Models & pricing

Model list, prices per 1M tokens, how a request is priced and what it holds.

Models

The same list with prices comes from GET https://api.egrtgghfghtytgb.space/v1/models, in a format both the Anthropic and the OpenAI SDK read.

Prices

Credits per 1M tokens. One credit is $1.

Prices are the provider's list prices plus 15%. There is no other fee.

• Haiku 5.5 has two tiers. A request's prompt is its input, cache read and cache write tokens together. Above 100,000 the whole request is priced at the second tier.
• Thinking tokens are output tokens.
• Minimum charge: 0.001 credit per request. It covers the two Solana transactions behind every request.
• Each request's cost is rounded up to the nearest 0.000001 credit.

Web search

11.50 credits per 1,000 searches. Search results join the prompt and are billed as input tokens.

Examples

These tables are computed from the same prices and code the API uses.

Sonnet 5.5: 2,000 input tokens, 500 output tokens

Opus 5.5: one Claude Code turn

3,000 new input tokens, 80,000 tokens read from the cache, 2,000 written to the 5-minute cache, 1,200 output tokens:

Haiku 5.5: a long prompt

120,000 input tokens and 1,000 output tokens. The prompt is over 100,000, so the whole request uses the second tier:

What a request holds

Before the model runs, a request holds its worst-case cost on your API balance:

• Input: text bytes divided by 3, at most the model's context window. It is priced at the 1-hour cache write rate if any cache_control has ttl: "1h", at the 5-minute rate if there is any other cache_control, and at the input rate otherwise. Requests with images or documents are counted exactly with the provider's token counter.
• Output: max_tokens at the output rate. Thinking is part of max_tokens.
• Web search: max_uses searches, or 5 if it is not set, each with 20,000 tokens of results.
• Haiku 5.5 holds at the second tier once the input estimate passes 70,000 tokens.

Sonnet 5.5 with 6,000 bytes of text and max_tokens: 1024:

When the request ends, the program burns its actual cost and releases the rest. You are never charged more than the hold: if the actual cost is higher, Unspent pays the difference.

One request can hold at most 25 credits. A request with a larger worst case gets 400 request_too_expensive: lower its max_tokens.