Every tier uses the same OpenAI-compatible API. You just pay 1×, 0.8× or 0.6× of the provider's list price as your monthly volume grows.
30%off provider list price
You pay 0.70× of what Anthropic / OpenAI / Google charge
Unlocks at $0 / mo
Drop-in replacement. Every model, every cache tier, zero commitment.
44%off provider list price
You pay 0.56× of what Anthropic / OpenAI / Google charge
Unlocks at $500 / mo
20% off everything once you cross $500 / mo. Auto-applied.
58%off provider list price
You pay 0.42× of what Anthropic / OpenAI / Google charge
Unlocks at $1,000 / mo
40% off plus invoices and dedicated concurrency once you cross $1k / mo.
| 1.0× | 0.8× | 0.6× | |
|---|---|---|---|
| Volume rebate | — | ✓ | ✓ |
| PDF receipts | ✓ | ✓ | ✓ |
| Auto-retry on 5xx | ✓ | ✓ | ✓ |
| Recommended concurrency | 32 | 128 | 512 |
| Requests / min | 120 | 600 | up to 2,000 |
| Target uptime | 99.9% | 99.9% | 99.9% |
| Activation threshold | $0 | $500 | $1,000 |
Tiers are evaluated on a rolling 30-day window. Top-ups never expire.
Tier is recomputed on every request from your rolling 30-day spend. Discount applied automatically — no promo codes, no invoices to edit.
Long system prompts, big documents, reused tool definitions — price them once, reuse them cheap.
| Model | Base in | 5m write | 1h write | Read | Savings vs base |
|---|
Included—the provider has no explicit cache-write fee — writes are covered by the base input price. Only Claude exposes separate 5-min / 1-hour write tiers; Gemini and GPT cache reads are free to write by design.
All prices USD per 1M tokens · Alloneia's final rate, official already discounted.
Mark a message block with cache_control and pick a TTL. The next request that starts with the same prefix gets charged at the cache read rate — typically 10% of the base input price.
"messages": [
{
"role": "system",
"content": [{
"type": "text",
"text": "You are a senior dev reviewer…",
"cache_control": { "type": "ephemeral", "ttl": "5m" }
}]
},
{ "role": "user", "content": "Review this PR" }
]OpenAI's cache is keyed on the prefix plus an optional prompt_cache_key. Same prefix + same key = cache hit at the discounted input rate.
{
"model": "gpt-5.4",
"messages": [...],
"prompt_cache_key": "assistant-v3"
}No cache-enabled models configured yet.
No. Same /v1/chat/completions. Add cache_control (Anthropic) or prompt_cache_key (OpenAI) to your existing request body.
The TTL of the cached entry. 5-minute cache is cheaper to write, perfect for chat sessions. 1-hour cache costs more to write but is worth it for long-lived prompts reused across many requests.
Anthropic supports up to 4 cache breakpoints per request, any contiguous prefix. OpenAI caches automatically on prefix matches of ≥1024 tokens.
No. Claude models expose full 5m + 1h writes; GPT-5 and Gemini 2.5 support cached reads only. The table above shows what's available per model.
No. We pass through the provider's cache pricing — same 0.7× discount applies to both base and cache tiers.
You get billed at the base input price — exactly what you'd pay without caching. No penalty, no surprise.