GPT-5.3-Codex

TL;DR Highlight

OpenAI launched GPT-5.3-Codex, scoring 77.3 on Terminal-Bench 2.0 and directly competing with Anthropic's Opus 4.6 for coding agent supremacy.

Who Should Read

Developers choosing between frontier coding models, and ML engineers tracking the GPT vs Claude performance race.

Core Mechanics

GPT-5.3-Codex is OpenAI's latest specialized coding model, benchmarked at 77.3 on Terminal-Bench 2.0 — a test of long-horizon command-line task completion.
The model is explicitly positioned against Claude Opus 4.6, with OpenAI claiming edge on Terminal-Bench while Anthropic claims edge on SWE-bench.
Terminal-Bench 2.0 tests multi-step terminal tasks (file manipulation, system configuration, code execution) — highly relevant for agentic coding scenarios.
Both models are at the frontier tier with premium pricing, making the choice primarily about which specific task distributions they excel at.
The benchmarking war between OpenAI and Anthropic is heating up — both are optimizing specifically for the measures the other model is leading on.
Key differentiator: GPT-5.3-Codex reportedly handles more diverse programming language coverage, while Opus 4.6 has stronger contextual coherence over very long tasks.

Evidence

OpenAI published Terminal-Bench 2.0 results with methodology details, enabling independent verification.
HN commenters ran both models on their own tasks and found results depended heavily on task type — no universal winner.
The community debated whether Terminal-Bench 2.0 is a good proxy for real agentic performance, with some noting it was designed shortly after GPT-5.3-Codex's strengths were known.
Engineers building production agents noted the right answer is: test both on your actual task distribution, benchmark scores are starting points not conclusions.

How to Apply

Run a head-to-head on your specific workload: sample 20-30 representative tasks from your production use case and compare both models on those, not on public benchmarks.
Consider a routing approach: use one model for tasks where it clearly leads and the other for different task types — the cost overhead of dual-model routing may be worth the quality gain.
Watch the benchmark arms race critically: when a benchmark is published shortly after a model launch, evaluate whether it's measuring general capability or the specific things that model happens to be good at.
For long-horizon agentic tasks (multi-hour, multi-file): Opus 4.6's context coherence advantage may outweigh Terminal-Bench scores; test with actual task length.

Terminology

Terminal-Bench 2.0A benchmark measuring AI model performance on multi-step terminal and shell tasks — designed to evaluate agentic coding capability in command-line environments.

SWE-benchA benchmark where the model must fix real GitHub issues end-to-end — complementary to Terminal-Bench, measuring different aspects of coding agent capability.

Task distributionThe specific mix of task types in your real-world use case — benchmark scores on general evals may not predict performance on your specific distribution.