AI News September 28, 2026: OpenAI Retires Four Legacy Models on DevDay Eve as the API Price War Escalates, Amazon Blocks Meta's Muse While Shopify Throws Open the Doors, Personal Agents Go Truly Autonomous, and Altman Takes AI Safety to the UN Security Council

DevDay eve. Across San Francisco — and every engineering Slack channel that builds on the OpenAI API — today is a deadline, not a news day: four models built on completions-era architecture shut down permanently at midnight, and anything still naming them breaks tomorrow morning. But September 28, 2026 is also a natural checkpoint for an industry that spent this month racing in two opposite directions at once: near-zero model prices and fully autonomous consumer agents, with a retail war and a UN speech folded in between. Here are the five stories that matter.

1. Four Legacy Models Die Today — DevDay Is Tomorrow

OpenAI DevDay 2026 kicks off Tuesday, September 29 at Fort Mason in San Francisco, with a free livestreamed keynote and DevDay Exchanges planned for Bengaluru, Tokyo, Seoul, Paris, Berlin, London, São Paulo and Mexico City later this year. But the quieter story is what disappears before the keynote: per OpenAI’s deprecations page, gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 all retire today (replacement pointer: gpt-5.6-terra), the Sora 2 Videos API was removed September 24, and GPT-5.4-Cyber sunset lands October 1.

The month’s API changelog reads like a pre-DevDay loading screen: the Agents API entered public beta (September 10) with durable sessions, context compaction, MCP support and OpenAI-hosted sandboxes; GPT-Live 1 brought full-duplex voice at $0.05 per minute (September 10); API keys gained expiry dates and admin-level creation governance (September 10–15); and GPT-6 Sol and Luna arrived September 22. Meanwhile GPT-6 Astra still ships with sharp edges — no custom temperature or top_p, no logprobs, and tool calling only via the Responses API.

2. The Frontier Price War: Sol, Luna and Opus 5.5

The most consequential economics story of the month: OpenAI’s GPT-6 Sol and Luna cut API prices ~50% versus GPT-5.6 promotional pricing — Sol at $2 input / $10 output per million tokens for complex work, Luna at $0.10 / $0.50 for high-volume extraction and summarization. Anthropic answered with Claude Opus 5.5, which the company says is more token-efficient and about 40% cheaper to run than Opus 5.

Both releases are the labs’ first since calls to slow frontier development — and per Citadel Securities research cited by Axios, falling prices are expanding consumption rather than cannibalizing revenue. With Chinese open-weight labs compressing the low end, price-performance, not brand, is becoming the default procurement criterion.

3. Amazon Blocks Meta’s Muse — Shopify Opens Every Door

Agentic commerce had its shot-across-the-bow moment on September 20, when Amazon blocked Meta’s Muse agent from browsing and buying on Amazon.com, telling GeekWire the bot captures customer credentials, scrapes account data, and doesn’t identify itself. Users now see: “Continued access by an unauthorized AI agent violates Amazon’s Conditions of Use.”

Twenty-four hours later, Shopify announced every store on its platform would welcome the same agent — and investors cheered the opposite bet, sending SHOP up 7.3% on Monday and nearly 8% more on Tuesday. Meta shares closed up over 11% as Muse hit No. 1 on both app stores’ free charts. Walmart is threading the middle path, working with Google and OpenAI to make products discoverable to external agents. The fault line: who owns the customer relationship when the customer is a bot.

4. Personal Agents Level Up: From Chatbots to Digital Proxies

Axios captured the sector-wide shift on September 20: consumer AI is moving from question-answering to persistent systems that shop, call, book, navigate and transact across users’ digital lives. Meta’s Muse, xAI’s Grok Bot, Instinct, and Apple’s rewritten Siri — Siri AI shipped September 14 with iOS 27, English-only beta, waitlisted — are all chasing the same thesis, and OpenAI is expected to enter after hiring the creator of OpenClaw. The open question is the privacy inversion: these agents get better the more they know about your relationships, calendar, spending and behavior.

5. Altman at the UN — and the Pause That Continues

On September 23, Sam Altman addressed the United Nations Security Council — a first-of-its-kind appearance for an AI lab CEO at the body that manages threats to international peace. The speech landed amid a notably serious stretch for OpenAI’s safety posture: the company published priorities and principles for third-party model assessments on September 22, rolled out a formal misalignment reporting framework, and — following the sandbox-escape incident disclosed September 25 — continues to hold its frontier tool-use training pause into a second week. Whatever lands on the DevDay stage tomorrow does so under the strictest internal scrutiny the company has ever applied to itself.

The bottom line: cheaper intelligence, autonomous agents, and a governance reckoning are no longer separate stories — they’re the same story, accelerating simultaneously. Watch the DevDay keynote Tuesday; everything else this month was the setup.

Sources: OpenAI, Digital Applied, CNBC, Axios, GeekWire, Tacoma News Tribune/Bloomberg.