Model Maximalism
Model maximalism is OpenAI’s internal product philosophy: do not build scaffolding around current model limitations; instead, build at the edge of model capability and trust that models will improve fast enough to make those limitations irrelevant.
The core argument
“Our general mindset is in two months, there’s going to be a better model and it’s going to blow away whatever the current set of limitations are.” — Kevin Weil on OpenAI and the Future of AI Products
AI models are improving at approximately 10x per year. A product that barely works today — because the model gets something right 70% of the time — will work well in three months when the model’s reliability on that task is 95%.
The implication: engineering time spent compensating for model limitations is largely wasted. The limitation will be gone before the scaffolding has earned back its cost. Build the right product; let the model catch up.
What it means in practice
What model maximalism recommends:
- If your product is right on the edge of what the model can do, keep going. You are doing something right.
- Invest in product design, evals, and user experience — these compound and remain valuable as models improve.
- Invest in fine-tuning models for your specific use case — this makes your model better, not less bad.
What model maximalism discourages:
- Building extensive guardrails or fallback logic for model errors that will be fixed in the next model release.
- Waiting for models to be “good enough” before building — the edge moves fast.
- Designing products around current limitations as if they are permanent.
The exception: safety-critical failures where errors are catastrophic warrant dedicated scaffolding regardless of future model improvement.
The “worst model you’ll ever use” corollary
“The AI models that you’re using today is the worst AI model you will ever use for the rest of your life.”
This framing is the empirical basis for model maximalism. The rate of improvement (approximately 10x/year in capability-per-cost) is steeper than Moore’s Law. Two orders of magnitude in API cost reduction in two years, with simultaneous improvements in capability, speed, and hallucination rate.
Applied to product strategy: evaluate products not on how they work today but on where the models will be in 6–12 months. Build for where the puck is going.
Tension with caution
Model maximalism is in deliberate tension with the instinct to compensate for model failures. Some counterpoints:
- Users experience the product today, not the better model in three months. A 70%-reliable product today damages trust even if it becomes 95%-reliable later.
- Not all limitations improve at the same rate — social/emotional intelligence, aesthetic judgment, and creative reasoning improve more slowly than factual retrieval and coding.
- Products shipped before models are ready can establish negative user perceptions that persist after the model improves.
The resolution Kevin implies: model maximalism applies where failures are recoverable and the use case is roughly in the model’s current capability zone. It does not apply where failures are catastrophic or where the use case is fundamentally outside current model capability.
Related thinking
- Scaling Laws — the empirical basis for projecting continued improvement.
- Jagged Intelligence — models have uneven capability profiles; model maximalism applies differently across the jagged frontier.
- Iterative deployment — the companion product philosophy: ship early, learn in public, co-evolve with users.
- Vibe Coding — an application of model maximalism: build by trusting the model and guiding rather than micromanaging.