Hands-On With GPT-5.3-Codex: OpenAI’s First High-Risk Model—So Restricted It Isn’t Even Available via the API Yet

The release of OpenAI GPT-5.3-Codex early this morning can be read as a decisive counterpunch to the recent surge of local Agent tools—most notably a direct response to Anthropic.

Paired with OpenAI’s newly launched Codex desktop app, many of the capabilities showcased by popular tools such as Skill, Cowork, Claude Code, and Openclaw can now be achieved within a single environment—powered by the Codex interface and the GPT-5.3-Codex model itself.

Inside the Codex app, users can directly select GPT-5.3-Codex and fine-tune the depth of its reasoning. To evaluate its real-world performance, we assigned it tasks similar to those we previously tested with Cowork: handling local files, converting between formats, orchestrating multiple Skills, generating Word, PowerPoint, and Excel documents, downloading videos, and even developing applications.

The results were genuinely impressive. Compared with the more involved setup required to install Claude Code from scratch, Codex offers a far more accessible entry point—especially for newcomers. This reflects a broader trend in the ecosystem: local Agents began in stark command-line environments, but are steadily converging toward polished, user-friendly graphical interfaces.

Public sentiment around Codex has also shifted noticeably in recent days. Many developers report migrating from Claude Code to Codex, while independent developers—particularly in China—note that Codex Plus is readily usable and avoids the frequent account bans that have frustrated some Claude users.

OpenAI CEO Sam Altman has openly celebrated this momentum, announcing that Codex has surpassed one million active users. In OpenAI’s own model update blog, the praise is equally explicit: GPT-5.3-Codex is described as the company’s first model capable of “self-construction,” enabling rapid iteration and deployment of subsequent versions.

Seen alongside Anthropic’s achievement of building Cowork in just two weeks using Claude Code with 100% AI-generated code—and OpenAI’s earlier case study on building Android-based Sora in 28 days—the conclusion is hard to avoid: the Agent era has truly arrived.

Replacing ChatGPT and Claude Code With Codex

Like most local Agent tools, Codex begins with a workspace. Users can create multiple Projects, each linked to a specific folder, and then open conversational Threads within those projects. The workflow feels immediately familiar to anyone accustomed to ChatGPT.

Starting with a simple example, we created an empty download folder, launched a new Thread, selected GPT-5.3-Codex, and issued a prompt. When asked to download a video from X, Codex automatically identified the appropriate Skills and invoked yt-dlp to handle the task. Even for a video over four hours long, Codex continuously updated progress directly within the conversation.

Once the download completed, we asked Codex to extract a full transcript, produce a bilingual document, and then package the entire workflow into a reusable Skill. The same session could then be extended to clip interesting segments, convert them into GIFs, or transcode formats—all without leaving Codex.

For instance, we instructed Codex to extract a 5-to-25-second segment into a new video. Thanks to GPT-5.3-Codex’s fast token processing, the turnaround was quick, with total time largely dependent on local hardware encoding and decoding performance. We could just as easily request a GIF from the first five seconds, constrained under 10 MB, with adjustable frame rates and a fixed width of 640 pixels.

Moments later, the GIF was ready. At the extreme end, Codex can even convert an entire video into image frames—30 frames per second, one image per frame. Combined with GPT-5.3-Codex’s strong showing on Terminal-Bench-2, these local file operations make Codex a credible foundation for a wide range of productivity and efficiency tools.

By comparison, the newly released Claude Opus 4.6 scores 65.4% on Terminal-Bench 2.0, while GPT-5.3-Codex reaches 77.3%.

From File Management to Full Applications

In another test, we asked Codex to rename a batch of images based on their content, limiting filenames to 20 alphabetic characters with no symbols. After completing the renaming, Codex seamlessly stitched the images together—vertically or horizontally—by invoking the appropriate tools.

Like Claude’s ecosystem, Codex supports a rich Skills marketplace. Out of the box, it already includes integrations for PowerPoint, Excel, Word, Canvas, and Notion.

