Claude Desktop pay-as-you-go pricing on Vana API: $3.33 per million

§ One address, three protocols, a short setup

Vana API is built around a single idea: one address, three protocols, minimal setup. Every model request goes to the same api.vanaapi.com address, and the only thing you have to decide is which protocol your client already speaks. There is no separate host to memorise per model family, no second account to open, and no per-model configuration file to hand-edit. The address answers API requests and nothing else. Keys, balance and the usage log live in the dashboard on vanaapi.com, so the account side of the service stays in one place while the traffic side stays on one endpoint.

Here is the whole routing map. Claude models are served over the Anthropic Messages path, /v1/messages. Gemini models are served over the Gemini native path, /v1beta/models/... . GPT, Grok, DeepSeek, GLM, Kimi, Qwen, MiniMax and MiMo models are served over the OpenAI format, /v1/chat/completions, and GPT models also answer on /v1/responses. One endpoint, three request shapes, and the client you already use decides the shape.

Getting started is deliberately unglamorous. Create a key in the dashboard, copy it, paste it into the base address field of your client, and send a request. New accounts receive a $0.10 credit gift on sign-up, which is enough to confirm that the path works end to end before you commit any money. Because billing is prepaid and charged per real token, there is no monthly fee to cancel and no subscription to forget about. You top up a balance, you spend from it, and the usage log shows you every call that was billed.

§ Claude Desktop and the Anthropic Messages path

Claude models on Vana API are reachable only from official Claude Code or Claude Desktop, version 2.1.285 or later, because those applications speak the Anthropic Messages protocol themselves. They do not work from any other client, any SDK, cURL or a script. If your daily driver is Claude Desktop, that is exactly the kind of constraint that turns setup into a two-minute job: update the app, point it at the address, and the model list appears.

For anyone comparing spend, the per-token numbers are where the value shows up. Prices below are per million tokens, shown as input / output, with the official list price alongside for reference:

• Claude Fable 5.1 (claude-fable-5-1): Vana API $3.33 / $16.74, official list price $10 / $50
• Claude Haiku 4.5 (claude-haiku-4-5-20251001): Vana API $0.3472 / $1.60, official list price $1 / $5
• Claude Opus 5.5 (claude-opus-5-5): Vana API $1.46 / $6.94, official list price $4 / $20
• Claude Sonnet 5.5 (claude-sonnet-5-5): Vana API $1.18 / $5.76, official list price $2 / $10

Read those rows as a decision, not a boast. If your work is mostly short exchanges and quick edits, Haiku 4.5 at $0.3472 / $1.60 keeps the meter slow. If a session is long and reasoning-heavy, Sonnet 5.5 at $1.18 / $5.76 or Opus 5.5 at $1.46 / $6.94 is the deliberate choice, and the same pattern holds on the official list price, where the cheaper tiers are also the lower ones. Fable 5.1 at $3.33 / $16.74 sits at the top of the Claude range on both columns. Vana API does not change how the models behave; it changes what each token costs.

One practical detail matters for a quiet month: the client, not the account, determines access. Claude Desktop 2.1.285 or later talks to the Anthropic Messages path directly, so there is no adapter layer to configure and no compatibility setting to toggle. When a request fails, the place to look is the protocol, not the brand name.

§ Pay-as-you-go or a monthly plan: three criteria

The choice between topping up and buying a monthly plan is not about excitement. It is about predictability, cap size and which models you actually use. Vana API offers both, and the decision criteria are simple enough to apply in a minute.

• Claude up to $1,000 / 30 days: $64.72 for 30 days, at most $320 a week, Claude models only
• Claude up to $2,400 / 30 days: $153.61 for 30 days, at most $750 a week, Claude models only
• GPT up to $4,000 / 30 days: $103.61 for 30 days, at most $1,200 a week, GPT models only

Criterion one is whether your usage is steady enough to fill a weekly allowance. The $64.72 Claude plan carries at most $320 a week against a 30-day ceiling of $1,000, which suits a regular working rhythm on Claude models. If a month needs more room, the $153.61 plan raises the weekly ceiling to $750 and the 30-day ceiling to $2,400, still Claude models only. Criterion two is whether you want a ceiling at all. Pay-as-you-go has no monthly fee and no weekly cap; you pay for the tokens you actually send, which suits an irregular month or a month of experimentation.

