Vercel Killed Its Flat $0.30 AI-Agent Fee — Here's the Exact Token Count Where the New Metered Rate Costs More

On June 30, 2026, Vercel quietly repriced its AI Agent product: the flat $0.30 fee charged for every Code Review or investigation is gone, replaced by a $0.25-per-million-token "Vercel Token Rate" layered on top of the same zero-markup provider pass-through that was already there. Existing code-review users keep the old flat fee for a 30-day grace window that closes around July 30, 2026 — about three weeks from today. Run the two pricing models against the same workload and the crossover is exact: 1.2 million tokens. Below that, the new metered rate is dramatically cheaper than the old flat fee; above it, the meter has no ceiling and just keeps climbing. For a realistic mix of mostly-light PR reviews with an occasional deep investigation, the new pricing comes out cheaper overall — but the part of the bill that scales with usage now scales without limit, which the flat fee never did.

What Actually Changed, and When​

Vercel's own changelog spells out the mechanics plainly. The old model charged a flat $0.30 for every Code Review or additional investigation, on top of provider token costs passed through at the underlying rate with no markup. The new model drops the flat fee and the pre-loaded credit wallet it required, replacing both with a per-token rate — $0.25 per million tokens, applied to input, output, and cached tokens — charged in addition to the same zero-markup provider pass-through. Vercel's framing is that cost should scale with the intensity of the work: "a quick question costs less than a deep investigation that reads your logs, deployments, configuration, and runtime data, spins up sandboxes, and writes across projects."

The rollout is split by cohort. New Vercel Agent users were on the token-rate model starting June 30, 2026. Teams already using Agent for code reviews keep the old $0.30-per-request pricing for 30 days, then auto-migrate to the new model with no action required — landing the full transition around July 30, 2026.

This repricing sits inside a wider billing surface that hasn't changed. Vercel's own pricing page still meters the Pro plan across eight core usage-based dimensions: Fast Data Transfer (from $0.15/GB), Edge Requests (from $2/million after 10M included), Function Invocations (from $0.60/million), Active CPU (from $0.128/hour), Provisioned Memory (from $0.0106/GB-hour), Image Transformations (from $0.05/1K), Developer Seats ($20/month each), and Build Minutes ($0.014–$0.105/minute by machine tier). AI Gateway usage — the zero-markup token pass-through for custom AI features that aren't the packaged Agent product — draws from its own credit balance, separate from the Pro plan's included usage credit. Agent's token-rate billing, post-repricing, no longer needs a separate wallet, but it's still a distinct rate applied on top of Gateway's already-separate pass-through. A team building an AI-agent feature on Vercel is touching the shared eight-dimension infrastructure meter, the AI Gateway's own credit pool, and now Agent's token-rate charges — three different billing surfaces stacked on one feature, even after the simplification.

For comparison, Netlify runs five usage meters — bandwidth, compute, web requests, AI inference, and production deploys — that all draw from one shared credit currency (3,000 credits for $20/month on Netlify's Pro plan, at 20 credits/GB bandwidth, 10 credits/GB-hour compute, 2 credits per 10K requests). AI inference credits convert at a fixed 180 credits per $1 of underlying model cost, the same provider list rate any other pass-through billing uses. Netlify's own docs let a team set a credit usage cap that pauses agent runs outright once hit, rather than stopping the meter per line item as Vercel's separate balances do.

Worked Example: A Nightly PR-Review Agent, Old Pricing vs. New​

Take a team running Vercel Agent's Code Review feature on every merged pull request — 600 reviews a month, a reasonable pace for a mid-sized engineering org. Hold the underlying model constant at Claude Sonnet 4.5 rates ($3/million input tokens, $15/million output tokens) so the only variable being tested is Vercel's own markup, not the provider's.

A typical, light review reads a moderate diff plus surrounding repo context: call it 50,000 input tokens and 2,000 output tokens, 52,000 tokens total.

• Provider token cost: (50,000/1,000,000 × $3) + (2,000/1,000,000 × $15) = $0.15 + $0.03 = $0.18
• Old pricing (flat fee + tokens): $0.30 + $0.18 = $0.48 per review
• New pricing (token rate + tokens): (52,000/1,000,000 × $0.25) + $0.18 = $0.013 + $0.18 = $0.193 per review

At 600 reviews a month, that's $288/month under the old flat fee versus $115.80/month under the new token rate — a 60% drop in the total bill, and a 95.7% drop in Vercel's own markup specifically ($180 in flat fees down to $7.80 in token-rate charges). For a team whose reviews are mostly light, the repricing is a straightforward win.

A deep investigation — the kind Vercel's own changelog describes, reading extensive logs, deployments, configuration, and runtime data, and spinning up a sandbox — burns far more context. A conservative estimate: 2.8 million input tokens and 200,000 output tokens, 3 million tokens total.

• Provider token cost: (2,800,000/1,000,000 × $3) + (200,000/1,000,000 × $15) = $8.40 + $3.00 = $11.40
• Old pricing: $0.30 + $11.40 = $11.70
• New pricing: (3,000,000/1,000,000 × $0.25) + $11.40 = $0.75 + $11.40 = $12.15

At this depth, the new model costs more — $12.15 versus $11.70, about 3.8% higher in total, even though the markup component alone is 2.5x higher ($0.75 versus $0.30). The reason the total barely moves is that provider tokens dominate the bill at real investigation depth; the crossover is real, but its dollar impact stays modest as long as token volume stays in the low millions. It stops staying modest once volume keeps climbing: at 10 million tokens for an exceptionally deep, multi-repo investigation, the token-rate markup alone reaches $2.50 against a flat fee that would have stayed $0.30 no matter how much context the investigation read. That's the actual shape of "no ceiling" — not a cliff at 1.2 million tokens, but a markup that keeps scaling with usage in a way the flat fee structurally couldn't.

