OpenAI's next major language model has cleared its first major milestone. The model, internally codenamed "Spud", completed pretraining on March 25 — and Sam Altman reportedly told staff that the pace of progress is "moving faster than many of us expected."
The Milestone
Pretraining is the foundational phase of building a large language model: the computationally expensive process of exposing the model to vast amounts of text data and learning to predict and generate language. It is necessary but not sufficient — a pretrained model is not yet safe, aligned, or ready for deployment.
What comes next is a lengthy post-training process: fine-tuning on curated data, reinforcement learning from human feedback, safety evaluations, and adversarial red-teaming. That pipeline typically takes several months for a frontier-scale model. Spud completing pretraining now does not mean it is close to release — but it does confirm the project is progressing.
Altman's comment to staff about pace is notable. OpenAI, like all frontier labs, operates under intense competitive pressure. Internal expectations are calibrated to a fast-moving industry. When the CEO says things are moving faster than expected even internally, it suggests the model is performing ahead of its training projections — a meaningful signal about capability level.
The Race Context
Spud arrives at a particularly charged moment in the frontier model race. Anthropic's next major model — internally called Mythos, with a smaller version known as Capybara — was accidentally exposed in a security lapse earlier this week. Anthropic describes Mythos as a "step change" in capability, the kind of language labs typically reserve for genuinely discontinuous improvements.
If both assessments are accurate — Spud ahead of internal projections, Mythos a step change — the industry is approaching a moment where two frontier labs simultaneously release models meaningfully more capable than anything currently available. The last time that dynamic played out, in late 2024, it produced a significant shift in what AI could demonstrably do.
GPT-5 launched earlier this year to strong benchmark performance and rapid adoption across enterprise use cases. Spud is positioned as the next tier above GPT-5, suggesting OpenAI is targeting another substantial capability jump rather than an incremental improvement.
What Post-Training Involves
The gap between pretraining completion and public release is not downtime — it is where a model becomes usable. Post-training alignment work teaches the model to follow instructions, refuse harmful requests, and behave consistently across a wide range of prompts. Safety evaluations probe for dangerous capabilities, misuse potential, and unexpected behaviours. Red-teaming involves adversarial testing by teams specifically tasked with finding failure modes.
For a frontier model with Spud's projected capability level, each of these stages is more complex than for smaller models. More capable models have larger surfaces for potential misuse and subtler failure modes that can escape automated testing. Anthropic, Google DeepMind, and other labs have all described increasing the depth of their safety evaluation processes as model capability grows.
The Acceleration Signal
The broader signal from Spud's pretraining completion — combined with Mythos, combined with the pace comments — is that the capability frontier is moving faster than the public conversation acknowledges.
Benchmark results lag reality by months. By the time a model is publicly released, evaluated on standard tests, and written up in the AI press, the lab has already moved on. Spud completing pretraining today means OpenAI's internal research teams are already thinking about what comes after Spud.
When Altman says things are moving faster than expected, he is speaking from inside that curve.
