GPT-5.4 arrives with a million-token memory

On 5 March, OpenAI released GPT-5.4, and the detail that mattered most was not on any benchmark leaderboard. The new model carries a context window of 1.05 million tokens, roughly twice what the previous generation offered and large enough to hold an entire multi-year legal case, a long financial filing or a whole codebase in a single pass. It ships in three variants with different latency, cost and reasoning profiles, a structural choice that says as much about the company as the model itself.

The context window is the quiet revolution of this generation. A 1-million-token window changes what a model can be asked to do: analyse a year of hospital records in one go, review a contract repository against a new clause, or take an agentic task that needs to remember hundreds of intermediate steps. The previous pattern, in which long documents were split into chunks and summarized, lost accuracy exactly where accuracy matters most, at the boundaries between chunks. A model that can hold the whole thing avoids that loss, and GPT-5.4 is the first OpenAI model built around the assumption that customers will want to do this routinely.

The three-variant strategy is the part worth watching for what it reveals about pricing. Frontier labs spent 2025 competing on single flagship models with a take-it-or-leave-it price. GPT-5.4 splits the difference: a fast variant for high-volume, low-latency work, a balanced option, and a Pro tier for the hardest reasoning tasks. That is a deliberate move toward matching a model to a workload, the same logic that drove cloud providers to offer compute in dozens of sizes. It also points at a crowded market. By March, the reasoning work that GPT-5.4 Pro targets has direct competitors from Google, Anthropic and xAI, all chasing the same enterprise contracts.

OpenAI’s financial position adds context. The company passed $25 billion in annualized revenue around the same time, an extraordinary number for an organisation that was, in public memory, a research lab four years ago, and it has disclosed plans for an eventual IPO. The revenue figure matters because it changes what the model race is for. A lab with that much recurring revenue competes on deployment economics, on serving the customers it already has, rather than on winning benchmark debates, and the three-variant design of GPT-5.4 is precisely a deployment-economics move.

The technical details are consistent with the commercial ones. The 1.05 million-token window, in particular, requires inference systems that most labs cannot run cheaply; long-context processing is expensive in memory and compute, which is why the model ships in tiers and why the company has invested so heavily in serving infrastructure. The gap between a model that can hold a million tokens and the pricing that makes it usable is where the real competition now sits, and it favours the labs with the largest fleets and the lowest unit costs.

None of this guarantees the model will be remembered as a breakthrough on capability. GPT-5.4 is, by the benchmarks published so far, roughly in line with the best of its generation, and the improvements over its predecessor are real but incremental. What it represents instead is a change in the product: the frontier has moved from how smart a model is to how much of your problem it can hold, and how cheaply it can hold it. That shift is what enterprises will feel, and it is the reason the release, for all its modest numbers, deserves attention.