LLM API pricing comparison (July 2026)

Per-token basics. Both providers bill input (what you send) and output (what the model generates) separately, per token. Output carries a large multiple: exactly 5x input across the current Claude lineup, and roughly 6x input on most current GPT models (8x on GPT-5.3 Codex). Both APIs are stateless, so a multi-turn conversation re-sends its whole history each call and pays input rates on all of it, which is why input usually dominates conversational bills despite the cheaper rate.

Caching discounts. Both providers discount repeated prompt prefixes to about 0.1x the input price, but the mechanics differ:

• Anthropic caching is explicit. You place cache_control breakpoints, and you pay a write premium: 1.25x input for a 5-minute TTL or 2x for 1-hour. Reads then cost ~0.1x. The write premium means caching can lose money on traffic that does not re-read the prefix within the TTL.
• OpenAI publishes a flat "cached input" rate per model (0.1x input on the current lineup, except GPT-5.5 Pro, which offers no cached discount). Its pricing pages list no separate write charge.

Both hinge on an exact repeated prefix. And notice what caching does not do: the history still grows every turn. Caching makes the old tokens cheaper, it never makes them disappear. That gap, the ever-growing history, is exactly what PromptCrunch closes.

Batch discounts. OpenAI documents batch processing at about half the standard rates (GPT-5.6 Sol batch input runs $2.50 versus $5.00). Anthropic offers batch too. If your work can wait for asynchronous replies, batch is the cheapest tier either provider sells.

Promotional pricing. Watch the time-boxed rates. Sonnet 5's introductory pricing (flagged in the table above) runs out August 31, 2026, then jumps to $3.00 / $15.00. Size your budget on the rate you will be paying in September, not the promo.

1. Compress the history before it is sent. This is the biggest recurring win, and the one the providers do not hand you. PromptCrunch shrinks the old turns of a conversation on the way to Claude or GPT, so you stop re-buying the same context on every call. On long prose chats it is worth up to 75% off your input bill. Proof and setup are in the box below.

2. Cache your stable prefix. If a big system prompt or a fixed tool block repeats on every call, provider caching discounts those repeated tokens heavily. It is built in and it stacks cleanly with compression: caching handles the parts that stay the same, compression handles the history that grows. On Anthropic, mind the write premium (1.25x for the 5-minute TTL, 2x for 1-hour); it only pays off if the prefix gets re-read inside the window.

3. Right-size the model. The conversation table above spans $0.05 to $2.57 for identical traffic, so test the cheaper models on your real workload before you reach for the flagship. This stacks too: a leaner model running a compressed prompt is the cheapest way to serve the same chat.