Yes, Vercel’s Token Pass is cheaper than paying list-price API credit for GLM 5.3 and Kimi K3, but only inside a narrow band. The subscription’s changelog states the terms plainly: $30 per month buys $80 of monthly credit for those two models. Any month your team’s list-price spend on the pair exceeds the $30 fee, the flat fee wins. Any month it falls below, you have paid the fee for credit you did not use, and that credit does not come back.
Before the math, two cautions that shape everything below. First, every subscription term here comes from a single vendor changelog for an early-access launch, so the terms may change without notice. Second, that changelog publishes no per-token prices for GLM 5.3 or Kimi K3, so token-level break-even cannot be computed and the comparison runs in dollars. Treat the arithmetic as solid and the surrounding feature claims as provisional.
What $30 actually buys
According to Vercel’s Token Pass announcement, the offer is in early access for Pro teams:
“Vercel Token Pass is now in early access for Pro teams. For $30 per month, your team gets $80 in monthly credit for GLM 5.3 and Kimi K3, with more models to come.”
Three properties of that sentence matter more than the headline ratio. The credit is monthly. It covers exactly two models today. And “more models to come” is a promise, not a contract term; if your workload splits across four models, the subscription covers half of it until Vercel says otherwise.
The mechanics of the credit, also from the changelog, define the real risk profile:
“The subscription renews monthly unless canceled. Unused credit doesn’t roll over. When your credit runs out, Token Pass requests stop until the next billing period, with no automatic overage charges.”
Non-rolling credit, a hard stop at exhaustion, and no overage path. Those three clauses turn a simple discount into a capacity-planning exercise, which is what the rest of this article is about.
The break-even you can compute today
Because no per-token prices are published, the cleanest comparison is dollar against dollar. The reasoning below is arithmetic on the stated terms, my inference rather than a vendor-stated figure.
- Spend under $30/month at list price on GLM 5.3 and Kimi K3: the subscription loses. You pay the fee for credit you only partially consume, and the remainder evaporates at renewal.
- Spend between $30 and $80/month: the subscription wins, from breaking even at $30 to an effective 62.5% discount against metered rates at full consumption.
- Spend over $80/month: the subscription wins on the first $80 but cannot serve the excess at any price. Requests stop until the next billing period.
The threshold that decides the question is 37.5% utilization of the $80 credit. Below it, stay metered. Above it, subscribe. That is the whole formula, and it is computable today from your existing gateway invoices if you already route these models, or from a metered trial month if you do not.
| Monthly list-price spend on GLM 5.3 + Kimi K3 | Subscription outcome | Action |
|---|---|---|
| $0–$30 | Fee forfeited on unused credit | Stay metered |
| $30–$80 | Cheaper than metered; up to 62.5% effective discount | Subscribe |
| $80+ | First $80 cheap, excess unservable | Subscribe plus preconfigured metered failover |
| Spiky or unknown | Idle months lose the full $30 | Measure one metered month first |
One demand-side wrinkle affects which band you land in. An author-reported study found that “Contracts cost +13% agent tokens and +38% wall-clock time” when the same coding agents ran under explicit delegation contracts with required evidence sections rather than plain issue-style prompts, with the effect larger for weaker model tiers. The lesson for budgeting is narrow but real: prompt and harness structure moves token consumption even with the model held fixed. If your setup enforces contract templates and evidence bundles, the $80 credit burns faster than a plain-prompt estimate would predict. Treat the figures as directional: the study was executed end-to-end by an AI research assistant under a prior written design, and it has not been independently audited.
The catches the headline hides
The $30-for-$80 framing invites a blanket “cheaper” verdict, and the changelog’s own terms are the strongest argument against it.
No rollover shifts idle-month risk to you. A team that subscribes in October and ships nothing in November has spent the $30 fee for zero requests. Metered billing never has this failure mode. If your usage is project-based rather than continuous, the subscription is a bet on your own consistency.
The hard stop is a reliability event, not a billing event. When the credit exhausts mid-month, Token Pass requests simply stop. There is no overage toggle, no grace buffer, no paid extension. For an interactive coding session that is annoying; for an overnight agent run processing a backlog, it is a silent failure that surfaces as missing work the next morning. The vendor provides no failover mechanism, so you must build one (more on that below).
“More models to come” is marketing language. The roster is GLM 5.3 and Kimi K3. If your routing strategy assumes a third model joining the pass, that assumption has no contractual basis. Plan against the two-model roster and treat any expansion as a pleasant surprise.
ZDR by default
The changelog’s compliance language is the strongest non-price argument for the pass:
“Token Pass has Zero Data Retention (ZDR) enabled by default. Requests run through selected providers with inference in the US and EU. AI Gateway and the provider delete your prompts and responses after each request, and your prompts aren’t used to train models.”
For teams routing proprietary code through agents, per-request deletion with no training use is a materially better default than retention-based API terms, and the US/EU inference constraint gives compliance reviews a concrete jurisdiction story. Default-on matters here: there is no toggle to forget, no plan tier to misconfigure.
