade attack surface, and vendors are naming names.

Agents: announced, downloaded, evaluated

OpenAI introduced "dots," described as a new AI personal agentIntroducing dots - OpenAISam Altman unveils "dots," OpenAI's new AI personal agent - CBS News. Meta's Muse AI agent was reported to have reached 5 million downloads faster than ChatGPTMeta's Muse AI agent reaches 5 million downloads faster than ChatGPT - qz.com. A third-party comparison published the same week lines up OpenAI Dot, Grok Bot, and Meta Muse on pricing, specs, and benchmarksOpenAI Dot vs. Grok Bot vs. Meta Muse: Pricing, Specs, Benchmarks and the Evidence - kingy.ai. The split is clean: "dots" is announced (press release, no deployment surface cited); Muse is available (download count is a real measurement); the comparison is a third-party reading, not an independent benchmark. Pick the one your procurement question actually asks about.

Bottom line for engineering decision-makers

• Frontier models: treat Gemini 4 ArgonGemini 4 Argon: our next era of frontier intelligence - blog.google and GPT-5.6 SolPreviewing GPT-5.6 Sol: a next-generation model - OpenAI as announced, not available, until the model cards and pricing surface.
• China silicon: the DeepSeek–Huawei integration is the week's most concrete infrastructure storyDeepSeek unveils Huawei AI chip software to challenge Nvidia - Seeking AlphaChina's AI industry closes ranks as Deepseek ships open-source software for Huawei's Ascend chips - the-decoder.comDeepSeek Announces a Stronger Software Integration with Huawei GPUs - analyticsindiamag.comDeepSeek opens tools to help Huawei chips supplant Nvidia in AI - South China Morning Post. The cost and latency trade-offs versus Nvidia are not in these headlines — measure them before any planning assumption.
• Security: the Anthropic GLM-5.3 disclosureGLM-5.3 and the spread of advanced cyber capabilities - AnthropicAnthropic says GLM-5.3-Flash built a working exploit chain for $20.40 - mixed-news.com and the OpenAI distillation takedownDisrupting a coordinated model-distillation campaign - OpenAI are both signals that the threat model is now part of the model spec, not a separate review.
• Agents: download numbers are the cleanest signal in the agent category this weekMeta's Muse AI agent reaches 5 million downloads faster than ChatGPT - qz.com; the rest is positioningIntroducing dots - OpenAISam Altman unveils "dots," OpenAI's new AI personal agent - CBS NewsOpenAI Dot vs. Grok Bot vs. Meta Muse: Pricing, Specs, Benchmarks and the Evidence - kingy.ai.