What are the strengths of this model?

Advanced coding capabilities Long context grasp of 128K Latest development by OpenAI

What are the weaknesses of this model?

Private closed license Instability due to research preview status Non-disclosure of source code

What are the best use cases?

Automatic generation of complex code Analysis of large-scale projects Efficient debugging work

Back to Models

OpenAIProprietary

GPT-5.3-Codex-Spark (Research Preview)

Name: GPT-5.3-Codex-Spark (Research Preview)
Author: OpenAI

GPT-5.3-Codex-Spark (Research Preview) is a coding-focused AI model developed by OpenAI. It is equipped with a 128K context window and is aimed at providing advanced programming assistance.

Parameters

Undisclosed

Context Window

128K

License

Proprietary

Release Date

2026-02-12

API Pricing

API pricing for this model is not yet available

Strengths

・Advanced coding capabilities
・Long context grasp of 128K
・Latest development by OpenAI

Weaknesses

・Private closed license
・Instability due to research preview status
・Non-disclosure of source code

Use Cases

・Automatic generation of complex code
・Analysis of large-scale projects
・Efficient debugging work

Deep Analysis

Release Date

February 12, 2026

Context Window

128K tokens

Speed

1000+ tokens/second

Speed Improvement

15x faster than GPT-5.3 Codex flagship

Hardware

Cerebras WSE-3 (Wafer Scale Engine 3)

Terminal-Bench 2.0

77.3%

vs GPT-5.2-Codex 64%

Pricing

No change from standard GPT-5.3-Codex

Availability

ChatGPT Pro subscribers, API coming soon

Strengths

・Blazing fast at 1000+ tokens per second — 15x faster than flagship GPT-5.3 Codex
・First OpenAI model running on Cerebras WSE-3 hardware with 4 trillion transistors
・Terminal-Bench 2.0 score of 77.3% — higher than GPT-5.2-Codex (64%)
・Enables real-time steering — developers can interrupt and redirect mid-generation
・80% reduction in round-trip time via persistent WebSocket connections

Weaknesses

・Scores lower on SWE-Bench Pro and Terminal-Bench 2.0 than flagship GPT-5.3 Codex
・May struggle with complex, multi-file architecture changes
・Does NOT meet 'High capability' cybersecurity threshold under OpenAI's Preparedness Framework
・Not suitable for sensitive security logic or autonomous authentication tasks
・Research preview — limited availability (ChatGPT Pro only, API 'coming soon')

Competitor Comparison

Model	Arena	SWE	GPQA	Price
GPT-5.3 Codex (flagship)	N/A	~80%	~90%	$1.75/$14 per 1M tokens
GPT-5.2-Codex	N/A	~75%	~88%	$1.75/$14 per 1M tokens
Claude Sonnet 4.5	~1420	~72%	~89%	$3/$15 per 1M tokens
Cursor (varies)	N/A	Varies	N/A	$20/month subscription

Overview

GPT-5.3-Codex-Spark is a research preview released February 12, 2026, delivering unprecedented coding speed at 1000+ tokens per second — 15x faster than the flagship GPT-5.3 Codex. It is the first OpenAI model to run on Cerebras WSE-3 hardware and achieves Terminal-Bench 2.0 score of 77.3%. It trades reasoning depth for speed, enabling real-time collaborative coding workflows.

Benchmarks & Performance

GPT-5.3-Codex-Spark achieves Terminal-Bench 2.0 accuracy of 77.3% (vs GPT-5.2-Codex's 64%) while running 15x faster at 1000+ tokens/second. It consumes approximately half the tokens of earlier models for certain output types. The model surpasses reasoning capabilities of both GPT-5.2-Codex and standard GPT-5.2. However, it scores lower than the flagship GPT-5.3 Codex on SWE-Bench Pro and Terminal-Bench 2.0. Software optimizations include 80% RTT reduction, 50% TTFT improvement, and 30% per-token overhead reduction via persistent WebSocket connections.

Detailed Comparison

GPT-5.3-Codex-Spark is 15x faster than the flagship GPT-5.3 Codex but trades some reasoning depth for speed. It outperforms GPT-5.2-Codex on Terminal-Bench 2.0 (77.3% vs 64%) while being dramatically faster. It does not meet the 'High capability' cybersecurity threshold that the flagship achieves. Compared to Claude Sonnet 4.5, it offers much higher throughput but may trail on complex multi-step reasoning. The Cerebras hardware advantage eliminates inter-GPU bottlenecks.

Community Feedback

The developer community has been highly enthusiastic about the 1000+ tok/s speed, describing it as shifting the workflow from batch-processing to live pair-programming. Early users praise the real-time steering capability (interrupt and redirect mid-generation). Some concern exists about the research preview status and limited availability. The self-development aspect (Spark helped create itself) has generated significant discussion about AI development acceleration.

Use Cases

Best for: rapid prototyping and iterative coding, real-time collaborative coding sessions, fast code generation and scaffolding, debugging with immediate feedback loops, quick code reviews, and interactive development workflows. Also good for: PRD writing, copy editing, user research, test generation, and metrics analysis. NOT suitable for: complex multi-file architecture changes, sensitive security code, autonomous authentication tasks, or tasks requiring maximum reasoning depth. Use GPT-5.3 Codex flagship for deep reasoning work.

Latest News

Released February 12, 2026 as a research preview. Available to ChatGPT Pro subscribers via the Codex App (select 'Spark'), VS Code extension, and CLI (`codex --model gpt-5.3-codex-spark`). API access is 'coming soon.' The model is described as 'the first step toward a Codex that works in two complementary modes: real-time collaboration and long-running tasks.'

Sources

Analysis generated: 2026-05-24