GPT-5.3-Codex: Coding Agents, Speed and Computer Use at Launch

In this February 2026 video, Matthew Berman discusses OpenAI’s GPT-5.3-Codex release alongside Anthropic’s Claude Opus 4.6, announced the same day. He focuses on long-running coding agents, responsiveness and examples from OpenAI’s release materials. The comparisons reflect announcements and benchmarks available at the time.

OpenAI reports that GPT-5.3-Codex runs 25% faster for Codex users. Berman welcomes the change because many users had found coding agents slow. He also mentions Peter Steinberger’s preference for Codex over Claude on his own codebase, while noting that he does not necessarily agree; that is an individual assessment, not a controlled comparison.

OpenAI says users can steer the agent during a task without losing context. It also says early versions helped the team debug training, manage deployment and diagnose evaluations. Berman calls this nearly autonomous self-improvement, but the description is assistance to a human-led development process, not evidence that the model independently designed or trained its successor.

Berman points to an OpenAI SWE-Bench Pro chart showing roughly 43,000 output tokens for GPT-5.3-Codex versus 91,000 for GPT-5.2-Codex at the displayed operating points. He interprets the reduction as the cause of the speedup. OpenAI’s release separately attributes the reported 25% faster Codex experience to infrastructure and inference-stack improvements; the token-count chart alone does not establish the cause of that figure.

OpenAI also reports improved Terminal-Bench results and presents two game examples: a racing game with different racers and a diving game with objectives, predators and an oxygen limit. Berman says the runs involved brief human prompts such as fixing bugs or improving the game over millions of tokens. These are selected demonstrations, not proof that arbitrary games can be built without supervision.

OpenAI says the model handles underspecified prompts by making more sensible defaults. Berman thinks that could help less experienced users who cannot specify every design detail. In the sponsored segment he describes Greptile’s code-review product; automated review and confidence scores may surface issues, but do not replace appropriate testing or human review.

A landing-page example compares GPT-5.2-Codex and GPT-5.3-Codex using the same Quiet KPI prompt. Berman prefers the newer page’s visual design, month-over-month figures and clearer pricing toggle with annual savings. The screenshots show one curated example, so his aesthetic preference should not be generalized into a comprehensive design benchmark.

Berman compares the push into professional work with Anthropic’s Claude Cowork, citing examples such as documents, spreadsheets and presentations. OpenAI’s examples show outputs for financial, retail and fashion tasks. These examples illustrate the intended scope of the product, not independent validation that their contents are correct.

For computer use, OpenAI reports 64.7% on OSWorld-Verified for GPT-5.3-Codex, compared with 38.2% for GPT-5.2-Codex and 37.9% for GPT-5.2 in its displayed results. OSWorld evaluates tasks in a visual desktop environment. Berman criticizes OpenAI’s chart for omitting competitors; the figures do not establish a ranking against them or guaranteed reliability on every desktop task.

Berman says the model was available in the downloadable Codex app and invites viewers to compare it with Opus 4.6. Availability and model rankings change over time; the video is a release-day perspective, not a current purchase or deployment recommendation.