Token Pass explained: what it changes for small teams
What launched, and what it changes for a small team
Token Pass is a monthly subscription add-on, not a new platform: it buys a fixed credit balance for two open-weight models, Kimi K3 and GLM 5.3, that run through Vercel's AI Gateway. For a small team, it trades per-token flexibility for a hard monthly ceiling, so this month's decision is whether your model usage is steady enough to stay under that ceiling.
Vercel's launch post, dated October 15, 2026, puts the fee at $30 per team per month and the included credit at $80 per billing period, according to Vercel (2026). The post says Token Pass is "now in early access for Pro teams," so eligibility is limited to Pro plans, and the Token Pass tab shows a waitlist when early-access spots are full.
In plain terms, a token pass is a prepaid, capped allowance of model usage sold for a flat monthly fee. It is not a seat license, and it does not replace the rest of your AI Gateway usage.
What changes is the billing pattern. Before this launch, a team that wanted Kimi K3 or GLM 5.3 through AI Gateway drew from a prepaid AI Gateway credit balance at each provider's rates. AI Gateway's pricing documentation says it charges no markup and no platform fee on tokens. Token Pass adds a second, separate balance. Its credit pays only for the featured models, and Token Pass requests do not touch your AI Gateway credits, as the Token Pass billing documentation explains.
So the change for a buyer is not a new model or a new gateway. It is a choice between a fixed monthly allowance for two models and metered spending across everything else.
Key Takeaways
Token Pass is a fixed monthly allowance for two featured models, Kimi K3 and GLM 5.3, and it does not pay for anything else.
Unused credit expires at the end of each billing period, and there is no overage: requests fail until the period resets.
A Token Pass key cannot fall back to other providers or to regular AI Gateway credits, so production traffic needs a separate regular key.
Token Pass is limited to Pro teams and is in early access; Enterprise teams cannot subscribe.
Zero data retention is on by default for Token Pass requests, which matters if your prompts contain customer data.
The Token Pass provider's per-token rate is not broken out on the model pages we opened, so a real cost comparison needs your own usage numbers.
How we evaluated these options
This is a desk comparison built from public pages. It uses no hands-on trials, and nothing in it is a measured result. We weighted seven criteria by how much each one changes a small team's monthly decision, and the weights add to 100%. Each criterion is scored from 1 to 5 from the evidence in the sections below, and the scores are our judgment. The two alternatives are Vercel's regular pay-as-you-go AI Gateway, which is the previous way to buy the same models, and OpenRouter, the main rival in this category.
| Criterion | Weight (%) | Score scale | Fits a small team when |
|---|---|---|---|
| Monthly cost ceiling | 20 | 1 to 5 | The fixed fee and cap sit inside a budget set in advance |
| Model coverage | 20 | 1 to 5 | Your agents use only the models the option covers |
| Behavior at the limit | 15 | 1 to 5 | Stopping requests is acceptable; billing on past the limit is not |
| Data retention defaults | 15 | 1 to 5 | Zero data retention is on by default or set per request |
| Eligibility and plan fit | 10 | 1 to 5 | Your team is on a plan the option accepts |
| Operational effort | 10 | 1 to 5 | Keys, routing and fallback fit existing code with little change |
| Provider and terms risk | 10 | 1 to 5 | The team accepts that models and providers can change at short notice |
The two largest weights go to the cost ceiling and model coverage, because those decide whether Token Pass is a budget tool or a poor fit for your stack. Data retention and behavior at the limit come next, because they decide what breaks in production. If you are weighing a different kind of vendor comparison, our property management vendor automation platform comparison uses a different set of criteria.
What Token Pass replaces, and what it does not
Token Pass does not replace regular AI Gateway usage. It sits beside it, and the difference shows up in the rows that matter when money or uptime is on the line.
| Feature | Token Pass (Vercel) | Regular AI Gateway (Vercel) | OpenRouter (rival) |
|---|---|---|---|
| Billing model | Fixed monthly add-on; credit drawn per request | Prepaid credits, per request at provider rates | Prepaid credits, per request at list price |
| Models covered | Kimi K3 and GLM 5.3 only; fast variants excluded | Any model in the AI Gateway catalog | Posted catalog; not checked model by model |
| When the budget runs out | Requests fail until the next billing period; no overage | Requests stop when credits run out; budgets and auto top-up can control this | Each request deducts from the prepaid balance |
| Fallback | None; Token Pass requests never fall back | Routing options can pick providers; BYOK fallback is charged to credits | Not checked on the pages we opened |
| Data retention default | Zero data retention on by default | Per-request ZDR at no extra cost; team-wide ZDR carries a fee | Not checked on the pages we opened |
| Eligibility | Pro teams only; Enterprise cannot subscribe | Every team account; free tier covers a subset of models | Accounts; Enterprise priced by contract |
| Setup effort | Separate Token Pass key, stored as an environment variable in deployments | Standard key and optional routing options | Not checked on the pages we opened |
The row that matters most is the budget limit. Regular AI Gateway keeps working as long as the credit balance covers it, and you can cap spend with budgets or top up automatically. Token Pass stops. Vercel's models and routing documentation confirms that Token Pass keys never fall back to other providers or to your AI Gateway credits, so a stop is a stop.
