Wednesday’s through-line is a cheaper near-flagship model, an AI agent aimed at shop owners, and the full text of a voluntary White House safety deal. OpenAI shipped GPT-6.1 Sol after shelving Astra. Meta expanded Muse for small businesses. And overnight reporting laid out what the “morally binding” accord actually asks the labs to do.
1. OpenAI ships GPT-6.1 Sol: near-Astra smarts, lower price
At DevDay on September 29, 2026, OpenAI introduced GPT-6.1 Sol, an upgrade to GPT-6 Sol aimed at coding, computer use, and professional work. In its launch post, the company says the model nearly matches GPT-6 Astra’s intelligence on those tasks at one-fifth of Astra’s standard input and output token prices. TechCrunch covered the same launch and noted that GPT-6.1 Astra is not shipping, after earlier safety concerns around that higher-tier release.
OpenAI’s own numbers (treat them as company claims, then verify on your workloads): on DeepSWE-style coding evals it says Sol matches Astra at roughly one-fifth the cost; on professional PDF work and multi-step business workflows it says Sol approaches Astra while beating prior Sol scores. Factual errors on hard prompts reportedly drop versus GPT-6 Sol at low reasoning effort. Standard API pricing listed at launch is $2 per million input tokens, $0.10 per million cached input, and $10 per million output. Availability: Plus, Pro, Business, Enterprise, and Edu in ChatGPT Work and Codex, plus the API as gpt-6.1-sol. It is not yet in regular Chat.
Why it matters: if you build agents or run heavy coding/document jobs, Sol is the practical follow-up to Tuesday’s Astra pullback. Benchmark it on your real prompts before you rip out Astra routes. Watch cache reuse, because the cheap cached-input rate only helps if you keep stable context. And keep a human gate on computer-use and tool actions. Cheaper tokens do not mean unsupervised spend or unsupervised clicks.
Sources: OpenAI – Introducing GPT-6.1 Sol, TechCrunch – OpenAI launches GPT-6.1 Sol.
2. Meta expands Muse for small businesses
Also on September 29, Meta announced Muse for Small Business, expanding its personal AI agent with skills and connectors meant for owners who already live in Meta apps and common business tools. Meta’s blog says you can give Muse a goal (run the business, find customers) and connect Facebook Pages, Instagram professional analytics, Meta ads, plus tools including Shopify, Slack, QuickBooks, Stripe, Canva, Notion, and more. TechCrunch reports a free tier with usage limits and paid plans for heavier use. Muse remains positioned for the U.S. and Canada.
Meta stresses approval gates: nothing publishes, sends, or spends without you saying yes. That is the right default for any agent that can touch ads, email, or bank-adjacent tools. The pitch is less “another chatbot” and more “an agent that already knows your brand voice and what customers keep asking.”
Why it matters: if you run a shop, clinic, or side hustle on Instagram and Shopify, this is one of the clearest SMB agent bets this week. Start narrow: analytics summaries and draft campaigns, not live ad spend. Connect only the accounts you need. Keep a weekly audit of what Muse proposed versus what you approved. And do not hand it your only QuickBooks login without 2FA and a recovery plan.
Sources: Meta – Muse for Small Business, TechCrunch – Meta expands Muse to small businesses.
3. The White House AI accord: four layers, no penalties
On September 30, The Verge published the substance of the voluntary deal announced after Tuesday’s White House lunch. The document, titled the Joint Commitment on Frontier Responsibilities, was shared publicly via presidential adviser David Sacks. Signers named in coverage include Google’s Sundar Pichai, Anthropic’s Dario Amodei, Meta’s Mark Zuckerberg, OpenAI’s Greg Brockman, xAI’s Elon Musk, and Nvidia’s Jensen Huang.
The accord lists four layers: (1) robust internal controls for cybersecurity, biosecurity, chemical threats, and unintended system access; (2) an internal team that keeps those controls working and fixes problems; (3) an independent external auditor or evaluator; and (4) an independent board committee that gets the reports and pushes remediation. The text leaves room to “codify these steps into laws or regulations” later. It does not spell out penalties, public disclosure duties, or hard deadlines. Separately, Trump ordered federal agencies to prefer the phrase “Super Intelligence” over “artificial intelligence” in executive-branch usage, per White House fact-sheet language and contemporaneous reporting.
Why it matters: this is self-policing with a press conference, not a regulator with teeth. Useful as a checklist for vendor questions (“Who audits you? Who on the board sees the report?”), weak as a guarantee that the next rogue agent gets reported fast. If you buy frontier APIs for customer work, ask for the external-eval partner and board oversight path in writing. Pair that with your own logs, allow-lists, and same-day incident clocks. Do not wait for Congress to finish an AI bill.
Sources: The Verge – How tech leaders will self-police under Trump’s deal, White House – Fact sheet on Super Intelligence terminology, The Guardian – “Morally binding” AI deal coverage.
Quick side note: Gemini Gems become Skills
Google is also reshuffling everyday Gemini habits. TechCrunch reports Gems will migrate to stackable “skills” starting November 17, 2026, with a slash-command style picker in chats. If you built custom Gems for work or school routines, export your instructions now and try the Skills flow early so November is a tweak, not a scramble.
What to do with Wednesday’s news
- On OpenAI? Pilot GPT-6.1 Sol on coding and PDF-heavy jobs. Keep Astra (or another flagship) as a quality check until you trust the cost/quality tradeoff.
- Running a small business on Meta? Try Muse on drafts and analytics first. Leave publish/send/spend locked behind your approval.
- Buying frontier models? Ask vendors which external auditor they will use under the new accord, and put incident reporting into the contract yourself.
- On Gemini? Back up Gem instructions before the November Skills migration.
That is Today in AI for September 30, 2026: a cheaper Sol landing while Astra stays shelved, Muse aiming at the Main Street workload, and a White House safety accord that finally has four bullet points and still no enforcement stick.
Share this article
Leave a Reply