Open vs. Closed Source AI: The Monetizable Spread Argument

Author introduces 'monetizable spread' to reframe closed-source AI valuations

The author argues the raw capability spread (benchmark delta between best closed and best open model) is the wrong metric for investors. What matters is the 'monetizable spread' — the subset of that delta customers will actually pay a premium for. This is declining faster than the headline capability spread, a divergence the author contends markets have not priced.

MMLU benchmark gap closed from ~17.5 pp to near zero in roughly 2 years

In late 2023, the best closed model scored ~88% on MMLU while the best open model scored ~70.5%. By early 2026, that gap is effectively zero on knowledge benchmarks and single digits on most reasoning tasks.

Epoch AI: open-weight models trail frontier by ~3 months, down from ~1 year in late 2024

The time lag between frontier closed models and the best open-weight equivalents compressed dramatically in roughly 18 months, implying continued rapid erosion of any capability-based moat for closed-source labs.

DeepSeek V3 trained at 2.6M GPU hours vs. Llama 3 405B's 30.8M — ~12x more efficient

DeepSeek's R1 reasoning model, built on V3, matched OpenAI's o1 at roughly 3% of the cost — illustrating how open-source teams are achieving capability parity through efficiency breakthroughs rather than proportional compute spend.

The 'good enough' line — where open models become interchangeable with closed ones — moves up each quarter

Each time the 'good enough' threshold rises to cover another task tier, a slice of paying customers loses economic justification for the closed-model premium. The monetizable spread is this gap between the threshold and the top of the task stack, multiplied by revenue density at each layer.

Anthropic's Economic Index: routine coding and math tasks = 36% of API usage

Menlo Ventures enterprise survey similarly found code completion and productivity tools — not agentic systems — drove category revenue, supporting the thesis that most current AI revenue sits in task tiers already reachable by open models.

Frontier closed models still lead on complex agentic coding and long-horizon reliability tasks

The author is not arguing open models have won overall. The capability spread at the top of the stack — complex multi-step tool chaining, long-horizon workflows — remains real. The argument is that the list of tasks where closed models meaningfully lead is shrinking each quarter.

Framework rests on unverified assumptions about revenue density by task complexity

No published dataset cleanly maps AI revenue to task difficulty. The revenue density claims at each tier are the author's assertion, not measurement. The author explicitly flags this caveat, noting proxy data is 'strongly directional' but not conclusive.