Prepaid LLM API Pricing: Per-Token Rates Compared (2026)

Prepaid LLM API pricing works the same way at every major provider: you buy credits up front, each request deducts its tokens at a per-million-token rate, and the API stops when the balance hits zero. You pay as you go, with no subscription. OpenAI, Anthropic, Google's Gemini API, DeepSeek, OpenRouter and SolCard Inference all sell API usage this way. What differs is the per-token price, the minimum top-up, any fee on the top-up, and whether unused credits expire.

This guide puts those numbers side by side, then prices one monthly workload on each provider so you can see what a bill actually looks like. All prices are in USD per 1 million tokens, taken from each provider's own pricing page on October 31, 2026.

A prepaid API account holds a dollar balance. Every request is metered in tokens -- input tokens you send, output tokens the model writes -- and the cost comes off the balance.

Google's billing docs describe it plainly: "your API usage costs are deducted from your Prepay credit balance in near real-time." DeepSeek's pricing page says fees "will be directly deducted from your topped-up balance or granted balance."

When the balance runs out, requests fail. Anthropic's billing article says: "If you run out of credits, you can no longer call the API or use the playground until you add more." Google returns "an HTTP 402 Payment Required error until you add credits."

That hard stop is the main reason developers pick prepaid billing. A runaway script can only spend what is already in the account.

Postpaid billing works the other way: usage accrues and you are invoiced later. Google offers both. Under Postpay, "your Cloud Billing account accrues costs" and "you are automatically charged at the end of the month," per the same billing docs. Claude through AWS Marketplace is postpaid too: Anthropic's pricing page lists its payment model as "Arrears only (postpaid); no prepaid credits."

For a solo developer or a small team, prepaid is the usual default. It needs no credit approval and caps your exposure.

A few details behind the table:

• OpenAI's help article sets the minimum purchase at $5 and says purchased credits expire after one year.
• Anthropic says its credits "expire one year from the purchase date, and the expiration date can't be extended," and that credit purchases are non-refundable, per its billing article. Its pricing FAQ says "Major credit cards are accepted for standard accounts."
• Google says "purchased credits expire after 1 year" and that refunds aren't allowed for Prepay accounts, except when switching account types.
• OpenRouter says its "fee is charged when you buy credits, 5.5% on Standard and 8% on Business, never on individual requests," and that "inference is billed at the provider's list price on every plan," per its pricing page. Its FAQ adds that "cryptocurrency payments are never refundable."
• SolCard Inference uses one prepaid USD balance for chat, every key and every model, with no card on file.

These are standard (non-batch) rates for a cross-section of current models, from cheap and fast to more capable.

Four pricing rules sit behind those rows:

• DeepSeek has off-peak rates. Its pricing page says "off-peak rates are half of the peak rates." Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays, excluding Chinese public holidays.
• Gemini 4 Argon is getting more expensive. Google lists $2 input and $10 output through November 30, 2026, rising to $4 and $20 on December 1, 2026.
• Anthropic charges for writing to the cache. A 5-minute cache write on Sonnet 5.5 costs $2.50 per million tokens, above the $2.00 base input rate. Cache reads are what's cheap.
• Batch is about half price. Anthropic's Batch API gives "a 50 percent discount on both input and output tokens," and OpenAI's Batch and Flex rates are about half of Standard. If a job can wait, batch it.

A low rate is only cheap if the model does the job. A small model that needs three retries costs more than a larger one that gets it right the first time. Token counts also differ by tokenizer: Anthropic says its newer tokenizer, used by Claude 4.7 and later, "produces approximately 30% more tokens for the same text." Test your own prompts before you compare bills.

Here is one workload priced on every model above. Say an agent sends 20 million input tokens a month, 15 million of which hit the cache (the repeated system prompt, tools and files), and writes 4 million output tokens.

The formula is: 5M x input rate + 15M x cached rate + 4M x output rate.

These totals leave out Anthropic's cache-write surcharge and Gemini's cache storage fees, so real bills on those two run a little higher. Through OpenRouter, the same model costs its list price plus the 5.5% fee on the credits you buy: $25.75 of gpt-6.1-terra usage takes about $27.17 of credits.

Two things stand out. Output is most of the bill on almost every row, so capping response length saves more than trimming prompts. And the cached share matters: without caching, the same 20 million input tokens on Sonnet 5.5 would cost $40 instead of $11.50.

SolCard Inference is not the lowest-priced option in this table. Small models from OpenAI, Anthropic and DeepSeek cost less per token. Where it fits is covered below.

Several providers accept requests in OpenAI's API format, so switching means changing a base URL, a key and a model name:

DeepSeek says its API "uses an API format compatible with OpenAI/Anthropic." OpenRouter says you can use "the OpenAI SDK pointed at OpenRouter as a drop-in replacement." SolCard Inference accepts OpenAI Chat Completions, OpenAI Responses and Anthropic Messages requests; its keys start with sk-si-.

Of the models in our table from these four providers, DeepSeek's have the lowest per-token rates. Payment is a different question: of these four, only OpenRouter and SolCard Inference say they take crypto.

Every provider above sells pay-as-you-go credits with no subscription required. Some also sell plans. SolCard Inference's pricing page lists three, paid from the same balance:

Plan limits reset every 5 hours and every week, and unused plan usage doesn't roll over. API keys spend credits first, then the plan.

