Introducing GPT-5.3-Codex

TL;DR

GPT-5.3-Codex is OpenAI’s closed-weights agentic coding model, released 2026-02-05, positioned as a single model that subsumes both GPT-5.2-Codex’s frontier coding performance and GPT-5.2’s general reasoning, while running 25% faster. The release post claims SOTA on SWE-Bench Pro and Terminal-Bench 2.0 plus strong showings on OSWorld (computer-use) and GDPval (knowledge-work), with the notable framing that “Codex goes from an agent that can write and review code to an agent that can do nearly anything developers and professionals can do on a computer.” It is the first OpenAI model trained partly with assistance from earlier checkpoints of itself (Codex used to debug Codex training, monitor deployment, and analyze alpha-test telemetry), and the first OpenAI launch classified High-capability on cybersecurity under the Preparedness Framework. Available in the Codex app/CLI/IDE/web with paid ChatGPT plans only; API access is “coming soon.” For Luma, the relevant datapoint is the consolidation of long-horizon coding + general reasoning + computer-use into one closed model and the “model that helps build itself” claim, both of which set the bar that the open-agentic-SWE cluster (Agentic Software Engineering) is being measured against.

Key claims

GPT-5.3-Codex is positioned as a single unified model that combines GPT-5.2-Codex’s coding ability with GPT-5.2’s reasoning and professional-knowledge capability, while being 25% faster than its predecessor [§“Introducing GPT-5.3-Codex”].
Reports state-of-the-art on SWE-Bench Pro (a 4-language, contamination-resistant, industry-relevant SWE benchmark distinct from SWE-Bench Verified’s Python-only setup) and far exceeds prior SOTA on Terminal-Bench 2.0, doing so “with fewer tokens than any prior model” [§“Setting a new standard on SWE-Bench Pro and Terminal-Bench”].
Benchmark table (selected, reading down from the body’s six rows of pairs of percentages): GPT-5.3-Codex 56.8/77.3/64.7/70.9/77.6/81.4 vs comparison columns at 56.4/64.0/38.2/—/67.4/76.0 and 55.6/62.2/37.9/70.9 (high)/67.7/74.6 across SWE-Bench Pro, Terminal-Bench 2.0, OSWorld, and GDPval-style splits [§Benchmark table at end].
First OpenAI model trained with material assistance from earlier versions of itself: alpha checkpoints debugged the training run, monitored deployment, identified context-rendering bugs, root-caused low cache-hit rates, dynamically scaled GPU clusters during traffic surges, and built ad-hoc data-pipeline/visualization tools that replaced the team’s standard dashboarding [§“Codex used to develop Codex”].
Interaction model is shifted from “wait for final output” to interruptible, real-time co-work — the model emits frequent progress updates and accepts mid-task feedback without losing context [§“A more capable, more interactive agent”].
Aesthetic / web-dev default behavior is upgraded: under-specified prompts default to sites with “sensible defaults” (e.g. yearly pricing rendered as discounted monthly with the discount made explicit; automatically transitioning testimonial carousels) [§“Better default web aesthetics”].
Demonstrates long-horizon autonomous iteration: when given a “develop web game” skill and generic follow-ups (“fix the bug”, “improve the game”), GPT-5.3-Codex iterated on two games (racing-v2 and a diving game) over “millions of tokens” across “the course of days” [§“Web development and long-running agentic capabilities”].
Computer-use claim: stronger performance on OSWorld than prior GPT models, framed as a step-change toward “a single, general-purpose agent that can reason, build, and execute” rather than a code-only agent [§“Beyond code: agentic computer use”].
Matches GPT-5.2 on GDPval (professional knowledge-work eval across 44 occupations) using custom skills, despite being optimized for coding — i.e. coding-specialization is reported not to cost knowledge-work generality [§“Professional knowledge work”].
First OpenAI launch classified High capability on cybersecurity under the Preparedness Framework, and the first directly trained to identify software vulnerabilities; routes some requests detected as high-risk to GPT-5.2 as a mitigation; paired with $10M API-credits commitment for cyber defense and a “Trusted Access for Cyber” pilot [§“Cybersecurity”].
Co-designed for and served on NVIDIA GB200 NVL72 systems [§“Hardware partnership”].
Available on paid ChatGPT plans across Codex app, CLI, IDE extension, and web; API access “coming soon” at the time of post [§“Availability”].

