GPT-5.3 Codex

GPT-5.3 Codex (officially styled GPT-5.3-Codex) is a frontier agentic coding model released by OpenAI on February 5, 2026, the same day as Anthropic's Claude Opus 4.6. OpenAI described it at launch as its most capable agentic coding model to date, combining the coding performance of GPT-5.2-Codex with the reasoning and professional knowledge of GPT-5.2 in a single model that also ran 25% faster (Source: openai.com). OpenAI stated it was the first of its models "instrumental in creating itself," and it was the first OpenAI model treated as High capability in the Cybersecurity domain under the company's Preparedness Framework (Source: GPT-5.3-Codex System Card).

Lineage and history

GPT-5.3 Codex extended OpenAI's Codex line of agentic coding models: GPT-5.2-Codex was introduced on December 18, 2025 (Source: openai.com), and a standalone Codex desktop app followed on February 2, 2026, three days before the model itself (Source: openai.com). At launch OpenAI positioned the model as a step 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" (Source: openai.com).

OpenAI stated that GPT-5.3 Codex "was instrumental in creating itself," reporting that early versions were used to debug its own training, manage deployment, and diagnose evaluations (Source: shumer.dev, The Shape of the Thing). The launch post gives specifics: the research team used Codex to monitor and debug the model's training run, track patterns during training, analyze interaction quality, and build applications for comparing the model's behavior against prior models; the engineering team used it to adapt the agent harness, identify context-rendering bugs, root-cause low cache-hit rates, and dynamically scale GPU clusters during the launch; a researcher had it write regex classifiers and run them over session logs during alpha testing, and a data scientist co-built data pipelines with it to analyze unusual alpha-test results (Source: openai.com). The system card characterizes this as the first confirmed public case of a deployed model being materially used to build the next generation of itself, a framing OpenAI presents as a recursive self-improvement milestone (Source: GPT-5.3-Codex System Card).

Within the GPT-5.x line, GPT-5.4, released in March 2026, was described by OpenAI as combining "the coding strengths of GPT-5.3-Codex with leading knowledge work and computer-use capabilities" (Source: openai.com), and GPT-5.4 Thinking was the first general-purpose OpenAI model to ship with the High-Cybersecurity mitigations first used for GPT-5.3 Codex (Source: GPT-5.4 Thinking System Card). The GPT-5.x coding line continued with the June 2026 GPT-5.6 family, whose Sol tier OpenAI positioned for the extended coding and security tasks GPT-5.3 Codex had targeted.

Architecture and training

OpenAI did not disclose the model's parameter count or architecture. The company stated that GPT-5.3 Codex was "co-designed for, trained with, and served on" NVIDIA GB200 NVL72 systems, and that infrastructure and inference-stack improvements let it serve the model 25% faster for Codex users (Source: openai.com).

Two training details are documented. First, it was the first OpenAI model directly trained to identify software vulnerabilities (Source: openai.com). Second, it was trained with a "user model" that made conflicting edits during reinforcement learning, with positive reinforcement for preserving user changes, through which it learned to handle user changes gracefully (Source: GPT-5.3-Codex System Card).

Capabilities and benchmarks

GPT-5.3 Codex combines the coding performance of GPT-5.2-Codex with the reasoning and professional knowledge of GPT-5.2, enabling long-running tasks involving research, tool use, and complex execution; users can steer and interact with the model while it works without losing context, and it provides progress updates during a task rather than only a final output (Source: openai.com).

Self-reported benchmarks

OpenAI reported the following scores, all run at xhigh reasoning effort (Source: openai.com, February 5, 2026; self-reported):

OpenAI characterized the SWE-Bench Pro and Terminal-Bench 2.0 results as new industry highs, noting that SWE-Bench Pro spans four programming languages where SWE-bench Verified tests only Python, and that the model reached its Terminal-Bench score using fewer tokens than any prior model (Source: openai.com). On OSWorld-Verified, an agentic computer-use benchmark on which humans score roughly 72%, OpenAI stated the model demonstrates far stronger computer-use capabilities than previous GPT models. On GDPval, an evaluation OpenAI released in 2025 that measures performance on well-specified knowledge-work tasks across 44 occupations (presentations, spreadsheets, and other work products), the model matched GPT-5.2 when using custom skills similar to those used for OpenAI's earlier GDPval results (Source: openai.com).

