## The Seven Force Pushes On February 27th, an OpenAI engineer submitted PR #13050 to the public Codex GitHub repository. Buried in the diff was a version check: minimum model `(5, 4)`. Within five hours, the PR was scrubbed via seven successive force pushes — the git equivalent of frantically shredding documents while the office watches. Three days later, PR #13212 dropped a `/fast` slash command with the description "toggle Fast mode for GPT-5.4." Gone within three hours. Then an employee named Tibo posted a screenshot of the Codex model selector dropdown. GPT-5.4 sat right there, selectable, between 5.3-Codex and the model picker's edge. That screenshot lasted minutes. Three leaks in five days, from the same public repository, at a company that just closed a $110 billion funding round. Either OpenAI has a serious operational security problem, or someone wants us talking about GPT-5.4. ## What the Code Actually Says Forget the headlines. The tech press ran with "2 million token context window" and "stateful AI" within hours of the leak. Neither claim has a single line of code evidence. Here's what the PRs actually contained: **Full-resolution vision.** PR #13050 added a `detail: "original"` flag for image passthrough — PNG, JPEG, and WebP files sent to the API without compression. Current models downscale everything. This change means pixel-level accuracy for technical schematics, UI mockups, and medical imaging. Not glamorous. Extremely useful. **Priority inference tier.** PR #13212 introduced `service_tier=priority` as an API parameter. A fast lane for paying customers who want lower latency. This is a pricing and infrastructure play, not a capability breakthrough. That's it. Two features. One makes vision actually work for professional use cases. The other lets OpenAI charge more for speed. The 2M context window? Speculation extrapolated from Gemini's 1M lead. Persistent stateful memory? A logical wish list item that appeared in zero code. The gap between what was leaked and what was reported tells you everything about how AI coverage works in 2026. ## The Cadence Tells the Real Story Zoom out from the leak and look at OpenAI's release timeline: - **GPT-5** — August 2025 - **GPT-5.1** — October 2025 (reasoning modes) - **GPT-5.2** — December 2025 (knowledge work + Pro tier) - **GPT-5.2-Codex** — January 2026 (coding specialist) - **GPT-5.3-Codex** — February 5, 2026 (agentic coding, 25% faster) - **GPT-5.4** — leaked March 2026 Monthly point releases since October. Each one more specialized, less general. OpenAI isn't building toward one god model. They're building a portfolio — the same strategic shift Microsoft made in the 2010s when they stopped trying to make Windows do everything and started shipping Azure services individually. GPT-5.3 general-purpose hasn't even shipped. Only the Codex variant exists. Which means either OpenAI skipped 5.3 general entirely (unlikely given their naming conventions), or they're decoupling the Codex and general tracks into separate release cadences. Prediction markets currently give GPT-5.4 a 55% chance of shipping before April and 74% before June. Given the monthly cadence, March or April feels right. ## Where This Leaves the Race The current competitive landscape, as of this week: **Coding:** Claude Opus 4.6 leads with 80.9% on SWE-bench — the first model above 80%. GPT-5.2 sits at 55.6%. That's not a gap, it's a canyon. **Reasoning:** GPT-5.2 dominates ARC-AGI-2 at 54.2% (Pro mode) vs Claude's 37.6% and Gemini's 45.1%. **Agentic tasks:** Claude edges ahead when models use tools. Better integration, more reliable execution chains. **Context:** Gemini's 1M window is 5x larger than everyone else. If GPT-5.4's rumored 2M is real, that lead evaporates overnight. **Price:** Gemini 3.1 Pro costs 7x less than Claude Opus 4.6 per request and leads on most benchmarks. The value play is increasingly hard to ignore. Notice what's happening: no single model wins across all categories. The AI industry has entered its specialization era. The question isn't "which model is best" anymore — it's "best at what, for whom, at what price." ## The Stateful Bet The most interesting claim about GPT-5.4 — persistent memory across sessions — has no code evidence but deserves attention anyway. Not because it's confirmed, but because it's the one capability that would actually change the competitive dynamic. Current AI interactions are goldfish conversations. Every session starts from zero. You re-explain your codebase, your preferences, your project context. The model that solves this — that genuinely remembers your work across days and weeks — creates a switching cost that benchmarks can't capture. Claude Code already does something like this with project memory files. Cursor and Windsurf maintain context across sessions through workspace indexing. But these are application-layer solutions. A model-native persistent state would be fundamentally different — and significantly harder to replicate. If OpenAI ships this with GPT-5.4, the conversation shifts from "which model scores higher" to "which model knows me better." That's a moat. ## The Strategic Read Three possibilities for why this leaked the way it did: **Negligence.** OpenAI's Codex team has their model references in a public repo and insufficient review processes for PRs that touch version strings. Boring but plausible — especially given the seven-force-push panic response. **Strategic leak.** March 2026 has been quiet. Claude Opus 4.6 dropped in February. Gemini 3.1 landed the same month. OpenAI needs to hold developer attention while GPT-5.4 bakes. A leak that generates a week of coverage costs nothing and delays nobody. **Accelerated timeline.** The Codex team is shipping so fast that internal model references are outrunning PR review cycles. The leak is a symptom of velocity, not strategy. Whatever the cause, the effect is the same: we're talking about a model that doesn't exist yet instead of evaluating the ones that do. OpenAI's most reliable product isn't AI — it's anticipation.