Data handling is the strongest point on the Token Pass side. The launch post says zero data retention is enabled by default, that inference runs with selected providers in the US and EU, and that prompts are not used to train models. If your team uses Kimi K3 or GLM 5.3 for work that touches customer records, that default is worth a review with your security owner.
Pricing and total cost for one scenario
Pricing checked October 17, 2026. The Token Pass fee comes from the launch post cited above. The pages we opened do not break out a separate Token Pass per-token rate, so the scenario below uses the list rate Vercel publishes for Kimi K3. The Pro plan is required for Token Pass and is left out of the comparison, because it is a cost you pay regardless of the option you pick.
Kimi K3 rates: $0.27 input, $11.4 output per million tokens according to Vercel (2026). Take a team that sends 40 million input tokens and 4 million output tokens through Kimi K3 in a month. The input side is 40 × $0.27, or $10.80, and the output side is 4 × $11.4, or $45.60, for a total of $56.40. Against the $80 monthly credit described above, that leaves $23.60, which covers about 2.07 million more output tokens at the same list rate. The scenario is illustrative, not a customer, and the Token Pass provider's own rate may differ from the list rate. To confirm which requests drew the credit, check providerMetadata.gateway.routing.resolvedProvider in the response, which Vercel's models and routing documentation says reads token-pass for Token Pass requests.
The main comparison most buyers will raise is OpenRouter. On OpenRouter's Standard tier the credit purchase fee is 5.5%, and it is 8% on Business, according to OpenRouter (2026). OpenRouter says that fee applies when you buy credits, not on individual requests, and that inference is billed at the provider's list price on every plan.
| Cost line | Token Pass (Vercel) | Regular AI Gateway (Vercel) | OpenRouter (rival) |
|---|---|---|---|
| Input: 40M tokens at $0.27 per 1M | $10.80 | $10.80 | Not checked |
| Output: 4M tokens at $11.4 per 1M | $45.60 | $45.60 | Not checked |
| Scenario total | $56.40 | $56.40 | Not checked |
| Fixed monthly fee | $30 | None | None |
| Credit left after scenario | $23.60 | Not applicable | Not applicable |
| Fee on credit purchases | None listed | None on tokens | 5.5% Standard; 8% Business |
Enterprise pricing is Quote-based on both sides. Token Pass is not open to Enterprise teams, and Vercel's AI Gateway documentation says Enterprise teams can pay by invoice, which removes payment processing fees. OpenRouter prices Enterprise by contract.
Vendor profiles: best fit, limits and evidence
Token Pass on Vercel
Token Pass fits a Pro team whose engineers already run Kimi K3 or GLM 5.3 in coding agents and want one shared monthly ceiling for that work. The launch post names fx, Command Code and Pi as compatible agents, and says a personal agent can also be built with the AI SDK.
The coverage is narrow. Only the two featured models count, fast variants of them do not, and image, video, audio and embedding models fall outside the credit. Vercel may change the featured models and the provider that serves them at any time, and the current provider is Wafer, a Non-Vercel Service that follows its own terms. Credit does not roll over. If you cancel, the change takes effect at the end of the billing period, and Vercel does not refund that period. Vercel can also remove unused credit without a refund if it believes Token Pass is being misused, including through multiple accounts.
Implementation is short. You subscribe from the Token Pass tab in AI Gateway, create a Token Pass API key, and connect your agent. A Token Pass key cannot be changed to a regular key later. For a deployed app, store the key as an environment variable, because a deployment's OIDC token cannot spend Token Pass credit. Plan the switch to a regular key before the credit runs out, since Token Pass does not top up. Evidence: Token Pass overview, credits and billing, models and routing.