On the core programming front, GPT-5.3-Codex represents a clear step up from GPT-5.2. We tasked it with building a “Word of the Day” app from scratch. Unlike ChatGPT’s Canvas—which produces a web preview that cannot be easily exported—Codex built the project locally and then deployed it via Skills such as Vercel or Cloudflare.

Using the Extra High reasoning mode, GPT-5.3-Codex paused before each major step to confirm next actions. This behavior reflects Codex’s internal orchestration, where different Skills—including brainstorming modules—are dynamically invoked as part of a continuous dialogue.

Ultimately, Codex delivered the requested functionality and even proposed extensions to macOS, iOS, and Android versions. When pointed at an existing codebase, it can analyze the project, identify bugs, and implement fixes directly.

A Shifting Balance in the Coding Model Wars

For a long time, developers gravitated toward Anthropic’s Sonnet and Opus models paired with Claude Code. OpenAI was widely perceived as lagging in complex, long-form code reasoning. GPT-5.3-Codex is clearly intended to close—and perhaps end—that gap.

Across benchmarks and hands-on testing, GPT-5.3-Codex not only surpasses its own predecessors but also shows early signs of outperforming competing models. It now convincingly demonstrates the ability to write, test, and reason about code as an integrated process.

Game development featured prominently in OpenAI’s release blog, so we asked Codex to build a simple physics-based pinball game. While the interface fell short of our RPG-style expectations, the result was functional—and playable. Online, developers have already shared other mini-games built with GPT-5.3-Codex, including Mario-like coin-collection experiences.

No Absolute Winner—Only Faster Cycles

From Anthropic’s perspective, OpenAI’s latest moves may feel familiar territory. In recent years, OpenAI has often appeared to follow Claude’s lead: when Claude doubled down on code, OpenAI pivoted from Sora, news digests, and browser experiments back toward programming. When Anthropic released Cowork in early January, OpenAI followed with the Codex app in early February.

The timing underscores this rivalry. At 1:45 a.m., Anthropic announced Claude Opus 4.6 on X. Shortly afterward, OpenAI unveiled GPT-5.3-Codex. Both models aim to serve as stronger foundations for Agents. The narrative has shifted from “vibe coding” to a simpler truth: powerful Agents depend on exceptional coding ability.

While Opus 4.6 trails GPT-5.3-Codex on Terminal-Bench 2.0 and SWE-Bench, it compensates by expanding context length to an unprecedented one million tokens. And the performance gap across benchmarks remains relatively narrow.

Anthropic itself hints that Sonnet 5 has yet to arrive—and that is where it expects its true strength to emerge.

Early user tests reinforce this competitive tension. Some report that Opus 4.6 can refactor entire codebases in a single pass, modularizing what were once unmanageable “spaghetti” projects. Others compare Opus 4.6 and 4.5 by letting both models play the same management game; while 4.6 spends longer on early strategy, it ultimately dominates in outcomes.

One developer even built a Pokémon-style clone, calling it the coolest AI-assisted project they had ever made. According to the report, Opus 4.6 thought for 90 minutes, consumed 110,000 tokens, and required just three iterations.

Anthropic’s own demos highlight similar strengths: Opus 4.6 reportedly closed 13 issues autonomously in a single day and correctly assigned another 12 to human teammates. Like Kimi K2.5’s agent swarm, Opus 4.6 can coordinate large codebases, supporting multi-agent teams where AI handles coding, review, and testing in parallel—boosting speed by as much as 2.5×.

Stronger Models, Stronger Users

The broader picture feels cyclical: Gemini had its moment, then Claude, and now OpenAI—or perhaps even Grok—may be next. Yet with each cycle, users benefit from tangible gains in capability.

OpenAI has not yet opened the GPT-5.3-Codex API, citing the model’s power and associated risks. The company is still evaluating how to enable access safely. Meanwhile, Claude Opus 4.6 is already available via Claude’s chat app, Claude Code, and API.

As two of this year’s most significant model releases, both GPT-5.3-Codex and Claude Opus 4.6 are well worth exploring. Looking ahead, enhancing Agent-centric workflows—so that Agents can truly work on our behalf—will remain a central focus of large-model development.