OpenAI has set a hard deadline: GPT-5.5 exits ChatGPT, ChatGPT Work, and Codex on October 14. This isn't a soft deprecation with a long tail. It's a production cutover, and if your team has Codex automations, CI/CD pipelines, or IDE integrations that explicitly target GPT-5.5, you have less than a month to act. The replacement lineup is already live: GPT-5-codex for agentic coding workflows, GPT-5.6 (Luna, Sol, and Terra variants) as the new ChatGPT backbone, and GPT-6 Astra for Pro and Plus subscribers who need maximum capability. Here's what changed, what it costs you operationally, and exactly what to do before October 14.

GPT-5.5 became the dominant workhorse for Codex's agentic coding capabilities after the earlier retirement of GPT-5.1 Codex Max and GPT-5.2 Codex variants. Engineers built around it because it was stable, capable, and predictable. That era ends in 29 days. According to OpenAI's Codex changelog, the replacement architecture splits across multiple models depending on your use case and subscription tier:

The headline addition is GPT-5-codex, a GPT-5 variant purpose-tuned for agentic coding. It becomes the default for cloud tasks and code review inside Codex, and it's selectable for local workflows through the Codex CLI and IDE extension. This isn't a rebrand of GPT-5.5 with a new name. It's a model trained specifically for tool use, multi-step code generation, and the kinds of autonomous agent loops that Codex runs in CI environments. GPT-6 Astra is the other major addition and carries the highest capability ceiling. Third-party coverage characterizes it as powerful but relatively expensive, and it remains gated behind higher-tier subscriptions. If your team is on Business or Enterprise, Astra is worth evaluating for complex refactoring and architecture-level tasks, but don't expect it to become a universal default anytime soon.

Most coverage will treat the GPT-5.5 retirement as a routine model upgrade story. That framing undersells the operational risk. Eighteen months ago, swapping AI models in your workflow meant changing a setting in a chat interface. In late 2026, Codex is embedded in CI pipelines, code review gates, IDE extensions, and automated workflows that run without human intervention. When OpenAI retires a model that underpins those systems, the event structurally resembles an API sunset: services break, behavior shifts silently, and teams scramble. The community reaction confirms this. An active OpenAI Community forum thread has already emerged with users asking OpenAI to retain GPT-5.5 as a legacy option for existing Codex Pro subscribers, even with usage caps. That request signals something important: engineering teams built operational dependencies on GPT-5.5 without building abstraction layers underneath it. When the model leaves, those dependencies are exposed. This is the pattern. GPT-5.1 Codex Max and GPT-5.2 Codex retired together earlier in 2026, replaced by GPT-5.5 and GPT-5.4 Codex Mini respectively. Teams that had just finished stabilizing on the 5.1/5.2 generation found themselves in another migration cycle. Now the 5.5 cycle ends. OpenAI is not slowing this rotation down. The model lifecycle for AI coding tools now runs faster than most teams' infrastructure release cycles. The implication is strategic, not just tactical: AI model dependencies must be treated like any other production infrastructure dependency, with version tracking, lifecycle monitoring, abstraction layers, and regression testing built in from day one.

The timing of this retirement also reveals something about OpenAI's competitive positioning against Anthropic's Claude Code and GitHub Copilot. Anthropic drew significant backlash earlier in 2026 when it attempted to push Claude Code users onto Anthropic-only model routing, limiting developer control over which models powered their coding workflows. OpenAI's Codex leadership responded by publicly arguing that developers should retain the ability to choose which models power their tools. GPT-5-codex being selectable across Codex CLI and IDE extensions, rather than forced as a single default, reflects that stated philosophy. But there's a tension here. OpenAI is simultaneously rotating models aggressively and gating its strongest options (GPT-6 Astra, GPT-5.6 Terra) behind higher subscription tiers. "Openness" that requires a Pro or Enterprise contract to access the best models is openness with conditions. Teams evaluating Codex against Claude Code should weigh the ecosystem flexibility argument against the reality that the most capable Codex models carry meaningful access overhead. GitHub Copilot, backed by its deep IDE integration and enterprise distribution, remains the volume play. But its model selection has historically lagged behind OpenAI's frontier capabilities. GPT-5-codex being purpose-tuned for agentic coding is a capability argument Copilot can't currently match with its default configuration. The honest competitive picture: Codex wins on capability ceiling and model flexibility for teams willing to manage the lifecycle complexity. Claude Code wins on model stability and workflow predictability for teams that prioritize fewer disruptions. The GPT-5.5 retirement accelerates that tradeoff becoming a real decision point for engineering leaders.

The migration work breaks into four distinct tracks. Run them in parallel, not sequentially, given the timeline.

Inventory every system that touches Codex or the ChatGPT API with explicit model targeting. Prioritize:

Any integration that doesn't explicitly specify a model may already be routed to newer defaults, but verify rather than assume.

Not every GPT-5.5 use case belongs on the same replacement. Use this logic:

The new models require updated versions of the desktop app, Codex CLI, and IDE extension. Don't assume your current installation supports GPT-5-codex or GPT-5.6 variants. Check the Codex changelog for minimum version requirements and push updates to your developer machines before October 14, not after.

This is the step most teams will skip and regret. Before the retirement date, run your critical Codex workflows against GPT-5-codex and your target GPT-5.6 variant in parallel with GPT-5.5. Diff the outputs. Measure:

Silent behavior shifts are the most dangerous migration outcome. A model that produces syntactically correct but subtly different code can introduce bugs that don't surface until production. Catch the differences now while you can compare against a known baseline.

The broader lesson from the GPT-5.1/5.2 retirement, the GPT-5.5 retirement, and every model rotation that follows is structural: teams that bind tightly to a specific model name in production code are accepting a recurring migration tax. The engineering response is straightforward in principle, even if the implementation takes work:

Teams that operate this way will treat October 14 as a scheduled maintenance window. Teams that don't will treat it as an incident.

The GPT-5.5 retirement is a useful stress test for how mature your team's relationship with AI tooling actually is. Immature adoption looks like individual engineers choosing their own Codex configurations with no organizational visibility into what's running where. Mature adoption looks like a platform engineering function that tracks model dependencies, owns the migration playbook, and treats OpenAI's lifecycle announcements the same way it treats AWS service deprecations. The model rotation cadence OpenAI is running is not going to slow down. If anything, the GPT-5.1 through 5.5 through 5-codex/5.6/6 progression that has played out across 2026 suggests the lifecycle will compress further as the capability curve steepens. The teams building organizational muscle around AI model governance now are building a durable competitive advantage. The teams treating each retirement as a surprise are paying a recurring operational tax that compounds. October 14 is the next deadline. Build the systems that make the one after it easier.