# DeepSeek Overtakes Google: What the Token Data Says > Source: > Published: 2026-08-17 08:09:27+00:00 In April, DeepSeek accounted for under 1% of token volume on Vercel’s AI Gateway. Google ran 40%. By July, those numbers had flipped: DeepSeek at 25%, Google at 10.7%. The [Vercel AI Gateway Production Index for August 2026](https://vercel.com/blog/deepseek-overtakes-google-on-volume-cost-per-token-falls) put official numbers on a shift that many developers missed while it was happening. If you haven’t looked at your model selection strategy recently, this data suggests you should. ## One Model Is Running the Show DeepSeek V4 Flash is not just popular — it ran more tokens in July than all Google models on the gateway combined. It accounted for roughly 19% of total gateway tokens, with 70% more volume than the next-closest model. Its pricing explains a lot: at $0.14 per million input tokens and $0.28 per million output tokens (before the new peak/off-peak structure that took effect August 16), V4 Flash costs roughly one-sixteenth the gateway average. At those rates, it is not a model developers are testing — it is running their production workloads. The benchmark story supports the adoption. DeepSeek-V4-Flash-0731, released July 31, [beat the flagship V4-Pro-Preview on all nine published agent benchmarks](https://artificialanalysis.ai/models/deepseek-v4-flash). On coding tasks specifically — the gateway’s largest workload by tokens — DeepSeek models collectively handle around one-third of volume. ## The Anthropic Paradox Here is where the data gets interesting. While DeepSeek is winning on volume, Anthropic is winning on money — and by a wider margin than before. Anthropic collected 65.1% of total gateway spend while running 29.8% of tokens. Its price premium compared to the average across other labs was 4.4x in July, up from 3.4x in June. Anthropic’s premium is expanding, not compressing, even as its volume share shrinks. Claude Fable 5 returned on July 1 after a three-week export suspension and immediately captured 13.2% of all gateway spend. Nine in ten teams that used Fable 5 in July had never used it before — that is new demand, not returning customers. Developers with high-stakes workloads are actively choosing to pay 4x more for Anthropic. The market is not moving to cheap models uniformly. It is splitting. ## 81% of July’s Tokens Ran on Models That Didn’t Exist on the Gateway Six Months Ago That single stat captures how fast this market is moving. [Kimi K3 from Moonshot AI](https://simonwillison.net/2026/Jul/16/kimi-k3/) launched July 16, published open weights July 27, ranked first in Frontend Code evaluation on Arena at 1,679 points, tripled its daily volume within a month, and was in the top ten by token volume by month end. The model did not exist on the gateway in January. It is now a meaningful share of production traffic. Three of four teams running ten million or more tokens per month changed their model mix by at least 10% in July. Three in five changed it by at least 25%. If you haven’t audited what models your application is routing to in the last six months, you haven’t audited it during the period in which the entire landscape changed. ## Switching Costs Have Effectively Disappeared The speed of this shift is not an accident. AI gateways — Vercel’s, OpenRouter, LiteLLM — have reduced model switching to a configuration change. In the Vercel AI SDK, the model string is the only thing that changes between `anthropic/claude-opus-4.8` and `deepseek/deepseek-v4-flash` . No migration, no refactor, no new API integration. The model you’re using is now an operating decision, not an architectural one. That is why three-quarters of teams are making active adjustments month over month. ## The Cost Math July’s aggregate numbers: token consumption grew 59%, total spend grew 37%, and the average price per token fell 13.6%. One in four teams cut costs by more than 30%. The market is getting larger and cheaper simultaneously — the signature of a commodity tier forming beneath a still-premium layer. Total gateway spend is up 74% since May. Usage is real and growing; costs are falling. Both are true at once. ## What This Means for Your Stack The split in this data points to a reasonable default strategy: use cost-optimized models — DeepSeek V4 Flash, Kimi K3, GPT-5-Nano — for high-volume routine work: code generation at scale, background agents, batch summarization. Reserve premium models (Claude Fable 5, Opus 4.8) for work where quality directly affects outcomes: complex debugging, security analysis, customer-facing reasoning. The infrastructure to support that split — an [AI gateway that lets you mix models per route or task type](https://www.truefoundry.com/blog/ai-gateway) — is available and mature. The Vercel index covers data through July 2026. DeepSeek’s new peak/off-peak pricing structure took effect August 16. The next index will show whether V4 Flash’s dominance holds under the new pricing model. Either way, the pattern is clear: this market is not settling. Build for the assumption that your model selection will change again.