> ## Documentation Index
> Fetch the complete documentation index at: https://docs.qingbo.dev/llms.txt
> Use this file to discover all available pages before exploring further.
# Text Models Overview
> The catalog of every text model: what each supports, how it is billed, and where to call it
All text models are served through the OpenAI-compatible Chat Completions endpoint: switch models by changing `model`, with no per-vendor URLs or keys. This page lists every model ID, what it supports and how it is billed, grouped by vendor.
## Entry points
`POST /v1/chat/completions`. The default entry; every model on this page is called here, except the Responses-only models.
`POST /v1/responses`. GPT Pro / Codex / o3-pro models accept only this entry.
`POST /v1/messages`. Native envelope kept for apps already built on the Anthropic SDK.
`POST /v1beta/models/{model}:generateContent`. Native envelope kept for existing Gemini clients.
New integrations should use Chat Completions. The native entries exist for compatibility with existing clients and expose no extra capability. All entries authenticate with a WaveAPI key (Chat / Responses via `Authorization: Bearer`, Messages via `x-api-key`, Gemini native via `x-goog-api-key`); an ecosystem login token is not a model key — see [Authentication](/en/docs/authentication).
## Billing rules
Per-model rates are in `price_config` on `GET /v1/models` and in the console's Model Market. What a specific call cost is the `usage.cost` field on the response.
These rules apply to every text model. Model-specific differences are in the vendor tables and the model notes below them.
* **Where rates come from**: 10% off the official list price or equal to it, never above; Batch, Flex and Priority prices are not used. For the actual numbers, read `price_config` from `GET /v1/models`.
* **Total input includes cache**: `prompt_tokens` is total input; reported cache-hit / cache-write tokens are billed at their cache rates, the rest at the ordinary input rate. Cache tokens are part of the total, never an amount added on top.
* **Reasoning is inside output**: `completion_tokens` already includes `completion_tokens_details.reasoning_tokens`; reasoning is billed once at the output rate.
* **Two-rate models**: models billed as input + output have no cache rate; any cache statistics in the response are charged at the ordinary input rate, and explicit `cache_control` returns 400.
* **Long-context tiers**: a tiered model picks its tier from **total input (cache included)**; once the threshold is reached the whole request — input, cache and output — moves to the higher rate, not just the excess. Thresholds and how the exact boundary value is treated are stated per vendor.
* **Time-of-day pricing**: DeepSeek V4 is priced by the UTC slot at request start; see the DeepSeek section.
* **Quota conversion**: the line items are summed and converted to an integer quota at **500,000 quota = 1 USD**; a request with positive usage that rounds below 1 quota is charged 1, then the account group multiplier is applied and the result truncated. `usage.cost` in the response is that integer quota, **not dollars**.
All rates are in USD per 1 million tokens:
```text theme={"system"}
ordinary input tokens = prompt_tokens - cached_tokens - cache_write_tokens
cost (USD) = ( ordinary input tokens × input rate
+ cached_tokens × cache-read rate
+ cache_write_tokens × cache-write rate
+ completion_tokens × output rate ) / 1,000,000
quota = cost (USD) × 500,000, truncated, minimum 1
```
The final charge is the text ledger entry in the console billing records. The states mean:
| Ledger state | Meaning |
| --------------------------------------- | --------------------------------------------------------------------------------- |
| `reserved` | Credit is held, result not yet confirmed; the held amount is not the final charge |
| `pricing_pending` | Usage received, cost under review; the credit stays held |
| `settlement_pending` / `refund_pending` | Waiting for settlement or refund to complete |
| `settled` / `refunded` | Charge or refund completed; reconcile against the final ledger amount |
A missing final `usage`, a dropped connection or a response without a cost field **does not make the call free**. Check the existing usage record in the console before resubmitting; never retry automatically on that basis.
## Caching
Caching is graded per model by which cache rates that model has configured. Cache reads and writes
**replace** the matching ordinary input charge at their own rate; they are not added on top.
| Grade | Test | Models |
| --------------------------------------- | ----------------------------------------------- | --------------------------------------------------------------------------------------------------------------- |
| No cache discount | `cache_billing: "input_output"` | every other model in the catalog; any cache statistics in the response bill at the ordinary input rate |
| Cache read only | has `cache_read` | `gpt-5.5` `gpt-5.4` `grok-4.6` `gemini-3.6-flash` `gemini-3.7-flash` `gemini-3.8-flash` `kimi-k3` `qwen3.8-max` |
| Cache read + write | also `cache_write` | `gpt-5.6-luna` `gpt-5.6-terra` `gpt-5.6-sol` `gpt-6-astra` |
| Cache read + 5-min write + 1-hour write | also `cache_write_1h` | `claude-opus-5` `claude-sonnet-5` `claude-fable-5` `claude-fable-5-1` |
| Time-of-day cache rate | `cache_read` inside each `text_schedule` window | `deepseek-v4-pro` `deepseek-v4-flash` |
### Automatic vs explicit caching
**Automatic caching** is decided upstream and needs no parameters from you: hit tokens appear in
`usage.prompt_tokens_details.cached_tokens` and bill at that model's cache-read rate. Hits are never guaranteed, so cost estimates should not assume one.
