GPT-5.3-Codex has been OpenAI’s primary agentic coding model since February 5, 2026. On May 17 it became the base model for all GitHub Copilot Business and Enterprise organizations, replacing GPT-4.1 as the default. On June 5, GPT-5.2-Codex was removed from most Copilot model pickers.

And on July 23, 2026, every earlier Codex API model was originally scheduled to shut down.

If you have any production calls to gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.2-codex, or gpt-5.1-codex-mini, those calls fail in 51 days.

Correction, 2026-08-22: An earlier version of this update (dated 2026-07-28) incorrectly stated that OpenAI had pushed this shutdown back to October 23, 2026. Re-checking the live OpenAI deprecations page directly: the five Codex model IDs below remain listed under the July 23, 2026 shutdown section (replacement gpt-5.6-sol, or gpt-5.6-terra for gpt-5.1-codex-mini), part of the GPT-5.6 family, which OpenAI first previewed in late June 2026 and moved to general availability on July 9, 2026. October 23, 2026 is a separate, later deprecation wave covering different legacy model snapshots (gpt-4, gpt-4o-2024-05-13, o1, o3-mini, o4-mini) — it does not apply to these Codex IDs. That means the July 23 shutdown for gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini, and gpt-5.2-codex already happened as originally scheduled: if you still have production calls to any of these model IDs, they are returning errors now, not on some later date. Move to gpt-5.3-codex or newer immediately.

What Is Being Shut Down

The following model IDs reached end-of-life on July 23, 2026, per OpenAI’s deprecations page (re-verified live 2026-08-22 — see the correction note above; this date did not move):

Model IDStatus (as of original publication)
gpt-5-codexShutting down July 23
gpt-5.1-codexShutting down July 23
gpt-5.1-codex-maxShutting down July 23
gpt-5.1-codex-miniShutting down July 23
gpt-5.2-codexShutting down July 23

The legacy codex-mini-latest model was removed from the API on February 12, 2026, replaced by gpt-5-codex-mini.

gpt-5.3-codex is not on this list. It has a long-term support commitment in GitHub Copilot Business and Enterprise through February 4, 2027.

The Migration

Change the model string:

# Before (any of these)
response = client.responses.create(
    model="gpt-5.2-codex",  # or gpt-5.1-codex, gpt-5-codex, etc.
    ...
)

# After
response = client.responses.create(
    model="gpt-5.3-codex",
    ...
)

GPT-5.3-Codex is backward-compatible with the same API surfaces as earlier Codex models. Existing prompts, tool definitions, and system messages transfer without modification.

The one thing to check: if your code passes a reasoning_effort parameter, gpt-5.3-codex accepts "low", "medium", "high", and "xhigh". The older models accepted a narrower set of values. If you were passing an unsupported string, you may get a different behavior — validate your reasoning_effort values after migrating.

What GPT-5.3-Codex Is

GPT-5.3-Codex combines two training stacks that were previously separate: the specialized coding training from the Codex lineage (GPT-5.2-Codex) and the broader reasoning and professional knowledge from the GPT-5.2 base. Earlier Codex models were optimized narrowly for code generation. GPT-5.3-Codex handles the full agentic coding loop — research, planning, writing, running, debugging — without needing a separate orchestration model for non-code steps.

OpenAI describes the model as working “much like a colleague” during interactive sessions: you can steer it during long multi-file tasks without losing context, ask it to pause and replan, or hand off partial work to a human reviewer mid-task.

The model was used during its own development — OpenAI describes GPT-5.3-Codex as the first model in the family “instrumental in creating itself,” with early versions used to debug training runs, optimize deployment infrastructure, and build data pipelines for model evaluation.

Capabilities Worth Knowing

Reasoning Effort Settings

GPT-5.3-Codex supports four reasoning effort levels via the API:

response = client.responses.create(
    model="gpt-5.3-codex",
    reasoning_effort="high",  # "low", "medium", "high", "xhigh"
    ...
)
  • low — fast iteration, good for autocomplete-style tasks and short edits
  • medium — balanced; default for most API calls
  • high — longer reasoning; appropriate for architecture decisions and complex multi-file refactors
  • xhigh — maximum reasoning budget; use for tasks where correctness matters more than latency (security audits, migration planning, critical path debugging)

Latency and cost scale with reasoning effort. For most build loops, medium is the right starting point. Use xhigh sparingly — it can take several minutes for complex tasks.

