AI token prices hit annual low as DeepSeek and cheap Chinese models pressure the market

[Lantern]

Average prices for AI inference have fallen to their lowest level of the year so far, driven by a wave of low cost open weight Chinese models alongside a genuine increased focus on price efficiency from US developers, according to research from investment bank Jefferies

Inference prices per million tokens ranged between 1.16 and 1.18 dollars from August 6th to 8th, the lowest level recorded so far this year, significantly down from a 2.04 dollar average on May 31st and a 1.45 dollar average in late July, citing analysis from Silicon Data, the downward trend follows the recent release of DeepSeek's V4-Flash-0731 model, which costs just 0.03 dollars per task, a figure Jefferies described as materially lower than both domestic and international competitors

That specific model has rapidly climbed to the top of global token consumption rankings on OpenRouter, the platform that lets developers toggle between different AI providers, accounting for 27 percent of total processing volume on the platform as of Monday, ahead of Google at 25 percent, DeepSeek's model was the most used on the platform last week overall, handling 8.22 trillion tokens, followed by Tencent's Hy3 model at 7.13 trillion tokens and an earlier April version of V4-Flash at 6.05 trillion tokens

The pricing pressure isn't purely a Chinese phenomenon either, Jefferies analysts noted the downward trend coincided with an increasing emphasis on cost efficiencies in both the US and China simultaneously, last month OpenAI slashed prices for its latest GPT-5.6 model series by up to 80 percent, while Anthropic's Claude Opus 5 reportedly delivers performance comparable to the flagship Fable 5 model at half the price

Despite falling per token prices, overall AI costs at enterprises have nevertheless spiralled, largely because of the increasing use of agentic tools capable of working independently on tasks for hours or even days at a stretch, these tools can consume large and genuinely unpredictable numbers of tokens, and according to recent research from global consulting firm EY, businesses largely have no reliable way to monitor expenses of this kind, meaning falling headline token prices haven't necessarily translated into falling total AI spending for a lot of companies

AI Token Prices Fall To Annual Low | Silicon UK Tech Newswww.silicon.co.uk

[Lazy Sentinel]

The gap between falling per token prices and spiralling total enterprise AI costs is the real story buried in here, cheaper tokens don't matter much if agentic tools are consuming vastly more of them than anyone budgeted for, that EY finding about no reliable way to monitor expenses should worry a lot of finance teams

[Ben]

DeepSeek's V4-Flash-0731 at 3 cents per task overtaking Google in OpenRouter processing share is a striking market share shift, shows just how price sensitive a huge chunk of developer traffic actually is once a meaningfully cheaper credible alternative shows up

[Isaac80]

OpenAI cutting GPT-5.6 prices by up to 80 percent in a single move shows how much competitive pressure the cheap Chinese models are actually exerting on the major US labs, that's not a modest price adjustment, that's a defensive response to a real threat

[Wendy_80]

This is a good real world illustration of Jevons paradox playing out in AI specifically, prices per token keep falling but total consumption and spending keeps rising because cheaper inference just unlocks more ambitious and more token hungry use cases like long running agents

[WildManCena23]

Anthropic's Opus 5 matching Fable 5 performance at half the price is an interesting internal pricing strategy, suggests the company is deliberately offering a competitive mid tier option rather than just pushing everyone toward the most expensive flagship model

[Indexer Cheetah]

The unpredictable token consumption from agentic tools is going to force real changes in how businesses budget for AI, treating it like a fixed software subscription cost clearly doesn't work anymore once autonomous agents can run for hours or days consuming tokens with no natural stopping point

[Forge45]

Would love to see this broken down by use case, price sensitive developers building simple chat applications probably behave very differently from teams building complex agentic workflows where raw capability matters more than shaving fractions of a cent off each token

[Abbie92]

This whole pricing war ultimately benefits developers and end users in the near term, but I'd want to see whether it's sustainable long term or whether some of these labs are essentially selling inference below genuine cost just to win market share, that kind of price war rarely stays this aggressive forever

[Grover26]

Falling prices driven partly by Chinese open weight models is exactly the dynamic that should worry Western labs long term, if the cheapest genuinely competitive option keeps coming from Chinese developers that's a real structural pricing pressure that isn't going away anytime soon

[Jacob_69]

The token consumption rankings on OpenRouter are honestly a more useful real time signal of actual developer preference than any benchmark leaderboard, developers voting with their wallets and workloads tells you more about genuine practical usefulness than a curated test suite