Vercel Token Pass at $150: What a Month of Kimi K3 Coding Costs

A month of Kimi K3 coding on Vercel’s Token Pass costs $30 flat, according to Vercel’s changelog, up to a $150 monthly token credit raised from $80 on November 13, 2026, with no change to the per-team price. The changelog states it directly: “Vercel Token Pass now includes $150 of tokens per month (up from $80) for the same $30 per team per month.” The credit applies to featured open-weight models served through AI Gateway, named in the same entry as Kimi K3, Kimi K3.1 Pro, GLM 5.3, GLM 5.3 Flash, and DeepSeek V4.1 Flash.

The harder question is what $150 of tokens actually buys, and that is where the evidence gets uneven. Vercel’s changelog does not state how the credit is denominated against gateway-billed tokens: provider list price, gateway rate, or something else. The cited documents publish per-token rates for only two of the five featured models, DeepSeek V4.1 Flash and GLM-5.3-Flash, and none for Kimi K3 itself. Every token count below is therefore arithmetic on public list rates with assumptions exposed, not a measured spend and not a Vercel guarantee. Benchmark figures are similarly mixed in provenance: the capability numbers are either vendor-reported (DeepSeek’s own results, Z.ai’s marketing) or independent (the CrossSWE preprint), while BenchLM’s aggregate mixes vendor-reported and independently run entries in one score; each is labeled as such where it appears.

What changed on November 13

Three things matter operationally, all from the single vendor changelog. First, the credit near-doubled while the price held at $30 per team per month. Second, the allowance does not accumulate: “Credit resets each billing period with no rollover.” A team that uses $40 of the $150 in January gets $150 again in February, not $260. Third, and most consequential, what happens at the cap depends on which key the agent uses: “When it runs out, requests using standard AI Gateway keys or OIDC tokens fall back to the team’s AI Gateway balance, while requests using Token Pass-scoped keys stop with no overage charges.”

There is also a supply constraint. Token Pass is, per the changelog, “an Early Access add-on for Pro teams, with limited spots and a waitlist when they fill.” Agents connect through vercel token-pass in the Vercel CLI or a Token Pass API key. Everything in this paragraph and the last is vendor-reported from one document published November 13, 2026; treat the terms as current but changeable.

What $150 buys: the reference table

The table converts $150 into token budgets at the published rates in the three cited pricing documents. Where no cited document publishes a per-token price, the cell says so rather than guessing.

The DeepSeek rows come from a review of DeepSeek V4.1 Flash 1017, which quotes first-party list pricing: “Off-peak API pricing is $0.15 per million cache-miss input tokens, $0.003 per million cache-hit input tokens and $0.60 per million output tokens. Peak rates are double.” Peak windows are narrow: the review defines them as 01:00-04:00 UTC and 06:00-10:00 UTC, Monday through Friday, excluding Chinese public holidays, with all other hours off-peak. A team running agents through the American workday is off-peak the entire time and should budget on the off-peak row; a European workday overlaps peak only in the early morning, roughly 08:00-10:00 UTC. The same review notes token prices did not change across the 1017 checkpoint.

The Kilo row is the warning. Kilo Code’s model page lists the identical model at $0.30 per million input and $1.20 per million output, double DeepSeek’s off-peak direct rates, with a $0.01 cache read price. The GLM row comes from a CometAPI quickstart for GLM-5.3-Flash, a reseller reference at $0.15 new input, $0.03 cached input, $0.50 output. That quickstart also documents a 50% launch promotion that ended September 9, 2026; anyone still quoting the promo rate is working from an expired discount.

Two assumptions the whole table hangs on

The first is the denomination question stated above: neither the changelog nor any other cited document says whether Vercel values the $150 at provider list prices or at its own gateway rates, which may differ. If the gateway bills the Kilo-style route at reseller rates, the same credit buys half the tokens of the direct-rate row. Before relying on any cell, verify what the gateway actually charges per model on your team’s plan.

The second gap is Kimi K3 pricing, and it matters because K3 anchors the headline. The Kimi K3 technical report establishes that the model is real and substantial, a “2.8T parameter Mixture-of-Experts model with 104 billion activated parameters, native vision capabilities, and a 1-million-token context window,” but it states no per-token API price for K3, for Kimi K3.1 Pro, or for the full GLM 5.3. Z.ai’s subscription page markets GLM-5.3 and GLM-5.3-Flash with a claimed “coding experience up 50% over the previous model” across a 1M context, but publishes plan details rather than per-token API rates, which is exactly why the GLM row above rests on a reseller. The honest statement of the headline math is: a month of Kimi K3 coding costs $30 until the credit runs out, and how many tokens that credit represents is a number Vercel’s changelog does not provide.