Beyond benchmarks, OpenAI demonstrated long-horizon web development — the model iterated autonomously over millions of tokens to build two browser games (a racing game and a diving game) from generic follow-up prompts — and described support for the full software lifecycle: debugging, deploying, monitoring, writing PRDs, editing copy, user research, tests, and metrics, as well as non-coding outputs such as slide decks and spreadsheet analysis (Source: openai.com).

Third-party measurements

METR added GPT-5.3 Codex to its task-completion time-horizon measurements on February 20, 2026, estimating a 50%-time-horizon of around 6.5 hours (95% confidence interval of 3 to 17 hours) with reasoning effort set to high rather than the xhigh setting OpenAI used for its published scores (Source: x.com; metr.org). On the Terminal-Bench 2.0 public leaderboard under the shared Terminus 2 scaffold, GPT-5.3 Codex scored 65% against Claude Opus 4.6's 63% — for GPT-5.3 Codex, lower than the 77.3% OpenAI reported with its own harness at xhigh effort (Source: lesswrong.com, February 16, 2026).

Availability and pricing

At launch the model was available on paid ChatGPT plans everywhere Codex runs — the Codex app, CLI, IDE extension, and web — with OpenAI stating it was "working to safely enable API access soon" (Source: openai.com). As of mid-2026, OpenAI's API price list carries gpt-5.3-codex under its specialized models at $1.75 per million input tokens ($0.175 cached) and $14.00 per million output tokens (Source: developers.openai.com, accessed July 2026), with a 400,000-token context window (Source: openrouter.ai).

On March 18, 2026, GitHub announced that GPT-5.3-Codex would become the base model for all Copilot Business and Copilot Enterprise organizations, replacing GPT-4.1; the change took effect May 17, 2026. GitHub designated it the first long-term support (LTS) model in partnership with OpenAI, guaranteed to remain available for twelve months from launch — through February 4, 2027 — to give enterprises stability for internal security and safety reviews. It carries a 1x premium request multiplier (GPT-4.1 remained force-enabled at 0x pending deprecation alongside usage-based billing on June 1, 2026), and GitHub stated its Copilot data showed the model had a "significantly high code survival rate" among enterprise customers (Source: github.blog).

GPT-5.3-Codex-Spark

On February 12, 2026, OpenAI released GPT-5.3-Codex-Spark in research preview: a smaller version of Codex optimized for fast inference, served on the Cerebras Wafer-Scale Engine at over 1,000 tokens per second for near-instant feedback in live coding. Cerebras described it as the first release in its collaboration with OpenAI and stated that on agentic benchmarks such as SWE-Bench Pro and Terminal-Bench 2.0 it produces more capable responses than GPT-5.1-Codex-mini while completing tasks in a fraction of the time; Sachin Katti, OpenAI's head of industrial compute, described fast inference as "a new platform capability" (Source: cerebras.ai). The preview rolled out to ChatGPT Pro users in the Codex app, CLI, and VS Code extension, with API access limited to select design partners; TechCrunch reported OpenAI calling it the "first milestone" in its relationship with the chipmaker (Source: techcrunch.com). OpenAI later reported making Spark about 30% faster, serving at over 1,200 tokens per second (Source: community.openai.com).

GPT-5.3-Codex-Spark was also among the workloads OpenAI reported running on engineering samples of Jalapeño, the custom inference accelerator it unveiled with Broadcom on June 24, 2026 (Source: openai.com).

Safety and evaluations

GPT-5.3 Codex was the first model OpenAI treated as High capability in the Cybersecurity domain under its Preparedness Framework, a classification that activated associated safeguards. OpenAI described the step as precautionary, stating "we cannot rule out the possibility that it may be capable enough to reach the threshold" (Source: GPT-5.3-Codex System Card). The system card also designates the model High capability in Biology/Chemical (with the same safeguards as other GPT-5 series models) and finds it does not reach the High threshold for AI Self-Improvement (Source: GPT-5.3-Codex System Card).

