DeepSeek V4 Doubled Its OpenRouter Token Share in 60 Days and Is Now the Platform's #1 Model
DeepSeek V4, released April 24, has turned a structural decline into a market-share recovery. OpenRouter’s analytics show the company doubled its weekly token share on the platform from 9% in January to 18% by June 2026. Since mid-May, it has been the single most-used model on OpenRouter, a result confirmed independently by the Wall Street Journal.
The path there was not linear. DeepSeek had held just under 10% of OpenRouter’s weekly token flow at the start of 2026. When agentic workloads began accelerating in February and March, that share collapsed to 5%. Proprietary models dominated the agentic tier. DeepSeek’s open-weight models, up to and including V3, could not handle the multi-step tool use and long-horizon task structure that agentic pipelines require at production scale.
V4 changed that calculation. OpenRouter characterises it directly: V4 is the first DeepSeek model sufficient for agentic workloads.
Why Agentic Sufficiency Matters
The token-share collapse in early 2026 was not about price. DeepSeek V3 was already cheap. The issue was capability at task completion rate. Agentic pipelines fail on different margins than simple completions: tool call accuracy across a 30-step trajectory, context retention across a 128K+ window, error recovery without human intervention. V3 fell short on enough of those dimensions that teams routing real agent workloads reached for Claude Opus 4.6, GPT-5.4, or Gemini 3.1 Pro instead.
V4 Pro’s benchmark profile addresses those gaps. SWE-bench Verified at 80.6% puts it in the same cohort as Claude Sonnet 4.6 and ahead of Gemini 3.1 Pro on raw coding capability. DeepSeek V4 Flash at 79.0% SWE-bench Verified and MIT licence gives teams a cheaper routing option for lower-priority agent tasks. The architecture paper documented 90% KV cache reduction at 1M tokens, which directly improves long-context agent economics.
The Token-Share Numbers
| Period | DeepSeek Token Share | Notes |
|---|---|---|
| January 2026 | ~9% | Pre-V4, V3 in market |
| February-March 2026 | ~5% | Agentic workload surge, V3 insufficient |
| April 24, 2026 | V4 released | Pro, Flash, and a third variant simultaneously |
| June 2026 | ~18% | #1 on OpenRouter since mid-May |
The 18% figure represents roughly double the share of the nearest competing open-source model family, based on OpenRouter’s overall distribution. Chinese open-source models collectively account for a significant share of the non-proprietary tier, and V4’s recovery pulled much of that share back to DeepSeek specifically.
Pricing Held the Floor
DeepSeek locked V4 Pro’s 75% price cut as permanent from June 1, setting the output rate at $0.87 per million tokens. V4 Flash runs below $1 input. For teams running high-volume agent pipelines, those numbers make a material difference compared to frontier proprietary pricing at $15-75 per million tokens.
The price-capability combination is what OpenRouter characterises as the agentic pivot: a model that can actually complete agent tasks, priced below the point where teams feel forced to route toward proprietary alternatives.
What the Shift Signals
A model becoming #1 on a multi-provider routing platform by token share is a different signal than leaderboard ranking. OpenRouter routes based on actual developer choices under real cost constraints. The shift reflects agentic pipelines being re-evaluated: teams that defaulted to proprietary models during the February acceleration are now mixing V4 in for price-sensitive agent tasks.
The competitive implications run in both directions. For Anthropic, OpenAI, and Google, it confirms that the open-weight tier has closed enough of the agentic capability gap that cost is again a variable buyers optimise on. For DeepSeek, the token-share lead is a validation of the V4 architecture decisions around KV efficiency and long-context handling, which were designed specifically for production agent deployment at scale.