The key type is the real cost control

Most subscription-budget advice reduces to watching token consumption. Here that advice misses the mechanism. Vercel’s stated semantics move cost control from monitoring to a one-time configuration choice:

• Token Pass-scoped key: at the cap, requests stop. No overage charges. The agent fails loudly and the invoice does not move.
• Standard AI Gateway key or OIDC token: at the cap, requests silently fall back to the team’s paid AI Gateway balance. The agent keeps working and the invoice grows.

For a coding agent, which can burn a long-context model’s window in a single session, the silent fallback is the dangerous default. I would treat the scoped key as the only acceptable configuration for any agent that runs unattended, and reserve standard keys for interactive use where a human will notice the balance draining. That judgment follows directly from the vendor’s own description of the two behaviors; it is not a tested finding.

The no-rollover reset cuts the other way. There is no benefit to conserving credit late in a billing period, and no penalty for heavy use beyond hitting the stop. A scoped key plus a flat monthly cap is close to a fixed-price contract for open-weight inference, which is a genuinely simpler budgeting shape than per-token monitoring, provided the capability tier is acceptable.

Can these models carry a coding agent?

The strongest independent benchmark is the CrossSWE preprint, which benchmarks coding agents on cross-repository joint development using mini-SWE-agent v2.4.6 at high effort with a $10 per-run cap. Its headline, author-reported: “Even the strongest model, GPT-6.1 Sol, achieves only 58.48% Pass@1 on CrossSWE.” Of the featured models, DeepSeek V4.1 Flash reaches 52.63% Pass@1 and GLM-5.3 Flash 43.86%, with Claude Sonnet 5.5 at 50.29% between them.

The per-instance costs from the same harness are more decision-relevant than the ranks: $0.2881 for DeepSeek V4.1 Flash, $0.3411 for GLM-5.3 Flash, and $0.3496 for GPT-6.1 Sol. At those measured rates, $150 covers roughly 520 DeepSeek-style instances or 440 GLM-style instances per month, a concrete sense of scale for agentic workloads similar to that harness. The catch, also author-reported in the same paper, is trajectory length: GPT-6.1 Sol and Claude Sonnet 5.5 average 32.9 and 29.2 steps per task, compared with 152.7 for DeepSeek V4.1 Flash and 141.7 for GLM-5.3 Flash. The open-weight models get comparable per-task costs by being cheap per token, not by being efficient per task, and four to five times the steps means proportionally more wall-clock time and more chances to hit a per-run cap.

Vendor-reported numbers paint a friendlier picture and should be weighted accordingly. DeepSeek’s own DeepSWE v1.1 comparison, cited in the BuildFastWithAI review, puts Flash 1017 at 77.0 against 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol; CrossSWE’s authors explicitly note GPT-6.1 Sol scores 58.48% on their harness versus a reported 75.2% on DeepSWE v1.1, which is a clean illustration of why harness choice moves numbers. BenchLM’s aggregate scores DeepSeek V4.1 Flash 1017 at 64.48/100, ranked 31 of 1033 models as of November 13, 2026, while flagging that partial benchmark coverage makes the overall score conservative. That aggregate mixes provenance in one ranking: its own tables mark most coding entries, including the 77.0 DeepSWE score, “Reported by DeepSeek,” while a minority, such as AA-SciCode and the Vals AI Terminal-Bench 2.1 run, are independent.

For Kimi K3, the technical report summarizes third-party evaluations as of July 23, 2026: Artificial Analysis Intelligence Index v4.1 of 57.1, fourth of 580 models, behind Claude Fable 5 (59.9) and GPT-5.6 Sol (58.9), and first of 99 on WebDev Arena at 1,678 Elo versus Claude Fable 5’s 1,634. The same report concedes the ceiling on coding: “On Kimi Code Bench 2.0, 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.” Those are vendor-reported figures from the model’s own report. The consistent pattern across independent and vendor evidence is the same: the featured open weights are competitive per dollar and per task, and they trail the strongest frontier closed models on absolute coding capability. The gap is not uniform down the frontier, though: on CrossSWE, DeepSeek V4.1 Flash’s 52.63% sits above Claude Sonnet 5.5’s 50.29%, so the open weights sit below the frontier leaders, not below every closed model.