For the Cybersecurity designation, OpenAI deployed what it called its most comprehensive cybersecurity safety stack to date: safety training, automated monitoring, trusted access for advanced capabilities, and enforcement pipelines including threat intelligence. Requests that its systems detect as having elevated cyber risk may be automatically routed from GPT-5.3-Codex to GPT-5.2, with security researchers able to apply for full access through Trusted Access for Cyber, a pilot program for cyber-defense research launched alongside the model. OpenAI paired the release with ecosystem measures: expanding the private beta of Aardvark, its security research agent; partnering with open-source maintainers on free codebase scanning for widely used projects such as Next.js, where a researcher using Codex found vulnerabilities (CVE-2025-59471 and CVE-2025-59472) disclosed the prior week; and committing $10 million in API credits for cyber defense, building on its $1 million Cybersecurity Grant Program launched in 2023 (Source: openai.com).

The system card reported the following safety evaluation scores (Source: GPT-5.3-Codex System Card):

The destructive-action-avoidance score of 0.88 compares with 0.66 for the original GPT-5-Codex, which OpenAI describes as a 33% relative improvement; the card identifies destructive actions such as rm -rf, git reset --hard, and push --force as a safety challenge specific to coding agents. On the conversational regressions (hate, extremism), OpenAI notes that conversational evaluations are not representative of real-world risk for a coding-focused model (Source: GPT-5.3-Codex System Card).

By default the model runs in isolated sandbox environments — cloud containers, or Seatbelt/seccomp/landlock policies locally — with network access disabled and file edits restricted to the workspace; users can expand network access to specific domains with explicit consent (Source: GPT-5.3-Codex System Card).

Reception

Released the same day as Claude Opus 4.6, the model's arrival was described as marking the "post-benchmark era" (Source: interconnects.ai). In comparisons with Claude Opus 4.6, GPT-5.3 Codex was described as holding a slight edge in pure coding — "a bit better at finding bugs and fixing things in codebases" — though the advantage was characterized as minor and most apparent in complex situations. It was reported to require more "babysitting" than Claude for mundane tasks such as "clean up this branch and push the PR," while being described as "much more Claude-like" than previous Codex versions, with faster feedback and broader task capability (Source: interconnects.ai).

Matt Shumer wrote that GPT-5.3 Codex was the model that made him feel "no longer needed for the actual technical work of my job," describing it as having something that "felt, for the first time, like judgment. Like taste" (Source: shumer.dev).

Analyst CharlesD, estimating METR time horizons for the two same-day releases before METR published its measurement, cautioned that model providers "have somewhat of a reputation for picking and publishing the benchmark evaluations that their models are going to do best on," and argued that both OpenAI's and Anthropic's published slates for the February 5 releases might therefore overstate the models' fitted capability (Source: lesswrong.com).

The model is discussed in The Shape of the Thing and in assessments of AI progress forecasts (Source: blog.aifutures.org).

Related models

GPT-5.3 Codex succeeded GPT-5.2 Codex and was directly compared with Claude Opus 4.6, released the same day (Source: interconnects.ai). GPT-5.4 (March 2026) folded its coding strengths into a general-purpose model (Source: openai.com), and the GPT-5.x coding line continued with the June 2026 GPT-5.6 family, whose Sol tier OpenAI positioned for the extended coding and security tasks GPT-5.3 Codex had targeted. Its GPT-5.3-Codex-Spark variant is a smaller, fast-inference model served on Cerebras hardware (Source: cerebras.ai).

Relationships

• developer: OpenAI
• instance-of: Frontier models
• depends-on: OpenAI Preparedness Framework — the High Cybersecurity designation activates safeguards specified there
• related: Recursive self-improvement — OpenAI frames the model's role in its own development as a milestone of this kind
• related: Claude Opus 4.6 — same-day release and direct comparison
• related: GPT-5.4 Thinking — first general-purpose model to extend the High-Cybersecurity mitigation approach introduced here
• related: GPT-5.6 — successor coding-capable line
• related: AI cybersecurity — first OpenAI model treated as High capability in the Cybersecurity domain
• related: GPT-5 family — covers the predecessor GPT-5 → GPT-5.2 span, including GPT-5.2-Codex