Research on AI API gateways — how they actually charge, what the models really cost per million tokens, and which levers move an LLM bill. Every figure we publish is dated and sourced.
News · 10 Aug 2026Every major gateway now passes tokens through at 0% markup — the money moved to platform fees. At 50M requests/month the spread runs from free (Cloudflare) to $10,300 (Kong), all on identical token cost. Plus Envoy AI Gateway’s v1.0 GA, and the benchmark everyone quoted that turns out to be a mock.
GuideA gateway with “0% markup” can still take 16% of a small top-up. The biggest savings in 2026 are not the gateway at all — they are 90% off cached input and 50% off batch, published by the model vendors themselves.
PricingEvery price, every multiplier, every trap — including the one Anthropic states but nobody budgets for: newer models produce ~30% more tokens for the same text, and up to 47% on technical English.
ComparisonTwo of these models cost the same per token. One costs nearly twice as much to run. Measured on the identical benchmark suite: Opus 4.8 $4,011, GPT-5.5 $2,159, Gemini 3.1 Pro $815.
Analysis$13 per developer per day, 3,500× the tokens of a chat call, and one documented $12 API call where 21 million cache-read tokens were billed for 4,000 tokens of real input.
GuideOne team went from a 7% cache hit rate to 84% and cut LLM cost 59%. The change that did most of it? Moving a timestamp out of the system prompt. Every vendor’s rules, and why caches silently miss.
ComparisonThey are not rivals — one sells tokens, the other is software you run. OpenRouter publishes its own break-even: self-hosting wins above $3,600/month of spend. Add an engineer and it moves to $36,400.
ComparisonPeople leave over the 2026 403 wave, not the fee. Verified pricing for 14 alternatives — 5% on tokens, 5.5% on top-ups, $49 per 100k logs, or genuinely zero markup.
Copyright © 2026 Best AI Gateways
Independent ranking. We may be rewarded for recommending the service we rate best and sending users to it — that reward pays for the research behind this comparison and never buys a ranking position, at no extra cost to you. Information is provided “as is” without warranty. See our Disclaimer & Terms.