How fast the credit burns

Two levers decide whether $150 lasts a week or a month. The first is reasoning effort. The DeepSeek V4.1 Flash paper documents three public API presets: “low, high, and max correspond to effort values of 50, 75, and 100, respectively,” and reports that mean output tokens per trajectory grow monotonically with effort across every scaffold-benchmark pair tested. Output tokens are the expensive direction in every published rate here ($0.50–$1.20 per million versus $0.15–$0.30 input), so the effort preset is the single biggest dial an agent operator controls. Running routine tasks at low effort and reserving max for genuinely hard problems is not a nicety; at the published ratios it can plausibly double how long the credit lasts.

The second lever is context length. Kimi K3 advertises a 1M-token window and Kilo’s page lists 1,048,576 tokens for DeepSeek V4.1 Flash. An agent that stuffs a repository into every request burns the input budget far faster than per-task averages suggest, and an independent study of million-token attention behavior adds a capability caveat, describing trust in a one-million-token window as “trusting a claim with no independent check behind it.” That paper’s own diagnostic pilot ran at 768 tokens on small models and does not extrapolate to frontier scale, so read it as a prompt to verify long-context retrieval on your own codebase rather than as a verdict.

The break-even: $30 against per-token spend and self-hosting

Against per-token API plans, the break-even is simple arithmetic. If your routed open-weight usage would cost more than $30 per team per month at the rates in the table, and your workload fits within the $150 credit, the pass wins on price; the only question is whether the capability tier above clears your bar. At DeepSeek off-peak input rates, $30 of credit-equivalent spend is about 200M input tokens a month. That is little headroom for a working agent: CrossSWE measured an average of 17.54M input tokens per instance for DeepSeek V4.1 Flash, enough to exhaust 200M in about eleven runs. Below that threshold, pay-as-you-go wins because the unused credit does not roll over.

Against self-hosting, the bar is concrete. A validated vLLM and LMCache study for GLM-5.3-Flash shows a featured model running locally (it served the RedHatAI GLM-5.3-Flash-NVFP4 quantized checkpoint) with preserved output equality across 120 requests, with CPU reload cutting time to first token by 46–64% in the measured trials. So local serving of at least one featured weight is a real option, and $30 a month is the number self-hosting must beat after GPU rental or hardware amortization, operator time, and the engineering cost of keeping a serving stack healthy. For a single team without existing GPU capacity, that is a hard number to undercut. For an organization already running inference infrastructure, the calculus reverses, and the pass’s Early Access limits and vendor-controlled terms count against it.

One more lineage note for the Kimi-curious: the Kimi-Dev paper reports 60.4% on SWE-bench Verified, best among workflow approaches in its comparison set, and 48.6% pass@1 adapted as an SWE-Agent, described by its authors as on par with Claude 3.5 Sonnet (241022 version). Those are author-reported results for a predecessor agent line, useful context for Moonshot’s coding pedigree, not evidence about K3’s agentic performance on your tasks.

Verdict

For Vercel Pro teams already routing a coding agent through AI Gateway, Token Pass at $30 a month is a defensible capped budget on DeepSeek V4.1 Flash or GLM-5.3-Flash, where $150 equates to roughly 250M to 1B tokens at published list rates, and roughly 440 to 520 CrossSWE-style agent tasks at independently measured per-instance costs. Make the Token Pass-scoped key a condition of adoption, because the standard-key fallback converts the cap into a meter you were trying to avoid.

Wait, and re-verify first, if you need Kimi K3-specific dollar math (neither Vercel’s changelog nor the Kimi K3 technical report states a per-token price for Kimi K3, and the changelog does not state how it denominates the credit), if you need a guaranteed seat (spots are limited and waitlisted), or if your workload needs frontier-level cross-repository coding, where the independent CrossSWE benchmark puts both featured Flash models behind GPT-6.1 Sol. All Token Pass terms cited here come from one vendor changelog published November 13, 2026, and every number in the table is a public-data calculation, not a measured spend.

Frequently Asked Questions

What happens to requests when the Token Pass credit runs out?

When it runs out, requests using standard AI Gateway keys or OIDC tokens fall back to the team’s AI Gateway balance, while requests using Token Pass-scoped keys stop with no overage charges.

Is Token Pass available to all Vercel Pro teams?

Token Pass is, per the changelog, “an Early Access add-on for Pro teams, with limited spots and a waitlist when they fill.”