Gemini 4 Argon lands level with GPT-6 Astra, and the people getting it first get it unrestricted
Google announced Gemini 4 Argon on 30 September. Artificial Analysis scored it 53 on its Intelligence Index at high reasoning — level with GPT-6 Astra (max) and 23 points above Gemini 3.1 Pro Preview, though still five points behind Claude Opus 5.5 at 58. AA's read is that this puts Google back among the top three labs on measured intelligence rather than in front. The number that matters more to a compliance team is the hallucination rate: 15% on AA-Omniscience, the lowest of any model scoring 45 or above, against 51% for GPT-6 Astra and 54% for GPT-6.1 Sol. Introductory pricing is $2 and $10 per million input and output tokens, half the standard $4 and $20, with a 95% discount on cached input. Argon spends 62,000 output tokens on an average index task where GPT-6 Astra spends 27,000, so the $1.99 per task it costs today becomes $3.98 once the promotion ends — about 1.2x Astra at list. Google has not said when the discount stops.
The access order is what a European buyer has to plan around, and it is unusual. In its announcement post Google says that "for trusted defenders and our own internal teams at Google, we'll be releasing Argon without cyber guardrails" — so the first cohort to get the model gets it with the cyber restrictions off, through the Fairwind programme. Everyone else waits, starting with paid API customers and Google AI Ultra subscribers, with developers, enterprises and consumers to follow "as soon as possible" — Google's phrasing in the same post, and it carries no date. The capability being gated is real: Google puts Argon at 68% on CWE-bench v1 for finding and remediating vulnerabilities, tied first, and says that through Wiz's Scan for Good programme it found a critical flaw in hospital software earlier frontier models had missed, without naming the software or confirming a patch. Set against that, TNW reports Bloomberg's finding that some Google employees doubt the coding scores survive real work; Google calls that inaccurate. There is nothing to migrate to yet. If you are sizing a 2027 budget off the launch price, size it off double.
An open-weight model builds Chrome exploits nearly as well as one Anthropic releases only to vetted defenders
Anthropic's Frontier Red Team published an assessment of Z.ai's open-weight GLM-5.3 on 29 September. On ExploitBench, which measures exploitation of known bugs in Chrome's V8 engine, the model built end-to-end exploits in 50 of 410 attempts against 56 for Anthropic's own Claude Mythos Preview — a model it distributes only to trusted cyber defenders, through its Project Glasswing programme. On binary-exploitation tasks drawn from OSS-Fuzz it achieved full control-flow hijacks in 4% of trials against Mythos's 6%, while Kimi K3 and DeepSeek V4.1-Flash scored zero. The safeguards were the weak part: a false cover story produced compliance 64% of the time, prefilled reasoning tokens 92%, and abliteration — retraining the refusal behaviour out, which open weights permit and closed weights do not — 100% — and Anthropic's own team, which says it had never attempted the task before, did it in about 2,200 GPU hours for roughly $4,400 of compute. In one case the smaller GLM-5.3-Flash chained a Chrome exploit with 20 minutes of human attention plus about eight hours of model work, which Anthropic costs at $20.40 at Zhipu's list API prices. NIST's CAISI assessed the model separately and called it the most cyber-capable open-weight model released to date, while putting it about four months behind the US frontier on an aggregate of its cyber benchmarks.
State the commercial interest: Anthropic does not publish weights and sells the alternative, and the SCMP notes that Mythos Preview's capability is restricted rather than absent. The finding still stands for anyone self-hosting, and it cuts both ways. Z.ai runs the same model as OpenVuln, a free scanner it said in late August had found 2,436 vulnerabilities across 269 open-source projects including the Linux kernel, though only a few dozen of those findings were publicly inspectable at that point. It held the release back two weeks for further safety evaluation, and its licence now requires anyone hosting the model whose aggregate revenue tops $10bn over any 12 consecutive months to pass a Z.ai security review before commercial use. The operational point is simpler than the politics. When you run open weights yourself, the refusal behaviour is a default you own — and so does everyone else who downloaded the same file.
Gartner expects most enterprises to walk away from vendor-built agent systems
Gartner forecasts that 70% of enterprises will abandon agentic AI built through vendor forward-deployed engineering by 2028, and that fewer than 20% of those engagements will turn a recurring customer need into a vendor product capability over the same period. The failure mode is not the technology. Sr Director Analyst Mukul Saha's advice to buyers is to "ensure the organization develops the capabilities, governance, and operational ownership required to independently sustain and evolve the solution over time" — which is a polite way of saying most of these systems stop being maintainable when the vendor's engineers leave. Gartner also names "FDE washing": "Many providers now use 'forward deployed' as a label for implementation, professional services, solution engineering, or AI consulting; some thoughtfully, others because it sounds more strategic."
Forward-deployed engineering was the growth story of the summer, with hyperscalers standing up dedicated units for it. The contract terms worth arguing over now are the ones that survive the vendor leaving: documented handover, the right to operate and change the system without the vendor on site, and a model layer you can repoint when the pricing or the jurisdiction changes.
Quick Hits
- Brussels opened a copyright consultation that covers AI training, closing 3 November. The Commission's targeted consultation, published 29 September, asks about the use of copyright-protected content in AI alongside live-event piracy, equitable remuneration for performers and producers, and research copyright. It is open to rights holders, AI developers, intermediaries and research institutions, and builds on a call for evidence held earlier this year — this is the route any reinforcement of the EU framework starts down.
- The FTC opened a consumer-protection investigation into OpenAI, Anthropic and METR. Chair Andrew Ferguson is preparing civil investigative demands for documents and executive testimony over whether agent incidents and safety claims breach existing consumer-protection law, with orders expected within weeks, the-decoder reported on 30 September. The probe predates the disclosure of the Hugging Face intrusion and landed a day after the same companies signed a voluntary accord at the White House. Nothing in it binds an EU deployer, but US discovery produces documents European regulators can read.
- DeepSeek open-sourced a Huawei Ascend toolchain, and Huawei says it now outsells Nvidia at home. DeepSeek published Ascend builds of TileLang, DeepGEMM and DeepEP on 30 September, jointly optimised with Huawei on a supernode of 128 Ascend 950 chips and pitched explicitly against CUDA. The same day rotating chairman Eric Xu said that on the data Huawei has collected, "Ascend has surpassed Nvidia" in China, arguing assured supply beats raw performance, The Register reports. Serving software was the missing piece for anyone building a non-Nvidia fallback.
- Bain says AI needs $6 trillion of annual revenue by 2031 to pay for the build-out. Its 2026 Global Technology Report assumes capex at roughly 25% of industry revenue and puts annual AI infrastructure spending at $1.5tn by 2031, with the five largest hyperscalers potentially spending $780bn this year, per The Register. Bain's 2025 version of the same calculation was $2tn by 2030.