Vercel's Token Pass Gives You $80 of Kimi K3 and GLM 5.3 for $30. No Rollover, No Overage, and a 403 When It Runs Out.

Vercel introduced Token Pass today, October 15, 2026, in early access for Pro teams: $30 a month buys $80 of monthly credit on two Chinese open-weight models, Z.ai's GLM 5.3 and Moonshot's Kimi K3, served "currently through Wafer," with zero data retention by default and inference in the US and EU. On the regular AI Gateway, team-wide zero data retention costs $0.10 per 1,000 requests, so the default alone has a price.

The arithmetic looks simple, $80 for $30, and the docs make it less simple in two places. First, the credit is spent "at the Token Pass provider's per-token rates, shown on the model page," which means inside the dashboard, after you subscribe. The nearest public number is Wafer's rate on the open Gateway, and we use it below, labeled as a stand-in. Second, the terms: no rollover, no overage, a 403 when the credit runs out with no fallback to other providers, no fast modes, and a clause that lets Vercel change the models or the provider at any time. We read the changelog, the Token Pass docs, the billing and routing pages, the Gateway endpoint data, and Z.ai's pricing page. Here is what held up.

The Offer, With the Terms Attached

The 403 is the operational catch. An agent loop that exhausts the credit mid-run does not degrade; it fails, with no fallback. Handle the 403 in your code and keep a non-Token-Pass key for the rest of the month.

"Currently through Wafer" is the quality catch. The provider can change. So can the models. Your rates and your outputs can move under the same plan name, which is a familiar pattern this month.

What the Credit Buys, at Wafer's Public Rates

The rates below are Wafer's rows on the public AI Gateway, which we take as the stand-in for the dashboard rate the docs do not print; first-party rates are from Moonshot and Z.ai.

By our arithmetic, Wafer against first-party is about 87% cheaper on GLM 5.3 input, 68% cheaper on output, and 31% cheaper on cache reads: a clean win. On Kimi K3 it is about 92% cheaper on input and 24% on cache reads, and about 7% more expensive on output, $16 against $15. Kimi K3's reasoning is always on and bills as output, as we noted when it launched, so real Kimi traffic leans toward the expensive side of that table.

Three illustrative Kimi K3 workloads, our arithmetic at the rates above:

• Output only: $80 of credit buys about 5.0M output tokens at Wafer's rate against about 5.3M at Moonshot's. You still come out ahead on the $30 you paid; the credit is just worth a little less than face value.
• 10M uncached input plus 1M output: about $18.40 at Wafer against $45.00 at Moonshot.
• 50M cached input, 2M uncached, 1M output (a cache-heavy agent loop): about $27.90 at Wafer against $36.00 at Moonshot.

So: GLM 5.3 is a strong deal across the board; Kimi K3 is a strong deal for input-heavy and cache-heavy work and a narrow one for reasoning-heavy output.

One thing on the launch graphic: a "DeepSeek V4.1 Flash" card appears under "More models coming soon." It is an image, not a line in the docs, so it stays unverified today.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Subscribe if you are a Pro team spending more than $30 a month on GLM 5.3 at first-party rates, or on input-heavy Kimi K3 work, and you can write a 403 handler. Check the rate on the dashboard model page before you commit; that is the number you are billed at, and this post could only print the public one.

Skip it if your Kimi K3 traffic is reasoning-heavy output, if you need fast modes, if you are on Enterprise, or if an agent that stops dead at zero credit is a problem you cannot engineer around. And treat the plan as a moving target by its own terms.

The frontier is elsewhere: on AA's index neither model is near the top, and that is not the point of a $30 pass. Our model comparison has the current standings side by side.

Update, October 25, 2026

• October 25: The Token Pass docs now read differently in two ways that matter. Featured models are Kimi K3, GLM 5.3, and, new, DeepSeek V4.1 Flash, so the launch graphic's card came true. And the billing page now says credit deducts at the model's list price, which it defines as the model provider's own API rate: Kimi K3 at $3.00 / $0.30 cached / $15.00, GLM 5.3 at $1.40 / $0.26 / $4.40, DeepSeek V4.1 Flash at $0.15 / $0.003 / $0.60. If that table is what Token Pass bills, the Wafer-rate comparisons above are historical, and the Kimi output catch no longer applies: $80 of credit at list simply buys $80 of first-party tokens for $30, on any of the three. Re-check the dashboard rate against the docs table before you rely on either. Still true on October 25: served through Wafer, zero data retention, no rollover, 403 when empty. Not on Token Pass: Kimi K3.1 Pro, which launched October 23 and scores 55 on AA's index; anyone quoting a K3.1 Token Pass rate is inventing it.

Sources

• Vercel changelog: introducing Token Pass
• Vercel docs: Token Pass
• Vercel docs: Token Pass billing
• Vercel docs: Token Pass models and routing
• Vercel AI Gateway: Kimi K3 endpoints
• Vercel AI Gateway: GLM 5.3 endpoints
• Z.ai: pricing

*Not sure whether a prepaid pass or first-party billing fits your traffic shape? Take the free 2-minute quiz and get matched. Then read about the model Token Pass was built around and its always-on reasoning bill: Kimi K3.*