A blended, realistic month. Most reviews are light; a minority are deep. Model 540 light reviews and 60 deep investigations out of the same 600/month:

Even with 10% of reviews running deep enough to individually cost more under the new pricing, the blended total is still cheaper under the new model — $833.22 versus $961.20, a 13.3% reduction — because the light reviews that make up 90% of volume get cut by more than 95% on their markup component. The new pricing is a net win for this mix. It just isn't a win on every individual review, and the reviews where it costs more are exactly the ones with no upper bound on how much more.

What a Chat-Driven Admin Panel Pays (and Why the Repricing Doesn't Touch It)​

Not every AI feature runs through the Agent product. A chat-driven internal admin panel — built as a Vercel Function that calls a model directly through AI Gateway — never touched the $0.30 flat fee or the new $0.25/million token rate, because neither applies outside Agent's packaged Code Review and investigation features. It only pays AI Gateway's pass-through and the Fluid Compute meter for the function itself.

Take a moderate-usage internal tool: 5,000 chat completions a month, each averaging 3,000 input tokens (system prompt plus retrieved context) and 500 output tokens, routed through AI Gateway to GPT-4o ($2.50/million input, $10/million output — zero markup, per Vercel's own Gateway pricing).

• Per-call token cost: (3,000/1,000,000 × $2.50) + (500/1,000,000 × $10) = $0.0075 + $0.005 = $0.0125
• 5,000 calls/month: $62.50

The Fluid Compute cost of the orchestrating function is close to a rounding error by comparison. A 512MB function handling each request in 3 seconds of wall-clock time (mostly waiting on the streamed model response) with 200ms of genuine active CPU, billed at Washington D.C. rates ($0.128/CPU-hour, $0.0106/GB-hour memory, $0.60/million invocations):

• Active CPU: (0.2s/3,600) × $0.128 = $0.0000071/call
• Provisioned Memory: 0.5GB × (3s/3,600) × $0.0106 = $0.0000044/call
• Invocations: 5,000/1,000,000 × $0.60 = $0.003 total
• Compute total: ($0.0000115 × 5,000) + $0.003 = $0.06/month

Total for the chat admin panel: $62.56/month, essentially all of it AI Gateway token pass-through. The takeaway isn't that this workload is cheap — it's that it was never subject to Agent's repricing at all, because "eight metered dimensions plus Agent's token rate" only applies to features actually running through Vercel's packaged Agent product. A hand-rolled feature calling AI Gateway directly sidesteps that specific markup layer by construction, while still paying every other dimension a Vercel Function touches.

What Both Numbers Skip: Owning the Inference Route​

Provider token pricing is the same wherever the request originates — Anthropic and OpenAI don't charge less because the call came from a script on owned hardware instead of through Vercel's or Netlify's pass-through. What changes is whether anything gets added on top of that list price. Running the same two workloads from infrastructure you own removes Agent's markup entirely, in either its old or new shape, along with the Fluid Compute meter — replaced by a flat monthly server that doesn't reprice based on how many milliseconds were active or how many tokens an investigation read.

Using the blended, realistic month from above (540 light + 60 deep reviews, plus the chat admin panel) and a Hetzner CX22 (2 vCPU, 4GB, $4.59/month post the June 2026 Hetzner repricing) to run both the review script and the admin panel's backend:

The gap between "own the inference route" and "pay Vercel's markup" shrank sharply with the repricing — from $175.47/month under the old flat fee to $47.49/month under the new token rate, a 73% reduction in what the packaging layer costs on top of identical provider tokens. That's a genuine improvement, not a rounding trick: Vercel's own logic (charge for the intensity of work, not a flat toll) closes most of the distance to a fleet that never charged a packaging fee on agent actions in the first place. What it doesn't close is the shape of the remaining gap — the new $47.49/month is proportional to token volume, so it re-opens as usage grows in a way the old flat-fee gap never did on a per-review basis, and never closes to zero on infrastructure someone else still owns.

The Actual Lesson of a Repricing, Not Just a Price Cut​

Vercel's move from a flat per-action fee to a metered token rate is a real, defensible improvement for the common case — light reviews get cut by more than 95% on the markup that used to cost a flat $0.30 regardless of how little work the request actually required. It's also, honestly, a repricing that makes Vercel's own bill harder to predict in advance: a flat fee tells a team exactly what 600 reviews will cost before running any of them; a token rate means the bill depends on how much context each investigation happens to read, discovered only after the fact in the dashboard.

Both numbers — the $0.30 that's disappearing and the $0.25/million that's replacing it — describe the same thing: a fee for not building agent tooling yourself. A team running its own PR-review agent and admin-panel backend against a model provider's API directly, from a Cluster API fleet on Hetzner nodes it already owns, was never charged either version of that fee. The tokens cost the same either way, because pass-through pricing means nobody marks those up. The only thing a self-hosted platform skips is the toll on top — whichever shape that toll currently takes.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own, with AI agents treated as first-class operators rather than a metered line item on someone else's invoice. Star the repo on GitHub or deploy your first agent-facing service today.

Sources:

• Vercel Agent has updated pricing — Vercel changelog, June 30, 2026.
• Vercel Agent Pricing — Vercel docs.
• Pricing on Vercel — Vercel docs.
• Vercel Pricing: Hobby, Pro, and Enterprise plans — Vercel.
• Fluid compute pricing — Vercel docs.
• Pricing for AI features — Netlify docs.
• Credit-based pricing plans — Netlify docs.
• Agent Runners overview — Netlify docs.