My bots need a new engine.
That is the short version of October 2026 for anyone who ships on OpenAI’s stack. GPT-5.5 retires from ChatGPT, ChatGPT Work, and Codex on October 14, 2026, across all plans. The API isn’t affected, which matters a lot depending on where your code actually calls from. If you’re running Codex with ChatGPT sign-in, you pick a replacement model or you get picked for.
Meanwhile GPT-6 Astra showed up with a pitch aimed squarely at people like me: better coding, better research, better handling of complex multi-step work. And GPT-6 Sol and Luna are available at lower token prices, which is the detail I circled twice.
Why the retirement date is the real story
New model launches get the headlines. Retirements get the 2 a.m. pages.
If you’ve built anything with a pinned model name, October 14 is a deadline, not a news item. The split between ChatGPT surfaces and the API is the part worth mapping carefully. Bots that authenticate through ChatGPT sign-in and bots that hit the API directly are now on different clocks, and if your architecture mixes both, you have two migration paths to track instead of one.
What I’d do this week:
- Grep your codebase and config for every hardcoded model string. All of them. Including the one in that script you wrote for yourself eight months ago.
- Separate your inventory into ChatGPT-surface calls and API calls, because only one group is facing the October 14 cutoff.
- Pick replacements per workload, not globally. A summarizer and a multi-step research agent do not need the same model.
- Keep a saved set of real prompts with known-good outputs so you can compare behavior before and after you swap.
That last one is the step people skip. Model migrations rarely break loudly. They shift tone, change how often a tool call fires, or quietly stop following one line of your system prompt. You catch that with a regression set, not with vibes.
Three models, three price points, one decision tree
Astra, Sol, and Luna being available together is the useful part of the GPT-6 rollout for builders. Sol and Luna sit at lower token prices. Astra is positioned for the heavy stuff: coding, research, multi-step work.
That’s a routing problem, and routing is where cost control actually lives. The pattern I keep coming back to is a tiered router with an escalation rule. Cheap model handles the first pass. If confidence is low, the task requires multiple tool calls, or the output fails a validation check, escalate to the stronger model. You pay for the expensive reasoning only when the task earns it.
A rough routing sketch
- Classification, extraction, short rewrites, routing decisions themselves → cheapest tier.
- Drafting, summarizing long documents, structured output with a schema → mid tier, with schema validation as the escalation trigger.
- Code generation across files, research chains, anything with more than three dependent steps → top tier.
Build the escalation logic before you tune the prompts. Prompt tuning on the wrong tier is wasted work.
Voice with plugins changes the interface question
ChatGPT added Voice with plugin support across web, iOS, and Android, including Voice in Work on web and mobile. For bot builders, voice plus plugins means your tool can be invoked by someone who is driving, cooking, or walking between meetings.
That changes design constraints more than it changes code. Voice responses need to be short, confirmable, and recoverable. If your plugin returns a table, a voice user gets nothing useful. If it performs a write action, a voice user needs an explicit confirmation step, because misheard input is a real failure mode. I’d design tool responses with two shapes from the start: one for screens, one for ears.
Data plugins and agents that stick around
The Data plugin for Work and Codex analyzes connected business data, and OpenAI also described agents capable of taking on ongoing responsibilities rather than one-shot tasks.
Persistent agents are a different engineering problem than chat completions. Ongoing responsibility implies state that outlives a session, scheduling, and some way to audit what the agent did while nobody was watching. Before you connect any agent to live business data, decide what it can read versus write, log every tool invocation with inputs and outputs, and give yourself a kill switch that doesn’t require a deploy.
Connected business data also means access scope is now a design decision, not an afterthought. An agent with broad read access to internal systems is a useful teammate and a sizable surface area. Scope it narrowly and widen it deliberately.
What I’m actually doing
Migration first, features second. Pin down the October 14 exposure, build the regression set, then wire up tiered routing against Sol, Luna, and Astra to see where the cost curve lands. Voice-shaped tool responses after that. Persistent agents last, once I trust the logging.
Boring order, fewer surprises. DevDay 2026 brought more than 20 major announcements, and the fun ones can wait until nothing is on fire.
🕒 Published: