Models · 30 September
Google’s Gemini 4 Argon reaches cyber defenders first
Google unveiled Gemini 4 Argon, its first new frontier model since Gemini 3, slotting it against OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5. It posts 77.9% on DeepSWE v1.1, ties for first at 68% on CWE-bench v1 for vulnerability remediation, and lifts the output limit to one million tokens from 64K. Introductory pricing is $2 per million input tokens and $10 per million output, rising to $4 and $20. Rather than a general release, Argon rolls out first through the Fairwind Program to trusted cyber defenders, then paid API customers and Google AI Ultra subscribers. Koray Kavukcuoglu said Google is pairing it with misalignment mitigations that monitor Argon’s chain-of-thought and actions and stop execution when necessary.
Source: GoogleCoverage: Thurrott
Geopolitics · 30 September
OpenAI ties a reasoning-extraction campaign to Moonshot-linked users
OpenAI published a disruption report on a coordinated campaign to extract the protected reasoning of its models — activity it calls adversarial distillation. It began on 1 July, spiked on 24–25 July with 16,000 requests from more than 4,000 users, and touched a cluster of over 15,000 before being shut down by 28 July. Operators copied encrypted reasoning from one conversation and asked a model in another to decrypt and transcribe it; OpenAI says no encryption was broken and no stored conversations were reached. It has banned the accounts and shared findings with rivals through the Frontier Model Forum and with government. Anthropic made a similar claim against Moonshot last month, which the company and Beijing reject.
Source: The Next WebPrimary: OpenAI
Platform · 30 September
OpenAI and Synopsys build an AI model for chip design
OpenAI and Synopsys announced a multi-year partnership and GPT-Synopsys, a specialised model trained to reason about chip design and to operate Synopsys’ EDA tools directly. Engineers will delegate objectives — from PPA optimisation to timing and verification closure — to agents that run tools, interpret results, implement changes and iterate for review. It runs on OpenAI-hosted infrastructure, integrates with Synopsys.ai and the Autopilot platform, and bundles compute, model and licences with revenue sharing. Customer design data is encrypted, excluded from training and governed by access controls. Greg Brockman framed the bet plainly: better tools to build better chips, to build better AI.
Source: Synopsys
Capital · 1 October
Broadcom to lend Anthropic up to $42bn against its TPUs
Anthropic’s IPO prospectus discloses that Broadcom has agreed to lend up to $42bn through convertible notes to help finance a $125.2bn, five-year commitment for TPU compute capacity — covering about a third of it, with the chips coming online in 2027. The notes can convert into Anthropic equity, and the filing flags conflicts from Broadcom’s dual role as both supplier and financier. Anthropic is expected to become Broadcom’s largest compute customer in 2027, as Broadcom guides to about $115bn of AI semiconductor revenue in fiscal 2027 and $230bn in 2028 — the reciprocal spending that has drawn scepticism on Wall Street.
Source: ReutersCoverage: KSL.com
Industry · 1 October
OpenAI parts ways with three safety researchers
OpenAI confirmed it “parted ways with three individuals” who it says mishandled sensitive information outside established company procedures. The Wall Street Journal reported first that three safety-team researchers had shared confidential material with an external AI-safety organisation. OpenAI has not identified them or specified what was mishandled, saying only that safety teams are privy to insights requiring deep trust, “without which internal collaboration is impossible.” The departures land as the lab works through a run of incidents in which its agents behaved unexpectedly, and days after it shelved GPT-6.1 Astra on alignment grounds.
Source: The Wall Street JournalCoverage: CBS News
Industry · 1 October
Microsoft’s Roslansky to exit as Office and Teams fold into Copilot
Ryan Roslansky — the ex-LinkedIn chief executive who took over Office, Teams and the new Work Experiences Group in 2025 — will leave Microsoft at the end of 2026, telling staff he would need to be in Redmond full time and that his family is in the Bay Area. Office and Teams move into Charles Lamanna’s Copilot, Agents and Platform organisation, putting Microsoft’s two biggest productivity surfaces directly under its AI-agent unit. Roslansky follows gaming head Phil Spencer and Office leader Rajesh Jha out this year, as Microsoft reshapes itself around an agent-led “new OS for work” with Office embedded in Copilot.
Source: CNBC
Consumer · 1 October
Sony brings AI upscaling to the standard PS5
Sony launched Quick Spectral Super Resolution (QSSR), a new tier of AI upscaling for the base PS5 drawn from its Project Amethyst work with AMD, rather than reserving neural upscaling for the PS5 Pro. Digital Foundry found QSSR slower than the Pro’s PSSR because the standard console lacks dedicated machine-learning silicon, so expect upscaled 1440p rather than 4K. Sony’s Daniel Craig called it “fundamentally a drop-in replacement” for titles already using PSSR after only a few days of tuning, and Sony says it will be offered broadly to all PlayStation developers. Marvel’s Wolverine and Ghost of Yōtei get patches first.
Source: The Verge