GPT-5.3-Codex

OpenAI introduced GPT-5.3-Codex on February 5, 2026 as its most capable agentic coding model, combining the frontier coding performance of GPT-5.2-Codex with the reasoning and professional knowledge of GPT-5.2 in a single model that is 25% faster. The Codex team used early versions to debug its own training, manage deployment, and diagnose evaluations, making it OpenAI's first model it describes as instrumental in creating itself. GPT-5.3-Codex sets new industry highs on SWE-Bench Pro and Terminal-Bench 2.0 while delivering strong performance on OSWorld and GDPval, reaching state-of-the-art results with fewer tokens than prior models. It supports mid-task steerability and live updates during long-running research, tool-use, and complex execution tasks, expanding Codex from a code-writing agent into a broader computer-use agent in the Codex app.

GPT-5.3-Codex moved to No. 1 in Quality on the Microsoft Foundry AI Model Leaderboard soon after release, while a cross-metric "podium" scoring method put GPT-5-Nano on top overall for efficiency.

GPT-5.3-Codex, OpenAI's latest agentic coding model, is now rolling out in GitHub Copilot. In early testing, GPT-5.3-Codex reaches new high scores on...

VentureBeat reported on February 5, 2026 that OpenAI released GPT-5.3-Codex simultaneously with Anthropic's Claude Opus 4.6, with Sam Altman calling it the most capable coding agent to date. OpenAI states the Codex team used early versions to debug training, manage deployment infrastructure, and diagnose evaluations, characterizing it as the first model instrumental in creating itself. GPT-5.3-Codex reportedly posts 57% on SWE-Bench Pro, 77.3% on Terminal-Bench 2.0, and 64% on OSWorld, using fewer than half the tokens of GPT-5.2-Codex for equivalent tasks while running 25% faster per token. The release added mid-task steerability and live updates, expanding Codex beyond code generation and review toward broader developer workflows during a week of competing launches.

DataCamp's analysis describes GPT-5.3-Codex as OpenAI's newest LLM replacing both GPT-5.2 and GPT-5.2-Codex, merging coding-agent and reasoning-model capabilities into a general-purpose agent. It pairs with the newly introduced Codex macOS app for interactive, context-preserving collaboration and nearly doubles the OSWorld-Verified score while setting new highs on SWE-Bench Pro and Terminal-Bench. OpenAI used GPT-5.3-Codex during its own development as a site-reliability-style agent, monitoring training runs, diagnosing results, and managing deployment, a step beyond models that merely generated their own training data. The launch emphasizes 25% faster inference from infrastructure and stack improvements, positioning the model for tasks beyond coding such as documentation, ticket updates, and pipeline management.

The llm-stats aggregator page for GPT-5.3 Codex, dated October 13, 2026, lists third-party benchmark scores including 0.81 on SWE-Lancer IC-Diamond, 0.78 on Cybersecurity CTFs, 0.77 on Terminal-Bench 2.0, and LiveBench results sourced from the model's official scorecard. Methodology notes xhigh effort evaluation across real-world freelance engineering, offensive security, and contamination-limited reasoning tasks. The page provides independent corroboration of GPT-5.3 Codex benchmark performance across categories including coding, cybersecurity, and reasoning, with confidence intervals shown for real-task performance. Scores are attributed to OpenAI's official blog and scorecard, serving as a third-party reference point alongside OpenAI's own launch benchmarks for the same model variant.

BenchLM's tracker entry for GPT-5.3 Codex, data as of September 3, 2026, lists the model as released February 5, 2026, with a 400K token context window, an aggregate capability score of 65.5 out of 100, and a rank of 42 of 231 tracked models. It reports API pricing of $1.75 per million input tokens and $14 per million The same BenchLM page publishes category-level percentiles for GPT-5.3 Codex against its eligible cohort: 86th percentile in Agentic (rank 21 of 144) and 83rd percentile in Coding (rank 26 of 149), while Reasoning, Knowledge, Math, Multilingual, Multimodal, and Instruction Following are not ranked. The tracker notes th