Models and pricing.
Choose a model by its capabilities and compare prices for the same number of tokens.
Found: 1 of 32
Reset filtersApproximate tokens per $1 of balance based on observed usage. Results depend on the mix of input, output and cache.
Model names are pinned. Scroll the price columns to the right →
Context and cache prices: expand the model with +.
| Model / Model ID | Context | InputInput tokens | OutputOutput tokens | Cached inputReading cache | Cache writeWriting cache |
|---|---|---|---|---|---|
GPT-6.1 TerraOpenAI ≈ 114 million tokens per $1gpt-6.1-terra | 1,050,000tokens | $0.06$1 | $0.3$5 | $0.006$0.05 | $0$1.25 |
GPT-6.1 Terra · OpenAI A fast model for visual reasoning, computer use, and writing. The row shows official Standard prices up to 272000 input tokens. Above 272000: input and cache rates ×2, output ×1.5 for the full request. Official cache writes per 1M tokens: $1.25 for 30 minutes of retention, $2.50 for 12 hours. These conditions do not change Etage prices. Model context: 1,050,000 tokens
Official* — comparison from the original catalog entries, without independent verification. “No data” does not mean free usage or support for an operation. | |||||
How request costs are calculated
How request costs are calculated
You pay for the amount of data the model receives and generates. This is measured in tokens — small pieces of text. A token can be part of a word, a whole word or punctuation.
The table lists prices in US dollars per 1 million tokens. You do not need to use a million at once: costs are calculated from the actual size of each request.
- Input — what you send to the model
Your message, instructions, conversation history included in the request and document contents. If you send a document and ask for a summary, both the document and your request count as Input. The more data you send, the more input tokens you use.
- Output — what the model generates
The model’s response: text, code, a translation or an analysis. Short answers usually use fewer tokens than detailed ones. Some models also bill reasoning tokens, depending on the model and its pricing.
- Cached input — repeated input costs less
When you keep working in the same chat or session in Claude Code, OpenCode, Codex and other AI tools, the app usually sends instructions and previous conversation history to the model again. If the beginning of the request matches cached data and the model supports caching, that part is billed at the lower Cached input rate instead of regular Input. This is called a cache hit.
The cache is not a saved answer. The model reuses input data but generates a new answer to your request.
- Cache write — creating the cache
Preparing input data for reuse. If the model prices this operation separately, its cost appears in the Cache write column. Cache creation rules and retention depend on the model.
How to configure your client for caching
Choose a suitable API format when your client lets you configure it:
- OpenAI — prefer the Responses API.
- DeepSeek, Qwen, GLM, Kimi and other Chinese models — usually Chat Completions, unless a different format is specified for the selected model.
- Anthropic / Claude — prefer the Messages API, with caching support in your client.
Using a suitable format helps avoid unnecessary request conversions and preserve model features, including caching.
Send a session identifier
If your client supports custom HTTP headers, configure X-Session-Id: use the same value for requests in one conversation and a new one for each new conversation. When supported by the API, this helps associate requests with the same session and use the cache more effectively.
In Hermes Agent, Cline and other clients, this setting depends on the version and selected connection. If the client already sends a session identifier automatically, you do not need to duplicate it.
Model ID — the model name for your connection
Copy the Model ID from the table and use it in the model field of your app or API request.