GPT-5.4 Accidentally Leaked in OpenAI Codex Repository: What the Code Reveals

A Model Name That Was Not Supposed to Be Public

On February 28, 2026, a pull request appeared in OpenAI's public Codex GitHub repository containing something developers were not meant to see: a version cutoff reading "GPT-5.4 or newer" for an image handling feature. The reference was quickly scrubbed via force pushes and edits, but not before X user @scaling01 flagged it publicly. Then on March 2, a second Codex pull request surfaced with a direct function call using gpt-5.4 as the model argument, plus a slash command explicitly described as "toggle Fast mode for GPT-5.4." An OpenAI employee named Tibo compounded the situation by accidentally posting a screenshot on X that included GPT-5.4 in the model picker before deleting it.

These are not speculative rumors. They are code artifacts from OpenAI's own repositories, providing concrete evidence that GPT-5.4 exists internally and is actively being developed for production deployment.

What the Leaked Code Reveals

Two-Million-Token Context Window

The most significant technical detail extracted from the leaked code is a context window of 2 million tokens. For perspective, GPT-5.3 Codex operates with a 256,000-token context window. An eight-fold expansion to 2 million tokens would place GPT-5.4 in territory currently occupied only by Google's Gemini 3.1 Pro, which offers a 1-million-token window. If the 2-million-token figure holds, GPT-5.4 would set a new record among commercially available frontier models.

A context window of this size has practical implications for enterprise users. Legal firms could process entire case files in a single conversation. Software development teams could load complete codebases for analysis. Research teams could feed entire paper corpora for synthesis. The shift from hundreds of thousands to millions of tokens is not incremental—it fundamentally changes what tasks are feasible in a single model interaction.

Full-Resolution Image Handling

The leaked pull request adds a feature flag that preserves original PNG, JPEG, and WebP image bytes, passing them directly to the Responses API without compression. Current GPT models compress uploaded images before processing, which can degrade analysis quality for tasks requiring pixel-level precision—medical imaging, satellite imagery analysis, document OCR, and design review among them.

The code sets the minimum model version to (5, 4), meaning this feature explicitly requires GPT-5.4 or newer. This is not a general capability being backported; it is a GPT-5.4-specific enhancement.

Fast Mode Toggle

The slash command for "toggle Fast mode for GPT-5.4" suggests OpenAI is building a speed-optimized inference path, likely similar to the existing approach where models offer different speed-quality tradeoffs. This aligns with OpenAI's recent pattern of releasing both standard and optimized variants of each model generation.

Timeline and Release Probability

OpenAI has maintained an accelerating release cadence within the GPT-5 series:

The interval between releases has compressed from three months to roughly one month. If this pace continues, a March or early April release for GPT-5.4 is plausible.

Prediction markets on Manifold currently give GPT-5.4 a 55% probability of shipping before April 2026 and 74% before June. However, shipping both GPT-5.3 and GPT-5.4 within the same month would require unprecedented acceleration even by OpenAI's recent standards.

OpenAI has made no official announcement regarding GPT-5.4. The company has not commented on the leaked code references.

Competitive Context

The timing of these leaks is notable. OpenAI is navigating a period of significant public relations pressure following its Pentagon classified network deal, which has triggered the "Cancel ChatGPT" movement and driven measurable user migration to Anthropic's Claude. A major model release could serve as both a technical milestone and a strategic counter-narrative.

Meanwhile, the competitive landscape has intensified. Google's Gemini 3.1 Pro already offers a 1-million-token context window with strong multimodal performance. Anthropic's Claude Opus 4.6 has been drawing praise for coding and reasoning capabilities. A 2-million-token GPT-5.4 with enhanced vision capabilities would represent OpenAI's bid to reclaim technical leadership on multiple fronts simultaneously.

What This Means for Developers

For developers building on the OpenAI API, the leaked features suggest several preparation steps. Applications that currently chunk documents to fit within context limits may need architectural revisions to take advantage of million-scale windows. Image processing pipelines that pre-compress inputs before API submission may benefit from passing raw bytes directly. Teams evaluating multi-model strategies should factor in the possibility that GPT-5.4 could narrow or close capability gaps that currently favor competitors in specific domains.

The Fast mode toggle also suggests that OpenAI is continuing to invest in inference optimization, which could translate to lower latency and potentially lower costs for real-time applications.

Conclusion

The GPT-5.4 leak is not a roadmap slide or a conference teaser. It is production code referencing a specific model version with specific capabilities—a 2-million-token context window, full-resolution image handling, and a fast inference mode. Whether the model ships in March, April, or later, the code confirms that OpenAI is building the next iteration of its flagship model with substantial capability expansions. Developers and enterprise users should monitor OpenAI's official channels for an announcement that, based on the evidence, appears to be a matter of weeks rather than months away.