A plan only beats pay as you go if you use most of it. If your usage is spiky or small, stay on prepaid credits: on SolCard Inference, the pay-as-you-go balance never expires, so a slow month costs nothing.

The per-token rate is not the whole price. Check these before you top up:

1. Expiry. OpenAI, Anthropic and Google credits expire after a year, and OpenRouter reserves the right to expire them. Buying a year's worth up front can mean losing some.
2. Top-up fees. OpenRouter adds 5.5% on Standard. If you fund a card-only provider with a crypto card, the card's own top-up fee applies too.
3. No refunds. Anthropic and Google say purchased credits are non-refundable, and OpenRouter says crypto payments are never refundable.
4. Auto-reload declines. Auto-reload charges your card when the balance drops below a threshold. If the card is declined, your app stops at zero. See why crypto cards get declined.
5. Minimums. OpenAI and Google set a $5 minimum. Small, but it matters when you just want to test.

There are two routes, and our full guide to paying for an LLM API with crypto walks through both.

Card-only providers (OpenAI, Anthropic). Load a crypto-funded card and buy credits with it. For these checkouts, use Virtual for AI & Ad Spend, SolCard's card built for AI and ad merchants: it costs $10 to issue and 5% per top-up. SolCard's standard Virtual and Platinum cards struggle at both checkouts in our data. The Anthropic API guide covers Console billing step by step.

Providers that take crypto directly. OpenRouter takes USDC. SolCard Inference takes SOL, USDC, ETH, BTC and more, on 9+ networks, from the dashboard's Top up page, or straight from a SolCard card balance. You sign up with an email or a Solana wallet, and new email sign-ups may get $1 of free credit for 30 days (abuse checks apply; wallet sign-ups get none).

If you code with Claude Code or Codex, the install page adds SolCard Inference's models next to your existing ones. Our guide to adding custom models to Claude Code and Codex covers the manual setup.

• Low per-token cost for simple tasks: Claude Haiku 5.5, gpt-6-luna or DeepSeek's deepseek-flash off-peak, all under $7 for the sample month above. OpenAI and Anthropic bill by card, and DeepSeek's pricing page doesn't list payment methods.
• Strongest models for hard tasks: Claude Opus 5.5 or Sonnet 5.5, or gpt-6.1-sol. Budget for output tokens, and use batch where you can.
• Many models under one key: OpenRouter, at list price plus 5.5% on credits, with USDC accepted.
• Pay in the coin you already hold, with no card and no expiry: SolCard Inference, at $1.25 input and $2.50 output per million tokens on its default model.

Plenty of teams mix them: a card-funded account for the one model they can't replace, and a crypto-funded API for everything else.

• Text only. Its models take no image, audio or file input, so keep a multimodal provider for screenshots and PDFs.
• Output is capped at 32,768 tokens per request.
• Its models are SolCard's own. You call solcard-inference-fast, -base or -large, not a Claude or GPT model. Test your prompts first.
• Small models elsewhere cost less per token, as the monthly table shows.

For SolCard's card fees and tiers, see SolCard fees.

It is pay-as-you-go billing where you buy credits in advance and each request deducts its token cost from the balance. When the balance reaches zero, the API stops working until you top up. OpenAI, Anthropic, Google, DeepSeek, OpenRouter and SolCard Inference all offer it.

Yes. All the major LLM APIs sell prepaid credits you draw down per token, with no monthly plan required. Plans, where offered, are optional.

You pay separately for input tokens, cached input tokens and output tokens, each at a rate per million tokens. Multiply each count by its rate and add them up: 5 million input tokens at $1.00 per million is $5.00.

It depends on the model you need. In our table, DeepSeek's deepseek-flash at off-peak rates came out at about $3.20 for the sample month. Older small models can cost less: Google lists gemini-2.5-flash-lite at $0.10 input and $0.40 output per million tokens on its pricing page. Prices change often, so check each provider's pricing page before you commit.

At most providers, yes. OpenAI, Anthropic and Google credits expire after a year, and OpenRouter reserves the right to expire them after a year. SolCard Inference says its pay-as-you-go balance never expires.

Requests fail until you add credits. Google returns an HTTP 402 error, and Anthropic says you can no longer call the API or use the playground. Auto-reload avoids this, as long as the card on file has funds.

Yes. OpenRouter accepts USDC, and SolCard Inference accepts SOL, USDC, ETH, BTC and more. For card-only providers such as OpenAI and Anthropic, load a crypto-funded card and buy credits with it.

Prepaid LLM API pricing is simple in structure and wide in range: the same month of work costs $2.65 on a small model and $103 on a frontier one. Pick the smallest model that does the job, cache what repeats, cap output, and buy only the credits you'll use before they expire. If you'd rather top up from a wallet than a card, SolCard Inference and OpenRouter both take crypto. New to wallet-funded cards? Start with what a crypto debit card is, or see what SolCard offers.

• Pricing, Claude Platform docs
• How do I pay for my Claude API usage?, Anthropic Help Center
• API pricing, OpenAI
• Setting up and managing prepaid API billing, OpenAI Help Center
• Pricing, Billing and OpenAI compatibility, Gemini API docs
• Models & Pricing and Your First API Call, DeepSeek API docs
• Pricing, FAQ and Quickstart, OpenRouter
• solcardinference.ai, Quickstart, Pricing, SolCard Inference