Open-Source Is Eating Frontier AI's Lunch, and the Price Chart Proves It
textak places the probability that open-source matches closed frontier performance at 75%, up from 72%. Today's OpenRouter data — Chinese providers capturing 45% of API traffic, up from 2% a year ago, with DeepSeek-V4-Pro-Max hitting 80.6% on SWE-bench while costing 28.7x less than Claude Opus — is the strongest directional evidence we've seen that the gap is closing in the dimension that actually matters to enterprise buyers: price-adjusted performance. We're holding at 75% rather than moving higher, and the reason is worth explaining.
Let's be precise about what 75% means and what it doesn't. The forecast target is 'open-source model matches closed frontier performance' — and our editorial standards require us to acknowledge that 'matches' is doing significant definitional work here. We operationalize this as: an openly-released, reproducibly deployable model achieves parity on at least two independent coding or reasoning benchmarks with the leading closed frontier model in the same evaluation window, as confirmed by independent evaluators rather than lab-released numbers. That threshold hasn't been crossed yet on the hardest benchmarks. The 55-point gap on Humanity's Last Exam — frontier models at 35%, human domain experts at 90% — is a genuine reminder that the most demanding academic reasoning tasks still differentiate open from closed systems significantly.
That said, today's evidence is genuinely strong directional signal, and we want to be honest about why. DeepSeek-V4's SWE-bench numbers and OpenRouter routing behavior are not direct evidence of parity — OpenRouter skews toward developers and researchers, not Fortune 500 procurement teams, and benchmark scores are proxies, not product performance. But the pricing dynamic is structural, not episodic. When the performance gap on developer-relevant benchmarks is small enough that 28x cost compression makes the choice obvious for a broad class of workloads, the market has effectively reached functional parity for those workloads — even if absolute benchmark supremacy hasn't transferred.
The part of our thesis that keeps us honest is the counterargument from unreleased capabilities. Anthropic's internal development pipeline and OpenAI's post-training techniques represent genuine information asymmetry that publicly available benchmarks can't capture. If a closed lab releases a model with step-change improvement in the next 90 days — think the kind of qualitative leap that moved GPT-4 to o1 — the gap could reopen faster than open-source can close it. We weight this risk at roughly 20% of the remaining probability mass: plausible, not dominant.
What would move us above 80%? A peer-reviewed or independent-evaluator confirmation that an openly licensed model matches or exceeds GPT-5.6 Sol on Terminal-Bench or an equivalent agentic task suite — not just pricing advantage on SWE-bench. What would drop us below 65%? A closed-lab release in Q3 2026 that reestablishes a 15+ point performance lead on multiple independent benchmarks simultaneously, with no open-source response within 60 days. We're watching the July-August release cycle closely.