Testing 6 fast tiers: only Qwen earns its price markup
The author calculates a speed tax as price ratio divided by p50 speed ratio, finding GPT-6 Astra Fast at 1.21, Sol Fast at 1.27, GLM-Prime slower than its standard tier, and Astra Ultrafast at 6x price with no public samples.
The AI Speed Tax: 5 out of 6 Fast Tiers Are Overpriced The formula for value is simple: Tax = Price Ratio / p50 Speed Ratio. A Tax <= 1.0 means you actually get what you paid for. The clocked numbers: - Qwen 3.8 Max Prime: 2x price / 2.03x speed -> 0.99 Tax (n=753, thin data) - Opus 5.5 Fast: 2x price / 1.80x speed -> 1.11 Tax (n=261, thin data) - GPT-6 Astra Fast: 2x price / 1.65x speed -> 1.21 Tax (n=51, dropped from 1.05x speed same day) - GPT-6.1 Sol Fast: 2x price / 1.57x speed -> 1.27 Tax (n=28,480) - Mimo Ultraspeed: 10x price / 3.74x speed -> 2.67 Tax (n=798, thin data) - GLM-5.3-Prime: 2.35x price / 0.79x speed -> 2.97 Tax Only Qwen earns its price markup. Opus 5.5 Fast is the closest contender behind it. The biggest red flags: 1. GLM-Prime ran SLOWER than its own standard tier across all 3 captured sessions, despite pitching "1.5-2x output throughput" on its landing page. 2. Astra Ultrafast demands a 6x price multiplier with zero public samples and zero clockings. Context note: p50 represents real OpenRouter traffic, not a controlled synthetic lab run. The Rule: Never buy a fast tier blind. Always check the p50 speed and sample size on that specific host first. Identical weights at the same baseline price can deliver a 3x speed difference depending on who hosts them.
This post quotes another post: x.com/victortaelin/status/2106359338407215322