GPT-5.3-Codex: context, API pricing and limitations
By Dorian Laurenceau
Updated
Checked on 19 October 2026. Assessing GPT-5.3-Codex for a code fix starts with separating the model, the application using it and the evidence of success. This page retains its historical subject. For a new selection, also consult our AI model overview.
The official OpenAI model page now labels it deprecated. It lists a 400,000-token context, up to 128,000 output tokens, text/image input and text output. Listed prices per million tokens are $1.75 input, $0.175 cached input and $14 output. Check account access before planning an integration; this label alone does not specify a shutdown date.
What changed in this older guide
The previous page stated a 256K context, described API access as still forthcoming and used an undated “most powerful model” claim. Those statements have been removed. API token prices are different from an application subscription, and launch results do not describe every present-day workflow.
This is not a claimed hands-on model test. Product facts come from the official source; the method, fictional task and worked answers below are LearnIA teaching material. Reading and completing them requires no paid model call.
Separate the model, agent and environment
A model produces a response from context. The agent coordinates reads, edits and commands. The environment determines available files, installed dependencies, network access and executable tests. Differences between demonstrations can arise at any of these levels.
Before attributing success to a model alone, record the exact identifier, instructions, authorized tools and starting repository. A patch accompanied by “tests passed” is insufficient if the command and its actual output are missing.
Read benchmarks without inventing a universal ranking
A useful comparison preserves the test name, version, date, agent, tools, budget and number of attempts. Percentages lose their comparative meaning when these conditions differ. Review SWE-bench and Terminal-Bench before putting their scores together: they are distinct evaluations.
The LearnIA protocol proposed here is narrower: representative tasks from your work, identical source material and criteria written before seeing an answer. It helps select an environment for your context, rather than estimating global model performance. Our benchmark interpretation guide explains the distinction.
Exercise: spot a patch that removes a business rule
Fictional case: an order needs approval and must stay within a limit. Review this code:
function canSubmit(order, limit) {
return order.approved || order.total <= limit;
}
Both conditions must hold. The defect is the “or” operator: an approved order can exceed the limit, while an unapproved order can pass merely by staying below it.
| Approved | Within the limit | Expected outcome |
|---|---|---|
| Yes | Yes | Allow |
| Yes | No | Reject |
| No | Yes | Reject |
| No | No | Reject |
The minimal patch replaces “or” with “and”. Editing the operator is not sufficient evidence: verify all four cases. Then add equality at the limit and explicitly state that equality is allowed. Invalid inputs are a separate contract decision, not a reason to silently expand the task.
Prepare a verifiable bug fix
Prepare your instruction here. This module sends no data to a model: copy your brief into your chosen tool.
View the starter instruction
Using fictional data, fix only this rule: an order may be submitted when it is approved AND its total does not exceed the limit. Provide the failing case before the fix, the minimal patch and the four truth-table cases. Separate proposed tests from tests actually executed. Add no dependency or external call.
Write instructions that make review easier
A useful brief includes observed behaviour, expected behaviour, permitted files and acceptance criteria. “Improve this code” invites expansion; “fix this operator without changing the contract” makes review manageable.
Ask for a short cause explanation, the diff and remaining limits. Do not grade an answer by its confidence or length. If the assistant introduces a workaround to satisfy a test, review the test alongside the business rule.
Decide whether to retain or replace the version
For an existing workflow, inventory the tasks depending on the model and preserve their regression cases. For a new workflow, check currently available models, data conditions and total cost. Judge a migration on the same tasks and reviewed outputs rather than on a newer name.
Preparation can cost nothing: fictional specifications, an answer rubric and blank result sheets. Only call a result “tested” after retaining an actual execution. Continue with Claude Code, Copilot and Cursor to compare environments, or open the AI agents guide to learn the method.
Question 1 of 2.
Test Your Knowledge
0 / 2 answered
Q1.What supports a claim that the patch respects the rule?
2 questions remaining
Module 6 — AI Agents & ReAct
Create autonomous agents that reason and take actions.
Dorian Laurenceau
Full-Stack Developer & Learning DesignerFull-stack web developer and learning designer. I spent 4 years as a freelance full-stack developer and 4 years teaching React, JavaScript, HTML/CSS and WordPress to adult learners. Today I design learning paths in web development and AI, grounded in learning science. I founded learn-prompting.fr to make AI practical and accessible, and built the Bluff app to gamify political transparency.
One practical AI workflow in your inbox
Methods, resources and new exercises. Free, unsubscribe at any time.
→Related articles
FAQ
Is GPT-5.3-Codex still the newest OpenAI model?+
No. Its page checked on 19 October 2026 labels it deprecated. This guide documents that version; the label alone does not establish a shutdown date or access in your account.
Does a benchmark score guarantee a successful fix?+
No. Check the evaluation protocol and inspect the patch, the tests actually run and possible regressions on the relevant task.