Regular AI Gateway, the previous way to buy these models
Regular AI Gateway fits teams that run several models, need production traffic to keep flowing, or want spend controls they can change. Requests keep working while the credit balance covers them, and you can cap spend with budgets or turn on auto top-up. Regular keys can route to other providers when you name them, and when a BYOK request fails and AI Gateway retries with system credentials, that fallback usage is charged to your credits.
The limit is that nothing caps spend by default. The ceiling is whatever budget you set, and a balance that runs out stops requests just as Token Pass does. Team-wide zero data retention carries a per-request fee, and that fee applies to Token Pass requests too, so check it before you turn the team-wide setting on. Per-request zero data retention has no extra cost on Pro.
Implementation means buying credits in the AI Gateway section, optionally setting auto top-up, and running the agent setup command without a key to create a regular key. Then remove any provider order that names token-pass. Evidence: AI Gateway pricing and the switch steps.
OpenRouter, the main rival
OpenRouter fits teams that want one prepaid balance, per-request billing and no subscription, and accept a purchase fee for that simplicity. Its pricing page says streaming is billed per token, the same as non-streaming, and that attempts that error are not charged for model tokens.
The costs sit in the purchase fee and in the BYOK terms. On Standard and Business, OpenRouter includes $25,000 of list-price BYOK inference per month with no fees and charges 5% above that, according to OpenRouter (2026). For a team that brings its own provider keys at high volume, that threshold matters more than the headline fee.
Implementation is the gap in this review. We did not open OpenRouter's setup documentation or check its model catalog model by model, so this article does not rate its coverage of Kimi K3 or GLM 5.3. Check that before you choose it for those two models.
When credit runs out: the operational risk
The failure mode is the part buyers underestimate. Once a Token Pass key's credit is used up, requests fail until the next billing period starts, and the application has to catch that failure and retry with a regular key itself. Setting a provider order alone does not switch keys. The team needs a rule for the last stretch of each billing period, and one named person who owns that rule. The models and routing documentation lays out the failover steps.
A configurable workflow could watch that rule instead of relying on memory. The trigger is a scheduled check of remaining credit. The prerequisite is that the team can read that balance or its usage programmatically or through an export; the Token Pass tab shows the balance, but we have not confirmed a programmatic read, so the team must check that first. The action compares remaining credit with a threshold the owner sets, such as a percentage of the $80 described above, and writes a dated line to a Slack channel using the pattern in our monday.com-to-Slack guide. The output is an alert, not a key change.
The human review points are the threshold, the alert recipients and the decision to switch. Moving from a Token Pass key to a regular key shifts spend from a fixed allowance to metered credits, so an owner should approve that step. If the team wants automatic failover, that is application code the team writes and tests. A proposed design from US Tech Automations would configure the schedule, the editable threshold and a log of each alert, with the owner's approval step kept in place. It would not change API keys on its own.
The do-it-yourself route is Zapier, Make or n8n, or a script your team writes. Those tools can support run histories, retries, error branches and audit evidence when they are configured. The buyer still has to design and own observability, idempotency, escalation, access controls and maintenance. A proposed US Tech Automations design would differ mainly by setting those review points up front, with the same prerequisites: API or export access, a named owner and a log of every alert.
Market context: usage billing is spreading
Token Pass arrives as buyers are getting used to metered AI tools. AI tool use or planned use: 84% according to Stack Overflow (2025), up from 76% in 2024. The same survey reports AI agents at work, used daily: 14.1% according to Stack Overflow (2025). Those are survey figures from respondents, not measures of spend.
GitHub moved Copilot to usage-based billing starting June 1, 2026, and its announcement sets out the individual tiers. Individual Pro plan: $10 per month, with $10 in AI Credits according to GitHub (2026). GitHub says base plan prices are not changing and that usage is counted on tokens, input, output and cached, at each model's listed API rates.
Read together, these point the same way: vendors are moving from flat seats toward metered usage, and Token Pass is a fixed-cap answer to that shift. Two data points are a signal, not proof of a market-wide change.
Signal vs speculation: what is confirmed and what is our read
Confirmed on vendor pages
The launch date, fee, included credit, featured models, eligibility, billing rules and data handling appear on Vercel's changelog and documentation pages linked above.
The current Token Pass provider is Wafer, and Vercel says it may change featured models and providers.
Unused credit expires, there is no overage, and cancellation takes effect at the end of the period with no refund for that period.
Vercel's changelog does not list an early-access spot cap, and its documentation says only that spots are limited.
Reported, not confirmed
Runtime Wire reports that Early-access cap reported: 500 Pro teams according to Runtime Wire (2026), citing a Vercel Developers post. Vercel's changelog does not list that cap, so treat the number as unconfirmed.
