Why Open-Weight AI Models Surged on Vercel’s AI Gateway

A cost-and-capability tradeoff is reshaping traffic on Vercel’s AI Gateway. Open-weight models accounted for 54% of gateway token volume on Tuesday, August 25, 2026, compared with 46% for proprietary models. On the preceding Saturday, their share reached a reported record of 62%, leaving closed models with 38%.4

The immediate catalyst was DeepSeek V4 Flash, whose updated weights improved performance on coding, tool use, and agentic workflows without turning it into a premium-priced model.19

The token-volume shift

Reported point in August	Open-weight models	Proprietary models
Preceding Saturday	62%	38%
Tuesday, August 25	54%	46%

That means open-weight models held an 8-point lead on Tuesday and a 24-point lead at the earlier record. The comparison is specifically about tokens processed through the gateway. It should not be read as a measure of revenue, profit, or the number of businesses using each model.

Why DeepSeek V4 Flash mattered

Vercel said DeepSeek V4 Flash began running updated weights by default on AI Gateway, with its Terminal-Bench score rising to 82.7 from 56.9 in the April preview—a 25.8-point improvement.19 The update required no change to the model ID or application code, reducing the friction for existing users to receive the newer weights.19

The model’s appeal was therefore broader than simply being inexpensive. Its improved agentic performance made it more suitable for high-volume coding and tool-use workloads, while its efficiency-tier positioning gave developers a way to control inference costs. Vercel describes the model as supporting reasoning, tool use, and implicit caching, with a context window of 1 million tokens.27

Pricing helped reinforce that positioning. Reuters reported that DeepSeek’s V4 Pro launched at roughly nine times V4 Flash’s input price and 14 times its output price, highlighting the gap between the company’s efficiency model and premium offering.17 Independent comparisons also show a substantial price advantage for V4 Flash, although benchmark results vary by test and configuration.25

Why proprietary models lost token share

The available reporting links the decline in proprietary-model volume to weaker business demand for Anthropic’s more expensive Fable 5.4 That explanation is plausible for workloads where developers care more about throughput and unit economics than maximum model capability, but the available gateway figures do not prove a single cause for every customer or request.

Fable 5 remains positioned for long-running, ambiguous, multi-step tasks, according to Vercel’s product description.3 In other words, the August traffic change looks less like a universal verdict that open models are better and more like a routing decision: use a lower-cost model for large volumes of suitable work, and reserve premium proprietary models for tasks that justify their higher price.

Token volume is not AI spending

The most important caveat is the difference between usage and revenue. Earlier Vercel reporting showed open-weight models rising to 29% of gateway token volume while accounting for under 4% of spend.44 That gap exists because token counts do not capture the price paid per input or output token.

So the August crossover signals that open-weight models were processing more of the gateway’s workload, not necessarily generating more business for providers. Proprietary models can retain a disproportionate share of spending even while cheaper open models handle more tokens.

What the shift means for developers

The data points toward a more deliberate model-routing strategy:

• Use efficiency models for scale: High-volume generation, coding assistance, and routine tool calls can be especially sensitive to per-token costs.
• Re-test after weight updates: DeepSeek V4 Flash’s upgrade changed its performance profile without requiring an application-level model-ID change.19
• Separate quality from economics: A benchmark improvement does not establish that a model is best for every production workload; real-world accuracy, reliability, latency, privacy, and provider availability still need evaluation.
• Track spend as well as tokens: A token-volume leader is not automatically a spending or value leader.

The broader lesson from Vercel’s August traffic is that open-weight models can win production volume when they combine adequate capability with sharply lower costs. DeepSeek V4 Flash supplied the clearest example, but the underlying shift is toward matching each workload with the least expensive model that performs reliably enough—not replacing every proprietary model at once.