Codex and OpenAI Changes, Aug–Oct 2026

If you set Codex up over the summer and haven't looked since, most of what you configured is still running — but the defaults around it have moved. A new model generation shipped in stages, the CLI's default model changed, a faster speed tier arrived, the Pro plan became a ladder, and several models left Codex for anyone signed in with ChatGPT. This is the period in order, grouped by what it changes for you, with the action each item calls for. For how the prices moved over a longer stretch, see the running table of pricing changes.

The new models

GPT-6 Astra arrived first, in early September, as OpenAI's most capable model. It's the model to reach for on the hardest end-to-end work, and it comes with migration notes worth reading before switching anything: it doesn't accept the none reasoning effort, custom temperature or top_p, or log probabilities, and tool calling requires the Responses API. Its launch also brought new controls for long-running work — async tool calling, mid-turn steering over WebSockets, and changing reasoning effort mid-conversation without losing the cached prefix — plus asynchronous misalignment monitoring that can pause a conversation for review.

GPT-6 Sol and GPT-6 Luna followed later in the month, at lower token prices than the GPT-5.6 models they succeed, and a fix for degraded image understanding in both landed days later — OpenAI recommends re-running evaluations that involve image inputs. GPT-6.1 Sol closed the month, pitched as near-Astra performance at a lower cost, with the deepest cached-input discount of any current model. It accepts neither none nor minimal effort. GPT-6 Astra vs GPT-6.1 Sol and GPT-6.1 Sol vs GPT-6 Sol put them side by side.

What to do: if you move a request to GPT-6 Astra or GPT-6.1 Sol, replace none with low and strip the sampling parameters — the effort error page lists exactly what to remove.

Codex itself

The CLI's default model is now GPT-6.1 Sol, replacing GPT-5.6 Sol. In ChatGPT, GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna are available in Work and Codex but not in Chat, and in Enterprise workspaces the new models arrived switched off until an administrator enables them. The CLI gained /import, which brings setup and recent chats over from Claude Code and Cursor. The codex mcp-server command was deprecated and then removed; integrations that launched it need the app server instead, which OpenAI describes as experimental and not for production. And the untrusted approval policy is gone — a config that still sets it can stop Codex from starting. Codex cloud also gained GitLab support, in beta on every ChatGPT plan, so tasks and merge-request reviews no longer require GitHub.

What to do: run /status to confirm which model your sessions actually use, and search your configs for approval_policy = "untrusted" — the fix takes a minute.

Retirements

For Codex with ChatGPT sign-in, GPT-5.4 and GPT-5.4 Mini left on August 31, 2026, GPT-5.3 Codex Spark on September 14, 2026, and GPT-5.5 leaves on October 14, 2026, across every plan. None of these touch the API or Codex used with your own API key. On the API side, GPT-5.4 Cyber was removed outright after a short deprecation. The retirement schedule lists each model's named replacement.

What to do: work through the GPT-5.5 checklist now, while you can still compare old and new on the same task.

Plans, credits and speed

Pro became a ladder of tiers ranked by included usage, with only the top tier including GPT-6 Astra Ultrafast; the middle tier reopened to new subscribers with a smaller allowance, and existing subscribers who qualify keep their previous allowance through October 29, 2026. OpenAI published a credit rate card covering every Codex model and speed, and the credits calculator turns it into a per-task figure. This site still doesn't print plan prices — the policy page explains why — but Codex plans and usage limits covers everything else about each plan.

Speed became a menu. Fast mode, renamed from Priority just before this period, started accepting long-context requests on the GPT-5.6 models in August. Ultrafast was announced in August as a limited preview for GPT-5.6 Sol and opened to all API customers for GPT-6 Astra on September 29, 2026, at low default limits and a steep price. Inside a ChatGPT plan, both speeds cost more against included usage than against credits — why Fast mode burns your plan faster explains the gap.

What to do: treat Fast as a per-session choice rather than a default, and read choosing Standard, Fast or Ultrafast before turning on either for unattended work.

Prices

GPT-5.6 Sol was cut to promotional rates in August, available at least through November 21, 2026 — a floor date, not a commitment about what follows. GPT-6 Sol and GPT-6 Luna launched below their GPT-5.6 predecessors, and every GPT-6 model carries a cache-write rate, so the write charge GPT-5.6 introduced is now the norm. Batch still costs 50% less than Standard on every model that offers it, and Flex still matches Batch.

The API platform

Several changes matter even if you never touch a new model. Organizations and projects can set hard spend limits, and the Usage and Costs dashboards filter by API key. Rate limiting gained two distinct error codes: a 429 with slow_down means traffic grew too fast, even inside your limits, while a 503 with server_is_overloaded means the model is temporarily overloaded — the 429 page and the 503 page cover the right response to each. Prompt caching gained a dashboard and generally available diagnostics that name the reason for a miss. Project API keys can now expire, administrators can restrict who creates keys, and a request can select regional processing through a prefixed domain. The Agents API entered public beta with a managed Codex harness, later adding computer use, and the Assistants API shut down. Lower-level changes landed too: mutual TLS and X.509 workload identity federation became generally available, and connections to api.openai.com can now use IPv6.

What didn't change

The long-context rule still reprices a whole request once input crosses 272K tokens. Rate-limit tables still differ by model group rather than scaling with price. And the gap between what Codex is named for and what it runs is still there: the CLI defaults to a general Sol model, not to the model with Codex in its name — why that is hasn't changed either.