Computer Use

GPT-5.3-Codex can interact with a desktop through screenshots, scoring 64.7% on OSWorld-Verified — a benchmark for vision-based desktop task completion. This applies to coding tasks involving legacy tooling that lacks an API: filling forms in a web IDE, clicking through a UI-only deployment pipeline, or navigating a GUI-based database tool. The model accepts screenshot inputs and outputs structured actions or Playwright scripts. Note that OpenAI’s subsequent GPT-5.4 release is billed as the first general-purpose OpenAI model with native computer-use built in, and is a stronger choice specifically for computer-use-heavy workflows.

response = client.responses.create(
    model="gpt-5.3-codex",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Click the 'Deploy to Production' button"},
                {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
            ]
        }
    ]
)

Terminal-Bench Performance

GPT-5.3-Codex scores 77.3% on Terminal-Bench 2.0, a benchmark for command-line agentic tasks — shell scripting, build systems, package management, file manipulation pipelines — up from GPT-5.2-Codex’s 64.0%. That was the top score on the benchmark at GPT-5.3-Codex’s February 2026 launch; GPT-5.5, released in April 2026, has since taken the top spot at 82.7%. If your agent runs in a terminal or interacts with shell commands as a core workflow, benchmark performance here is directly relevant.

Specifications

Per OpenAI’s Codex model documentation and GitHub’s Copilot pricing docs:

ParameterValue
Model IDgpt-5.3-codex
Input context400,000 tokens
Max output128,000 tokens
Input pricing$1.75 / MTok
Cached input$0.175 / MTok
Output pricing$14.00 / MTok
Reasoning effortlow, medium, high, xhigh
API availabilityAll paid tiers (Tier 1–5)
Supported endpointsResponses API, Chat Completions, Batch, Fine-tuning, Assistants

Pricing is identical to GPT-5.2-Codex. Migrating does not change your API costs.

GitHub Copilot Status

GPT-5.3-Codex is now the default model across GitHub Copilot. Timeline of changes:

For teams using Copilot via GitHub’s UI, this migration happened automatically on May 17. If you had explicit model selection set to gpt-5.2-codex in VS Code or JetBrains settings, check whether those overrides are still active — they may cause inconsistent behavior between your IDE and the Copilot API.

GPT-5.3-Codex vs. GPT-5.5

GPT-5.5 is OpenAI’s current top model. OpenAI’s Codex documentation lists GPT-5.5 among its “Recommended models” for complex coding, computer use, knowledge work, and research workflows, with GPT-5.3-Codex positioned as the Codex-tuned alternative for coding-heavy sessions.

The distinction is specialization vs. breadth:

  • GPT-5.5 handles coding plus everything else (writing, analysis, multimodal tasks, professional documents). It is the right choice when your agent does coding alongside non-coding work.
  • GPT-5.3-Codex is optimized for the agentic coding loop specifically. On Terminal-Bench 2.0 it trails GPT-5.5 (77.3% vs. 82.7%) but at a lower per-token input cost ($1.75 vs. GPT-5.5’s $5.00).

If your agent is exclusively writing, running, and debugging code — no document generation, no multimodal input — GPT-5.3-Codex remains cost-competitive at the task it was designed for.

GPT-5.3-Codex-Spark

OpenAI has released gpt-5.3-codex-spark as a research preview. It is optimized for near-instant coding iteration at ultra-low latency (1,000+ tokens/sec via Cerebras hardware). Currently available to ChatGPT Pro users and in the Codex app, Codex CLI, and VS Code extension — no general API access as of June 2026. Watch this space for a broader rollout.

Builder Actions

  1. Audit your API calls now. Search your codebase for gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini, gpt-5.2-codex. Every instance of these model IDs breaks on July 23.
  2. Swap to gpt-5.3-codex. No other code changes required for most workloads.
  3. Validate reasoning effort values. If you pass a reasoning_effort parameter, confirm your value is one of "low", "medium", "high", or "xhigh".
  4. Check Copilot IDE settings. If you have explicit model overrides in VS Code, JetBrains, or CLI settings, verify they point to gpt-5.3-codex or gpt-5.5.
  5. Evaluate GPT-5.5 if you are on a mixed workload. For agents that combine coding with document generation, analysis, or multimodal inputs, GPT-5.5 is now recommended as the primary model.

At original publication, July 23 was 51 days out. That date held — per the correction note above, the shutdown happened as scheduled and is not October 23. If you haven’t migrated yet, the affected model IDs are already returning errors. The migration is still a one-line change.