Two cautions. First, this is vendor wording in a changelog, not an audited control. If your compliance posture requires contractual guarantees rather than product-page language, ZDR-by-default in early access is a starting point for procurement, not the end of one. Second, the changelog covers Token Pass only; whatever else you route needs its retention terms checked on their own merits.
What you are actually routing
The subscription is only interesting if GLM 5.3 and Kimi K3 are worth routing agent traffic to in the first place. What is known about the two models is asymmetric, and that asymmetry should temper the decision.
On the Kimi side, the record is substantial but self-reported. According to the Kimi K3 technical report, K3 is “a 2.8T parameter Mixture-of-Experts model with 104 billion activated parameters, native vision capabilities, and a 1-million-token context window”, a large step up from Kimi K2 at 1 trillion total parameters with 32 billion activated, and from the intermediate Kimi K2.5, built on a 1.04 trillion-parameter, 384-expert backbone with 8 activated per token. More relevant to the harness question, K3’s authors trained it in an RL environment that “can instantiate mainstream harnesses such as Kimi Code [56], Claude Code [14], Codex [19], OpenClaw [86], and Hermes [44].” If you are routing into Claude Code or Codex, K3 was quite literally trained against that shape of workload.
The authors’ cost claims are the reason this subscription exists. On Kimi Code Bench 2.0, they report that “Kimi K3 is 4.0 points behind Claude Fable 5 at 38% of its cost, and at high effort it already matches Claude Opus 4.8’s maximum-effort score at roughly one third of the cost.” On BrowseComp, the same report gives the best score (91.2%) at $2.03 per task, half the cost of GPT-5.6 Sol and an order of magnitude cheaper than the Claude models at maximum effort. Read those numbers with both hands on the caveats: every benchmark is self-reported, Kimi Code Bench is the authors’ own benchmark, and even there K3 trails Claude Fable 5 by 4.0 points. The claim “frontier capability at a third of the cost” is a hypothesis the authors have argued, not a result independent evaluation has confirmed.
On the GLM 5.3 side, the changelog offers nothing beyond the name: no technical report, no benchmarks, no pricing, no usage data. Half the subscription’s roster is, for now, a name on a changelog. That does not make GLM 5.3 bad; it makes any capability claim about it unverifiable today. If GLM 5.3 is central to your routing plan, evaluate it metered first and let your own traces decide.
Planning for the hard stop
Since the vendor stops requests at exhaustion and provides no failover, the subscriber has to. The pattern is straightforward in principle: run Token Pass as the primary route for GLM 5.3 and Kimi K3 traffic, monitor credit consumption through the billing period, and configure a second route for the same models on metered billing that activates when the pass stops responding. Verify what that metered route charges before assuming the failover is cost-neutral.
The design detail worth getting right is where the switch happens. A failover that requires a human to notice dead requests at 2 a.m. is not a failover. Route selection should key on request failure or a credit-exhaustion signal, not on a calendar reminder. And because unused credit never rolls over, the failover should not be shy: hoarding the $80 credit to avoid touching metered billing is how teams end the month having paid the fee for a fraction of its value.
I would treat the first subscribed month as a measurement exercise: leave the metered route live, log what fraction of spend lands on the pass, and only tighten the routing once you know whether your team’s natural volume sits in the $30–$80 sweet spot or blows through the ceiling in week two.
Verdict: when to subscribe, when to stay metered
Subscribe to Token Pass for months when your team’s list-price spend on GLM 5.3 and Kimi K3 reliably exceeds $30, which is 37.5% utilization of the $80 credit. Below that threshold, metered billing is strictly cheaper because idle credit is forfeited at renewal. Above the credit ceiling, the subscription still wins on the covered spend but cannot serve the rest, so a preconfigured metered failover is a precondition, not an option. Treat “more models to come” as a promise, and treat the ZDR-by-default terms as valuable vendor-stated posture that still needs contractual confirmation for regulated workloads.
One final plain statement on the capability side. Kimi K3’s cost-performance numbers are self-reported on its authors’ own benchmark and include a measured 4.0-point deficit to Claude Fable 5; GLM 5.3 has no capability evidence available to check; and the delegation-contract overhead figures are author-reported and directional. The dollar break-even is arithmetic you can trust. Almost everything around it deserves a verification pass before you build a routing strategy on top of it.
Frequently Asked Questions
What happens to unused Token Pass credit at the end of the month?
The subscription renews monthly unless canceled. Unused credit doesn’t roll over. When your credit runs out, Token Pass requests stop until the next billing period, with no automatic overage charges.
What is the break-even point for Vercel’s Token Pass?
The threshold that decides the question is 37.5% utilization of the $80 credit. Below it, stay metered. Above it, subscribe. That is the whole formula, and it is computable today from your existing gateway invoices if you already route these models, or from a metered trial month if you do not.
Does Token Pass include Zero Data Retention by default?
Token Pass has Zero Data Retention (ZDR) enabled by default. Requests run through selected providers with inference in the US and EU. AI Gateway and the provider delete your prompts and responses after each request, and your prompts aren’t used to train models.
Join the discussion
Share a useful perspective or ask a question about this article.