An 11x price spread for the same open-weight model on OpenRouter

Brett Harrison, founder and CEO of the derivatives exchange group Architect, published a snapshot of OpenRouter's open-weight inference market showing an 11x gap between the cheapest and most expensive host of a single model — DeepSeek V4 Flash. His framing question is whether inference costs can be commoditized at all when identical weights sell for wildly different prices at wildly different speeds.

Open Weight Inference Metrics, source data from OpenRouter. Credit: Brett Harrison / Architect.

#The numbers

• Baidu served DeepSeek V4 Flash at $0.049 per million input tokens at 124 tokens per second — the cheapest and nearly the fastest endpoint on the board. 26 of the 30 other providers were both more expensive and slower. As Harrison puts it, "it's not a tradeoff of speed vs cost."
• Across the 117 models on OpenRouter with three or more providers, the median price spread between cheapest and priciest host is 2.2x; the most extreme is 14.4x, for DeepSeek v3.2.
• 37% of endpoints are strictly dominated on cost, speed and uptime — beaten on all three by some other host of the same model. Considering only cost and speed, that rises to 52%.

The right-hand panel of his chart ranks the spreads: DeepSeek V3.2 at 14.4x across 14 hosts, Gemma 4 31B at 12.4x across 18, DeepSeek V4 Flash 0731 at 11.0x across 30, GLM 5.2 at 6.9x across 34, down to Hy3 at 1.6x across 7.

#Checking it against OpenRouter today

We pulled OpenRouter's public endpoints API to see whether the dispersion holds. It does, though the individual rankings churn:

On DeepSeek V4 Flash 0731 the cheapest input price is now $0.05/M at OpenInference and the most expensive $0.44/M at Cloudflare, Phala and AtlasCloud — an 8.8x band. Baidu no longer sits at the bottom: it now lists at $0.14/M, in the crowded middle. Quantization varies across that band too — fp4, fp8 and bf16 endpoints are priced side by side — which is one reason the cheapest number is not automatically the right buy.

#Why it matters

The standard story about open weights is that they commoditize inference: anyone can serve the model, so price converges to the cost of GPU time. Harrison's data says the convergence hasn't happened. A router that picks purely on model name, or a default endpoint left untouched, can plausibly cost an order of magnitude more than the frontier of the same market — for bytes that are, nominally, identical.

The dominated-endpoint figure is the sharper one. If more than a third of endpoints are worse on every axis that matters, the market is not clearing on quality; it is clearing on inattention, defaults and routing policy. That is also why the integrity of router metrics matters so much — providers have incentives to game the scores that determine who gets traffic, and OpenRouter's own numbers are the thing everyone is optimizing against now that Stripe has agreed to buy it.

One disclosure worth keeping in view: Architect, Harrison's company, is building derivatives markets for "AI commodities." A world where inference is a fungible, priced commodity with a visible spread is the world his exchange is designed for. The underlying OpenRouter figures are checkable — we checked several of them — but the framing is not disinterested.