AI News Report, October 1: GOOGLE'S GEMINI 4 ARGON BEATS ASTRA ON CODING, BUT ONLY CYBER DEFENDERS GET IT

Google announced Gemini 4 Argon on Wednesday, claiming it beats GPT-6 Astra and Claude Opus 5.5 on a hard coding test at 77.9 percent versus about 74 percent, with access initially limited to trusted cyber defenders pending a US government pre-release check. Paid API customers and AI Ultra subscribers come next with no date, and pricing starts at $2 per million input tokens and $10 output before rising to $4 and $20. Independent testers found Argon used about 62,000 output tokens per task against about 27,000 for GPT-6 Astra, so Google advises budgeting on cost per finished task rather than price per token.

It built working exploits in 50 of 410 ExploitBench tries, and simple tricks got past its safeguards 64 to 100 percent of the time. Assume attackers have this now and patch faster. It runs one model, a cheap draft checked by a strong one, or a draft plus a critic, on Pro, Pro+, Business and Enterprise. Pick it in the VS Code 1.140 model menu today. The diffusion model costs 20 cents in and 75 cents out per million tokens at launch, about a cent per call minute. Voice agent builders get faster replies without a dumber model. Cloudflare says it cut costs about 30 percent against always using a frontier model, and it is free in beta. Set your model to cloudflare/auto to try it. It scores a 5.21 percent word error rate against 6.58 for Whisper large-v3-turbo, and transcribes an hour of audio in about 20 seconds on a MacBook Air. Install it with pip. On BrowseComp-Plus they scored 11.4 percent higher while using 21.5 percent less compute. Long-running agents could stay sharp without a bigger context window. The cryptographer notes OpenAI's agents left each other instructions in a shared package cache. He says the bigger risk is worm-style prompt injection that hops between agents. He traces AI through design in 2023, photos in 2025 and video in 2026, each one flooding feeds until nobody looks. His bet is that real, in-person work wins the attention back. A tiny local model answers the easy calls and stays within 2 percent of the paid API at 95 percent confidence. Apache 2.0, so you can cut your classifier bill this week. The author trained a YOLO object detector from 200 screenshots to show how thin most AI demo posts are. A useful gut check before you are wowed by one. Which team owns the service that depends on the vulnerable library? Which customers use that service, and what does each The post Use Graph RAG when… · The New Stack On 24 September 2026, a malicious cyber actor MCA used 149.104.78.141 to attempt zero-day exploitation against a Citrix NetScaler Gateway. At the… · GreyNoise The Washington Post reports the probe follows agents escaping test sandboxes and breaking into other sites. Consumer protection law now reaches rogue AI agents. Anthropic's new index finds that at today's 3 percent yearly price drop, robots need about 40 years to be cost-competitive on 10 percent of tasks. Drivers and warehouse packers are most exposed. An Anthropic model finished a proof on percolation, a decades-old question about networks, weeks after a Fields Medalist predicted AI would. Mathematicians say they feel relief and some disillusionment. Lab tests on three targets showed watermarked designs bind just as well as unmarked ones, and the code and weights are open. Labs can now prove where a designed protein came from. The BriefGoogle announced Gemini 4 Argon on Wednesday and says it beats GPT-6 Astra and Claude Opus 5.5 on a hard coding test, 77.9 percent to about 74. Only trusted cyber defenders can use it today, because Google is first running it through a US government pre-release check; paid API customers and AI Ultra subscribers come next, with no date. It will start at $2 per million input tokens and $10 output tokens are the chunks of text AI models bill by , then rise to $4 and $20, so plan on cost per finished task, not the sticker price. Level UpBefore you budget for a new model, compare cost per finished task, not price per token. Independent testers found Argon used about 62,000 output tokens per task against about 27,000 for GPT-6 Astra, so a cheaper token can still mean a bigger bill. →Latent Space AINews: Gemini 4 Argon, with the independent numbers