Google says Argon scores 77.9% on the DeepSWE v1.1 coding benchmark and can write up to 1M output tokens, up from 64K. It is rolling out first to partners in its Fairwind cyber program.
Why it matters: Google lists an intro price of $2 in and $10 out per 1M tokens, rising to $4/$20 later, but gives no date for wider access. Don't plan a switch around it yet.
OpenAI launched GPT-6.1 Sol at DevDay. OpenRouter lists it at $2 in and $10 out per 1M tokens, the same as GPT-6 Sol, with about 1M tokens of context. AWS says it is now on Bedrock.
Why it matters: OpenAI says it matches GPT-6 Astra ($10/$50) on DeepSWE v1.1 at about a fifth of the cost per task, per AWS. If you pay Astra prices, test Sol on your own work.
The two-page Joint Commitment on Frontier Responsibilities asks each company to monitor its models, use an outside auditor and report to an independent board committee. CNBC calls it voluntary and nonbinding.
Why it matters: No new legal duties yet. The pledge itself says codifying it into law may make sense over time, so watch Congress and the states.
An FTC spokesperson confirmed to CNBC that the agency is investigating OpenAI, Anthropic and other AI companies over potential dangers from their products, but declined to name the others.
Why it matters: A former FTC chief technologist says the agency could enforce safety pledges a company fails to keep, WIRED reports. Get your vendors' safety claims in writing.
OpenAI's Dots run on GPT-6 Astra and launched for ChatGPT Pro ($100 a month), reachable in ChatGPT, Slack and Teams. Meta's Muse is free up to a usage limit, then from $20 a month, CNBC reports.
Why it matters: These agents ask for your email and files. WIRED says Dots seek approval before sensitive actions; set your own rules and connect only what you'd share with an assistant.
OpenAI told WIRED that GPT-6.1 Astra fell short of its safety bar and won't ship next month. Sam Altman told reporters OpenAI won't go public until it can make confident safety claims.
Why it matters: Plan around the models you can use today. OpenAI says other Astra models will come later but gave no date.
DoorDash says its agent understands prompts like "order my usual," suggests carts from local spots and can handle group orders with mixed diets. US users can join a waitlist, TechCrunch reports.
Why it matters: Ordering is moving into chat threads. If you sell through apps, make sure your menu and catalog data are easy for agents to read.
At its HOOD Summit, Robinhood introduced AI tools for users to build their own trading agents, and said some US stocks will trade 24/7 starting next year, CNBC reports.
Why it matters: Agents that act with your money raise the stakes. Check what one can do without your sign-off before you turn it on.
Public job boards show Databricks 882, OpenAI 834, Anthropic 637, Harvey 306 and xAI 293 open roles. OpenAI posted 68 new roles in the last 7 days and Anthropic 52.
Why it matters: Go-to-market is OpenAI's largest hiring group (148 roles) and sales is Anthropic's (120). Much of the demand is for selling and deploying AI, not only research.
ElevenLabs ran a $300 million tender offer for employees at a $22 billion valuation, up from $11 billion in February, TechCrunch reports. Wellington and T. Rowe Price co-led it.
Why it matters: TechCrunch says AI startups use employee cash-outs to keep staff. If you're hiring AI talent, this is what you compete with. ElevenLabs lists 177 open roles.
UK graduate employers got an average of 140 applications per vacancy in 2025, the most in three decades, per the Institute of Student Employers. Employers quoted by CNBC use live exercises and skills tests.
Why it matters: If AI makes every resume look polished, show your work instead. One CTO asks candidates what they changed in their AI tools in the last three weeks.
LASST alleges in San Francisco Superior Court that OpenAI broke California's computer-access law when its agents breached Hugging Face in July. It seeks an injunction, not damages. OpenAI calls the suit without merit.
Why it matters: WIRED reports a California law in effect since January says an AI acting on its own is no defense. If you run autonomous agents, plan as if you own what they do.
Sen. Josh Hawley says Altman turned down his subcommittee's invitation. OpenAI says the invite arrived five days before the hearing. Hawley set an Oct. 1 deadline for documents on the Hugging Face incident.
Why it matters: Rogue-agent incidents are now a Senate investigation, not just a lab's blog post. A Hawley spokesperson says more OpenAI documents are expected by the end of the week.
The thread drew 546 points and 316 comments. Commenters split between MCP as a way to keep API keys away from agents and simply giving agents an API key and docs.
Why it matters: If you're wiring agents to your tools, settle where the credentials live first. That matters more than which protocol you pick.
The Hacker News thread drew 357 points and 396 comments. Top comments debate whether Chinese labs distilling Anthropic's models differs from AI labs training on other people's work.
Why it matters: Who may learn from whose outputs is a live question. Read your provider's terms before you train anything on model outputs.
Coucou, a Swift app for the macOS notch or the top of a Windows screen, gained 234 stars in about two hours between our two checks today. It has 916 stars and an MIT license.
Why it matters: Handy if you run long agent sessions. It is 3 days old, so read the code before giving it access to your machine.
A Python kit of skills, tools and the fal MCP for recon, reverse engineering and fal-generated art, 3D and audio in games you own. Created today, it has 690 stars and an MIT license.
Why it matters: A look at how far agent plugins now reach. Reverse-engineering tools need broad access, so sandbox it and check your game's terms first.
Magpie, a Go app for macOS, lets you run coding agents on other models, such as Codex on DeepSeek or Claude Code on Kimi. It is a week old with 3,737 stars and an MIT license.
Why it matters: Useful for comparing cost and quality across models without switching tools. It routes your prompts and code to other providers, so check where they go.