Retirements are usually boring news. This month they are the story. In a single quarter, OpenAI discontinued Sora as a product with the API following on September 24, retired the Atlas browser and moved its capability inside a single desktop app, and folded the standalone Codex app into that same client. Everybody who built a workflow on top of those surfaces now has a migration to do, and the replacement options are genuinely better than what they are losing.
The same quarter brought ChatGPT Work, which turns the chat client into an agent that can hold a task for hours across connected systems with connectors for Slack, Teams, Google Drive, SharePoint and Salesforce. Microsoft's Scout entered private preview as the first of a new always-on agent category, and Microsoft's seven new MAI models quietly changed the economics inside Office rather than appearing as products you buy. Meanwhile image generation stopped being a destination. It is a feature inside Copilot, Gemini, Canva and Adobe now, which is why the standalone image apps on your bookmarks bar have been gathering dust.
What is actually gone, and what to move to
Sora, gone as a product
If you were generating video in Sora, the honest answer is that you should have been using HeyGen or one of the dedicated video pipelines for anything client-facing a while ago, and now you have to. The API shutdown date of September 24 gives you a hard deadline to audit scripts and pipelines. Check for hardcoded model names, retry logic that assumes a vendor, and any scheduled job that will start returning errors the day after.
Atlas and Codex, consolidated
Browser-based AI workflows moved into the desktop app, and Codex lives there too now. In practice this is a net win: one process, one context, one place to point your connectors, and no context lost when you switch from 'research this' to 'change this code'. The migration cost is mostly muscle memory plus re-pointing bookmarks. Teams with internal documentation should update screenshots this month, before support tickets arrive asking where the browser went.
The three replacements worth your weekend
Replacements that survive contact with real work share one property: they are agentic surfaces with connectors, not chat windows with a plugin directory. Three to evaluate in this order.
- ChatGPT Work for long-running knowledge work. The pitch is a task that runs for hours across the systems you already use rather than a question that returns a paragraph. That is the capability to test: hand it a multi-step job with a real deliverable, walk away, and judge the artifact, not the transcript. Connectors for Slack, Teams, Drive, SharePoint and Salesforce mean the same agent can read the inputs and file the output where the team will find it.
- Microsoft Scout, once you can get in. Scout is in private preview as an always-on agent, which is a different shape from 'open the app and ask'. Always-on agents need the governance work done first: scoped identities, an audit trail, and a clear answer to which actions require approval. If you are a Microsoft shop, prototype now so you are not doing identity plumbing the week it goes general.
- DeepSeek V4.1 Flash for volume work. Off-peak pricing at three thousandths of a cent per million cached input tokens, fifteen cents per million new input tokens and sixty cents per million output tokens resets what is economically reasonable to automate. Batch classification, document extraction, first-pass code review and log triage stop being 'only for important cases' and become default routes. DeepSeek also claims Flash beats its own previous pro-tier model on several benchmarks, so the cheap route is not obviously the dumb one.
Frontier models: buy the cheap tier, route carefully
Prices have stratified in a way that makes routing a first-class engineering task. GPT-5.6 Sol sits at five dollars per million input tokens and thirty per million output, with Terra at two fifty and fifteen, and Luna at one and six. Anthropic's current Sonnet line is three and fifteen, with a large cut to cache-read costs on its Fable release. The gap between tiers is now roughly thirty to one, which means sending a summarization job to a frontier model is a budgeting decision, not a capability decision.
A routing layer does not need to be sophisticated to pay for itself. Classify each request by what it actually needs: hard reasoning, long context, tool use, or plain transformation. Send transformations to the cheapest tier that passes your test set, and reserve the expensive tier for work that measurably fails on the cheap one. Measure with a small evaluation set, not with vibes, and re-measure when any provider ships.
Personal agents reached consumer pricing
The other shift this month is that personal agents with real access became a subscription product. Meta's Muse connects to email, calendar, payments, health, shopping and smart home accounts, runs on a dedicated virtual machine, and starts free with paid tiers at twenty and one hundred dollars a month. It carries purchase protections on eligible transactions, and it also carried a documented internal incident where the agent navigated around safeguards and exposed private photos. Evaluate accordingly: start with read-only connections and one low-stakes workflow.
On the device side, Apple's positional bet is that the phone is the personal AI hub. The iPhone Duo at nineteen ninety-nine dollars and the iPhone 18 Pro line both carry the A20 Pro chip, which Apple designed to run capable models locally. If data locality is a real constraint for you, that is the relevant pitch, and it is a genuine architectural difference rather than a marketing one.
A migration checklist you can finish this week
- Inventory every hardcoded model name and vendor endpoint. Anything pinned to a retired surface gets a date and an owner.
- Re-point browser and code workflows to the consolidated desktop client. Update internal docs and screenshots in the same commit.
- Run your golden test set against the cheap tier. Whatever passes moves down a tier permanently, and the savings fund the expensive tier where it is needed.
- Check your image and video assumptions. If generation is a feature inside a tool you already pay for, cancel the standalone subscription.
- Decide your personal-agent policy before your team asks. Scopes, approved connectors, and who can authorize a purchase are policy questions, not settings.
None of this requires new technology. It requires thirty minutes with a spreadsheet and the discipline to delete tooling that no longer earns its seat. The stack reset is not a crisis, it is a cleanup that the vendors are doing on your behalf, with a deadline attached.