Criterion three is model family. The GPT plan at $103.61 for 30 days, at most $1,200 a week, up to $4,000 / 30 days, covers GPT models only, so it does not mix with the Claude plans. If your work spans both families, prepaid pay-as-you-go is the neutral option, since one balance funds whichever protocol the request uses. A useful example: say you run a daily Claude Desktop habit and an occasional GPT workflow in the same month. Pay-as-you-go lets both draw from the same prepaid balance and appear in the same usage log, while a single-family plan asks you to pick one lane for that period.

§ Gemini and the OpenAI-format families

Gemini models are served only from Gemini CLI over the Gemini native protocol, /v1beta/models/... . The criterion here is simply the client you use. Gemini prices per million tokens, input / output, are $0.2083 / $1.60 for Gemini 2.5 Flash, $1.39 / $7.92 for Gemini 3.1 Pro, $1.04 / $4.65 for Gemini 3.8 Flash, and $1.46 / $6.88 for Gemini 4 Argon. Flash-tier models are the economical starting point; the Pro and Argon entries cost more in both columns and are chosen for output quality rather than for the routing.

The OpenAI-format group covers GPT, Grok, DeepSeek, GLM, Kimi, Qwen, MiniMax and MiMo over /v1/chat/completions, from Cursor, Cline, Roo Code, OpenCode, Cherry Studio, any OpenAI SDK or cURL. GPT models also answer on /v1/responses, which is what Codex uses, so a Codex workflow stays on the same address. Representative prices: GPT-6.1 Sol at $0.1389 / $0.8333, GPT-5.6 Sol at $0.3472 / $1.81, GPT-6 Astra at $0.6944 / $3.54, Grok 4.7 at $0.4861 / $1.53, DeepSeek V4 Flash at $0.2778 / $1.04, DeepSeek V4 Pro at $1.11 / $3.33, GLM-5.3 at $0.5556 / $1.67, Kimi K3 at $1.04 / $5.21, Qwen3.7 Plus at $0.1278 / $0.4167, Qwen3.8 Max at $0.6944 / $2.08, MiniMax M3 at $0.1389 / $0.5556, and MiMo V2.6 Pro at $0.2083 / $0.3472.

Because all of these share one address and one balance, the practical criterion becomes which model tier fits the task and its per-token price, rather than a per-provider setup cost. That is the quiet benefit of the single-endpoint design: adding a second model family costs a model id, not a new integration.

§ Money mechanics that decide the month

Billing on Vana API is prepaid, charged per real token, with no monthly fee. Every call is listed in the usage log in the dashboard, so the question of what you spent is answered by a record rather than an estimate. Keys are created in the dashboard, which keeps credentials and balance under the same roof as the log that explains your spending.

Top-ups are reversible within a narrow, clear window: unused balance can be refunded within 3 days of the top-up, minus what was used and the payment processor's fee. If you are deciding between a large prepaid balance and a smaller one, that 3-day rule is a real criterion, not a footnote. It gives a short window for correcting a mistaken top-up, and after it closes the balance simply behaves as prepaid credit.

There is also a referral mechanism. You earn 14% of what people you invite pay, and 5% one level further, with earnings withdrawable from $13.89. For a quiet month, that threshold is the number to keep in view: it is the point at which referral earnings can actually be taken out rather than merely accumulated.

§ A short decision guide

Start with the client. If it is Claude Desktop or Claude Code at version 2.1.285 or later, Claude models are a short configuration away on Vana API. If it is Gemini CLI, the Gemini native path is already the one it speaks. If it is one of the OpenAI-format clients, the /v1/chat/completions path is there, with /v1/responses available for GPT work including Codex.

Then start with the money. For irregular use, pay-as-you-go on a prepaid balance keeps the commitment small and the log complete. For steady Claude use, the $64.72 plan with its $320 weekly ceiling may be simpler than watching a balance; for heavier Claude months, the $153.61 plan raises the ceiling to $750 a week and $2,400 / 30 days. For steady GPT use, the $103.61 plan with a $1,200 weekly ceiling and $4,000 / 30 days applies to GPT models only.

Finally, pick the tier by cost per million tokens and by what the month actually needs. Haiku 4.5 at $0.3472 / $1.60, Qwen3.7 Plus at $0.1278 / $0.4167 and MiniMax M3 at $0.1389 / $0.5556 sit at the light end of the list; Opus 5.5 at $1.46 / $6.94 and Fable 5.1 at $3.33 / $16.74 sit at the heavy end. One address, three protocols, minimal setup, and a number attached to every token.