Picking a model API usually comes down to two questions: is it good enough, and what will it cost me? This page answers the second one, using each provider's own published prices as of Oct 22, 2026. Prompt caching alone drew about 800 posts in our X sample. People are clearly trying to cut their bills.
The price table
All prices are US dollars per million tokens, standard API, no batch discount.
| Model | Input | Output | Cached input (cache hit) | Notes |
|---|---|---|---|---|
| Claude Opus 5.5 | $4 | $20 | $0.20 | 5-min cache write $5 |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 | 5-min cache write $2.50 |
| Claude Haiku 5.5 (prompts ≤100K) | $0.10 | $0.50 | $0.01 | over 100K: $0.50 / $2.50 |
| Gemini 4 Argon (until Nov 30) | $2 | $10 | 95% off input | "introductory" |
| Gemini 4 Argon (after Nov 30) | $4 | $20 | — | per Google's footnote |
| GPT-6 Astra | $10 | $50 | separate rates | fast mode: 2x speed, 2x price |
| MiniMax M3.1 Flash | $0.15 | $0.60 | $0.015 | flat pricing |
| Mercury 3 (list) | $0.20 | $0.70 | 90% off input | 50% off at launch: $0.10 / $0.35 |
Sources: provider pages linked below. Model guides: Opus 5.5, Gemini 4 Argon, GPT-6, MiniMax M3.1 Flash, Mercury 3.
Not in the table: Google's post-November cached rate, OpenAI's exact cache rates and Inception's launch-discount end date weren't on the pages we checked, so we left them out rather than guess.
Three things that move your bill more than the list price
1. Output tokens
Every model above charges 3.5–5x more for output than input. A coding agent that writes whole files is mostly paying the output price. Compare that column first.
2. Caching
If you send the same big prefix again and again (system prompt, codebase map, docs), cache hits are the cheapest tokens you'll buy:
- Claude Opus 5.5 and Sonnet 5.5: cache hits are 0.05x base input. Most other Claude models use 0.1x. Writing to the cache costs extra: 1.25x base input for 5 minutes, 2x for 1 hour.
- Gemini 4 Argon: cached input 95% off at launch pricing.
- Mercury 3: cached input 90% off.
- MiniMax M3.1 Flash: $0.015, a tenth of its input price.
How to make caching actually hit: prompt caching explained.
3. Batch and long context
- Batch: Anthropic's Batch API is 50% off input and output, if you can wait for results.
- Long prompts: Claude models from 4.6 on bill the full 1M context at standard rates, except Haiku 5.5, which costs more above 100,000 prompt tokens. See context windows explained.
A worked example
Here's a small agent workload: 100 requests, each with a 15,000-token cached prefix, 5,000 fresh input tokens and 1,000 output tokens. That's 1.5M cached, 0.5M fresh input and 0.1M output tokens.
| Model | Cached (1.5M) | Fresh input (0.5M) | Output (0.1M) | Total |
|---|---|---|---|---|
| Claude Opus 5.5 | $0.30 | $2.00 | $2.00 | $4.30 |
| Claude Sonnet 5.5 | $0.15 | $1.00 | $1.00 | $2.15 |
| Claude Haiku 5.5 | $0.015 | $0.05 | $0.05 | ≈$0.12 |
| Gemini 4 Argon (intro) | $0.15 | $1.00 | $1.00 | $2.15 |
| MiniMax M3.1 Flash | $0.0225 | $0.075 | $0.06 | ≈$0.16 |
| Mercury 3 (list) | $0.03 | $0.10 | $0.07 | $0.20 |
| GPT-6 Astra (no cache discount applied) | — | $20.00 (all 2M input) | $5.00 | $25.00 |
Caveats:
- We left out cache-write charges (Claude charges them on the first request that writes the prefix).
- Astra's cache rates weren't published on the page we checked, so its row is pre-discount.
- Tokenizers differ. Anthropic says Claude 4.7 and later models produce about 30% more tokens for the same text than earlier ones. Other providers' tokenizers differ too, so the same prompt isn't 20,000 tokens everywhere.
- Per-task cost is what counts. A model that solves it in one try can beat a cheaper one that needs three.
How to choose
- Prototype on something cheap (Haiku 5.5, MiniMax M3.1 Flash or Mercury 3), and switch up only where quality fails.
- Cache your prefix before switching models to save money.
- Mark Nov 30 if you're building on Gemini 4 Argon: the per-token price doubles.
- Run your own five-task test and compare real bills, not list prices.
- Watch the providers' pages. Prices here were checked on Oct 22, 2026 and change often.
Running locally is another way to cut costs: see Ollama for beginners and what it costs to rent a GPU.