Our read (speculation)
A fixed ceiling suits steady usage. Spiky agent work will reach the ceiling early in a period, and requests will stop until the reset.
The two-model list is probably a starting point. Vercel says it plans to add featured models over time, so a team that builds around exactly these two should expect the list to change.
Wafer is a dependency to put in front of your legal or security reviewer, because its terms apply alongside Vercel's.
Whether Token Pass is cheaper than pay-as-you-go depends on your token mix and on a Token Pass per-token rate the pages we opened do not break out. We cannot give a break-even point from public information.
Who should choose which option
| Your situation | Start with | Why | Watch for |
|---|---|---|---|
| Steady use of Kimi K3 or GLM 5.3 in coding agents, on a Pro team | Token Pass | A fixed monthly ceiling you can budget | Early-access spots, no rollover, requests stop at zero |
| Several models or providers in one product | Regular AI Gateway key | Covers every model and can route across providers | Credits must stay funded; set budgets |
| Production traffic that cannot stop mid-month | Regular AI Gateway key | Token Pass has no fallback | Keep Token Pass for non-critical work only |
| One prepaid balance and no subscription | OpenRouter | Per-request billing from one balance | Purchase fee; model coverage not checked here |
| Enterprise procurement | Quote-based talks | Token Pass is not open to Enterprise teams | Confirm invoiced billing terms |
Who this is for: an engineering or operations lead at a small B2B software company that is already on a Pro plan, runs coding agents on Kimi K3 or GLM 5.3, and wants a hard monthly limit on that work more than flexibility across models.
Red flags: your production traffic must not stop when a monthly cap is reached; your team needs Enterprise billing or contract terms; or your usage spans many models beyond the two featured ones.
When NOT to use US Tech Automations
Skip a US Tech Automations design in three cases. If you only want a message when credit runs low, a reminder in the chat tool your team already uses is less to maintain. If one engineer checks the Token Pass tab once a week, a calendar reminder does the same job. And if the team cannot read credit or usage data from its account at all, the workflow has no trigger, so a manual check is the honest option.
Frequently asked questions
Is Token Pass a replacement for AI Gateway credits?
No. Token Pass is a separate balance that pays only for the featured models, and regular AI Gateway credits keep covering every other request, per the billing documentation.
What happens when the credit runs out mid-month?
Requests made with the Token Pass key fail until the next billing period. Vercel does not charge overage or top up automatically, so the team must switch to a regular key and keep it funded.
Can a deployed app use Token Pass in production?
Yes, with a Token Pass key stored as an environment variable, but production traffic that must stay up should not depend on it alone. Token Pass does not fall back to paid usage, so a regular key is the safer path for that traffic.
Does Token Pass keep prompts out of model training?
Vercel says Token Pass requests have zero data retention and are not used for training, and the launch post says zero data retention is on by default. Confirm those terms with your own legal or security reviewer before you send regulated data through it.
Who can subscribe, and can an Enterprise team use it?
Pro teams can subscribe, Hobby teams must upgrade to Pro first, and Enterprise teams cannot subscribe to Token Pass. Early access has limited spots, and a full tab shows a waitlist.
Should we switch this month or wait?
If your usage is steady and limited to the two featured models, run Token Pass on non-critical agents first to learn your real burn rate. If your usage is spread across models or must never stop, stay on regular AI Gateway usage and revisit when the model list is wider.
What to do this month
A short decision checklist keeps the choice concrete:
Pull the last full month of usage by model, and confirm how much of it is Kimi K3 or GLM 5.3.
Confirm the team is on a Pro plan, or decide whether the upgrade is worth it for this workload.
Check that agent requests use only the two featured model IDs, with fast mode off.
Name one owner and one threshold for remaining-credit alerts.
Keep a funded regular key ready for any production path.
Three common mistakes are worth avoiding. Do not assume unused credit carries into next month, because it expires. Do not put a Token Pass key on production paths with no fallback. And do not set a provider order that leaves out token-pass, because the request will fail. If model choice is your biggest driver of token spend, our comparison of token efficiency between two Codex models covers that question separately.
Token Pass is worth a look this month only if your team's Kimi K3 or GLM 5.3 use is steady and a hard stop is acceptable. For everything else, the regular pay-as-you-go setup remains the safer default. To see how the alert design above would be configured, see how US Tech Automations configures this for a proposed workflow.
Sources:
About the Author
Helping businesses leverage automation for operational efficiency.
