Enterprise AI Inference Costs Hit New YTD Low, Driven by Price War and Chinese Open-Source Models
Miles Bennett
Enterprise AI inference prices fell to roughly $1.17 per million tokens in early August — a 2026 low — as a global price war and the spread of cheap Chinese open-source models squeeze service providers from both sides.
How far have inference prices fallen?
Jefferies, citing US research firm Silicon Data, reports that average inference prices dropped to $1.16–$1.18 per million tokens between August 6 and 8 — the lowest this year.
That marks a roughly 43% decline from the $2.04 recorded on May 31, and a further step down from $1.45 in late July.
In plain terms = the same AI query now costs an enterprise almost half what it did three months ago.
What is driving this round of cuts?
Force one: a global AI price war. OpenAI last month slashed prices on its latest GPT-5.6 series by up to 80%; Anthropic's Claude Opus 5 matches the performance of its flagship Fable 5 at half the price.
Force two: Chinese open-source models going mainstream. Low-cost open-source models — code-public AI models companies can deploy themselves — led by DeepSeek, are being adopted at scale, pulling down the pricing anchor for commercial APIs.
This means → the decline is not one vendor's promotion; two structural forces — closed-source giants undercutting each other + open-source alternatives flooding in — are compressing margins simultaneously.
What does this mean for enterprise users and AI providers?
For enterprises making heavy API calls, operating costs drop — the same budget now buys more inference.
For AI service providers, per-unit revenue shrinks. This means → they must either scale volume to offset lower prices, or find value-added revenue streams beyond inference.
Jefferies' team notes that both the US and Chinese tech ecosystems are "increasingly emphasising cost efficiency." This reflects a shift: cost reduction has moved from competitive tactic to industry consensus.
Will prices keep falling?
With the price war compounded by open-source proliferation, a floor has not been confirmed — further penetration of open-source models could force commercial API prices lower still.
In plain terms = the endgame depends on how far open-source models can substitute for paid APIs; the higher the substitution rate, the weaker the pricing power of paid services.
This is the key variable to watch next: when does the floor arrive, and who exits first.
Content is for reference only, not financial advice.