Method

The post is a product announcement, not a technical report — no training-data, architecture, or post-training-recipe details are disclosed. The most concrete methodological breadcrumbs are: (a) the model is the first OpenAI release whose training was directly assisted by earlier checkpoints of itself, with named workflows including training-run debugging, deployment monitoring, dynamic GPU-cluster scaling, alpha-test telemetry analysis via Codex-authored regex classifiers, and ad-hoc data-pipeline construction; (b) the system is evaluated on a fixed quadruple of benchmarks (SWE-Bench Pro, Terminal-Bench 2.0, OSWorld, GDPval) framed as “coding, agentic, and real-world capabilities”; and (c) deployment-time mitigations for cybersecurity include automatic routing of high-risk-detected requests away from GPT-5.3-Codex to GPT-5.2, plus an opt-in “Trusted Access for Cyber” program for legitimate security researchers. Underlying training methodology — data synthesis, RL recipe, scaffold diversity, verifier design — is not described. See the system card for the safety-eval details (not filed here; pointer only).

Results

The body of the post quotes one table of six benchmark rows × three model columns (presumed GPT-5.3-Codex / GPT-5.2-Codex / GPT-5.2 or similar), with values 56.8/56.4/55.6, 77.3/64.0/62.2, 64.7/38.2/37.9, 70.9/—/70.9 (high), 77.6/67.4/67.7, 81.4/76.0/74.6. Without the row labels (lost in the fetch), the directionally robust statements are: GPT-5.3-Codex wins on every row; the largest absolute gaps are at the second and third rows (≈13 and ≈26 points respectively) — consistent with the post’s verbal claim of large gains on Terminal-Bench and on a third agentic axis (likely OSWorld). On SWE-Bench Pro the headline number is 56.8 with a marginal ≈0.4 lead over the comparison column. The 70.9 row matches between GPT-5.3-Codex and the third column (at “high” reasoning), suggesting one benchmark where GPT-5.2 already saturates. Other reported wins: long-horizon game-development autonomy (“over the course of days”, “millions of tokens”); 25% lower latency vs predecessor; GDPval parity with GPT-5.2 despite coding specialization; first model to pass all three cybersecurity Preparedness thresholds (per the linked system card, not the post body).

Why it’s interesting

The release closes the loop the open-agentic-SWE cluster (Agentic Software Engineering) has been racing against: OpenAI is now framing the headline benchmark as SWE-Bench Pro rather than SWE-Bench Verified, exactly the move Qwen3-Coder-Next made when filing its own positioning (Qwen3-Coder-Next Technical Report) — i.e. SWE-Bench Pro is now the consensus arena. The “model helps build itself” claim is the most interesting research signal in an otherwise marketing-shaped post: it is the closed-source analog of the closed-loop self-synthesis pattern that InfTool (Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing) and Qwen3-Coder-Next (Qwen3-Coder-Next Technical Report) document at the data-generation level, but extended one layer up to training-process automation (debugging the training run itself, analyzing alpha-test telemetry, building bespoke evaluation dashboards). For Luma, the relevant track to watch is whether the consolidation of “long-horizon coding + general reasoning + computer-use” into a single model is replicable in the open: the open-cluster has matching SWE-Bench-Pro numbers but not yet matching computer-use (OSWorld) numbers, and no filed open model claims GDPval parity. The cybersecurity High-capability classification is also the first concrete instance of an OpenAI deployment activating cyber-domain mitigations, which is a precedent worth tracking for any downstream agent Luma ships.