GPT‑5.3‑Codex‑Spark is OpenAI’s first real-time coding model for Codex, built for near-instant iteration. It delivers over 1,000 tokens per second, supports a 128k context window, and is available as a research preview for ChatGPT Pro users in the Codex app, CLI and VS Code extension. It’s tuned for fast, targeted edits rather than long, autonomous runs.

“Faster AI” only matters if it changes how you build.

OpenAI’s GPT‑5.3‑Codex‑Spark is designed to do exactly that: it’s the first Codex model built for real-time coding, where latency is low enough that you can iterate like you’re pairing with someone sitting next to you.

Released as a research preview on 12 February 2026, Codex‑Spark is a smaller version of GPT‑5.3‑Codex, optimised for ultra‑low latency inference—reportedly delivering more than 1,000 tokens per second, with a 128k context window at launch. It’s currently text-only.

What’s new: real-time coding inside Codex

Codex‑Spark is built for the moments where speed matters as much as intelligence:

• targeted edits,

• rapid refactors,

• quick interface iterations,

• and tight feedback loops where you want to steer mid-stream.

OpenAI notes that Codex‑Spark keeps its default working style lightweight: it prefers minimal, targeted changes and doesn’t run tests automatically unless you ask. That’s exactly what you want for fast iteration: fewer “big rewrites”, more precise loops.

Who can access GPT‑5.3‑Codex‑Spark?

At launch, Codex‑Spark is rolling out as a research preview for ChatGPT Pro users in the latest versions of:

• the Codex app,

• the Codex CLI,

• and the VS Code extension.

During the preview it uses separate rate limits, and OpenAI notes you may see limited access or queuing when demand is high.

Where Codex‑Spark fits in a modern dev workflow

Think of Codex‑Spark as the “fast lane” for coding:

Best for

• tight edit → run → refine loops

• small, safe refactors

• targeted bug fixes

• quick code generation and scaffolding

• interface tweaks and incremental improvements

Less ideal for

• long-running, multi-hour autonomous work

• complex architecture redesigns that need deeper reasoning

• high-risk changes without strong test coverage

In other words: use Spark to stay in flow, and switch to deeper models when the problem demands it.

How to prompt Codex‑Spark for better results

Speed doesn’t replace clarity. Strong prompts make the model faster and more accurate.

Use this simple structure:

1. Goal (what you want done)

2. Constraints (what must not change)

3. Context (file/module purpose, relevant APIs, expected behaviour)

4. Acceptance criteria (how you’ll verify it worked)

5. Output format (diff-style edits, list of changes, tests to run)

Prompt template (copy/paste)

“Make a targeted change to achieve [goal]. Do not change [constraints]. Context: [brief]. Acceptance criteria: [bullets]. Output: propose the smallest diff, then list tests I should run.”

Example prompts

• “Refactor this function to remove duplication without changing behaviour. Keep the public API identical. Output the smallest diff and add unit tests for the edge cases.”

• “Fix the bug where null values crash the parser. Don’t change the parsing rules. Provide a minimal patch and a regression test.”

• “Improve performance for this loop. Keep output identical. Suggest micro-optimisations first; only restructure if needed.”

Practical steps: getting value in week one

If you want early wins, focus on measurable “developer minutes saved” rather than chasing perfect end-to-end automation.

Step 1: Choose one workflow to accelerate

Pick something frequent:

• code review fixes,

• small refactors,

• test generation for a module,

• documentation updates,

• repetitive UI changes.

Step 2: Standardise your prompting patterns

Create a shared set of “known-good” prompts (bugfix, refactor, tests, docs) and publish them.

Step 3: Add lightweight guardrails

Require:

• diffs rather than full rewrites,

• explicit acceptance criteria,

• and a test plan for any functional change.

Step 4: Track outcomes

Measure:

• cycle time for small tasks,

• PR turnaround time,

• fewer review iterations,

• and reduction in repetitive work.

How teams can operationalise adoption

Codex‑Spark becomes more valuable when teams use it consistently and safely.

• Use Asana to run an adoption plan (owners, rollout waves, metrics).

• Use Miro to map your “AI-assisted dev workflow” and agree review gates.

• Use Notion to host prompt libraries, coding standards, and examples.

• Use Glean to make those standards searchable so people (and AI) can find the right patterns fast.

Summary

GPT‑5.3‑Codex‑Spark is a meaningful shift: real-time coding assistance with near-instant generation speed and long context, designed for interactive iteration rather than long autonomous runs.

If your teams want to move from occasional AI usage to consistent productivity gains, the biggest lever is operational: standard prompts, small safe workflows, and measurable outcomes.

Next steps

• Try Codex‑Spark on one repeatable task category.

• Save the best prompt as a team template.

• Require acceptance criteria and a test plan.

• Measure time saved, then scale.

FAQ

Q1: What is GPT‑5.3‑Codex‑Spark?
GPT‑5.3‑Codex‑Spark is OpenAI’s real-time coding model for Codex, optimised for ultra-low latency and rapid iteration. It’s a smaller version of GPT‑5.3‑Codex.

Q2: Who can access GPT‑5.3‑Codex‑Spark?
It’s currently available as a research preview for ChatGPT Pro users in the Codex app, Codex CLI and VS Code extension.

Q3: What are the headline improvements?
OpenAI positions Codex‑Spark as 15× faster than GPT‑5.3‑Codex for generation speed, with a 128k context window and near-instant responses.

Q4: Is Codex‑Spark available via the API?
OpenAI has stated it’s available in the API for a small set of design partners during the research preview.

Q5: What’s the best way to use Codex‑Spark?
Use it for fast, targeted edits and quick iteration loops. Ask explicitly for tests to be run or for a test plan, since the model defaults to lightweight edits.