**Explicit caching** marks a reusable prefix with `cache_control`. It is allowed per model:
| Models | Explicit `cache_control` |
| -------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- |
| `claude-opus-5` `claude-sonnet-5` `claude-fable-5` `claude-fable-5-1` | ✅ Allowed; all three cache rates configured (read / 5-minute write / 1-hour write) |
| every `gemini-*`, and `qwen3.8-max` | ❌ Automatic caching only; sending `cache_control` returns 400 |
| two-rate models | ❌ No cache billing; sending `cache_control` returns 400 |
| everything else (`gpt-5.4/5.5/5.6-*`, `gpt-6-astra`, `grok-4.6`, `kimi-k3`, `deepseek-v4-*`) | Upstream caches automatically; `cache_control` is neither needed nor used |
Unsupported requests return `400` and are not billed.
On first integration, send a small request and check that the cache line items in your console usage record look as expected before scaling up.
### Writing a Claude explicit cache request
On the OpenAI-compatible endpoint, `content` **must be an array of content blocks**. A string
`content` cannot carry `cache_control`; the marker is ignored and the whole prefix bills as ordinary input.
```json theme={"system"}
{
"model": "claude-sonnet-5",
"max_tokens": 1024,
"messages": [
{
"role": "system",
"content": [
{
"type": "text",
"text": "Put the long prefix you want to reuse here...",
"cache_control": {"type": "ephemeral"}
}
]
},
{"role": "user", "content": "Answer using the content above."}
]
}
```
The native [Messages endpoint](/en/api-reference/text/claude-messages) takes the same shape with the
prefix in the `system` array. `ttl` defaults to 5 minutes; for the 1-hour tier send
`"cache_control": {"type": "ephemeral", "ttl": "1h"}` together with the
`anthropic-beta: extended-cache-ttl-2025-04-11` header — the gateway forwards that header upstream as-is.
### Hit conditions and reading usage
* **Prefix length** — usually at least \~1,024 tokens; shorter prefixes create no cache.
* **Byte-identical prefix** — any difference in text, spaces, newlines or content-block order misses.
* **Reuse within the TTL** — 5 minutes or 1 hour, counted from the most recent hit.
In Chat responses the hit count is `usage.prompt_tokens_details.cached_tokens`. Native Messages reports
`cache_read_input_tokens` and `cache_creation_input_tokens` instead, and its `input_tokens` **excludes**
both — total input is the sum of the three (whereas Chat's `prompt_tokens` already includes them).
Written tokens bill at the cache-write rate, which is higher than ordinary input, so explicit caching suits prefixes that are stable and reused within the TTL.
## Not supported
The following return `400` and are not billed:
* **Vendor built-in tools** — any entry in `tools` whose type is not `function` or `custom` (hosted web search, hosted execution, and similar). Client-executed function tools are not in this group and work normally.
* **`web_search_options`**.
* **`service_tier`** set to anything other than `standard` / `default`.
* **Explicit caching** — see the allow-list above.
Each model serves only the protocols it supports (Chat / Responses / Claude Messages / Gemini native). Sending a request to a protocol a model does not serve, or sending `tools`, `tool_choice` or `response_format` to a model that does not support them, also returns `400`. The protocols and capabilities each model supports are in the catalog below.
## Model catalog
**How to read the tables**: every model supports non-streaming and streaming text output, so those are not given a column. In the JSON / Tools column the left mark is JSON Schema structured output and the right one is function calling — **the platform returns the function name and arguments; your application runs the tool**. `✓` means supported, `—` means not supported. The Billing column names which dimensions a model is billed on; for rates see [Catalog API](#catalog-api). A model with a long-context tier picks its tier from **total input** (cache included) and switches the whole request once the threshold is reached; the thresholds are in the model notes under each table. Model notes also carry each model's input and output limits and any model-specific restrictions; a limit is the maximum the endpoint accepts, and input and output cannot both be maxed out at once.
### OpenAI
| Model ID | Billing | JSON / Tools |
| --------------- | ------------------ | :----------: |
| `gpt-6-astra` | + cache read/write | ✓ / — |
| `gpt-5.6-luna` | + cache read/write | ✓ / ✓ |
| `gpt-5.6-terra` | + cache read/write | ✓ / ✓ |
| `gpt-5.6-sol` | + cache read/write | ✓ / ✓ |
| `gpt-5.5` | + cache read | ✓ / ✓ |
| `gpt-5.4` | + cache read | ✓ / ✓ |
| `gpt-5.4-pro` | input + output | — / — |
| `gpt-5.4-mini` | input + output | ✓ / ✓ |
| `gpt-5.4-nano` | input + output | ✓ / ✓ |
| `gpt-5.3-codex` | input + output | — / — |
| `gpt-5.2` | input + output | ✓ / ✓ |
| `gpt-5.2-pro` | input + output | — / — |
| `gpt-5.1` | input + output | ✓ / ✓ |
| `gpt-5` | input + output | — / ✓ |
| `gpt-5-mini` | input + output | ✓ / ✓ |
| `gpt-5-nano` | input + output | ✓ / ✓ |
| `gpt-5-pro` | input + output | — / — |
| `gpt-4.1` | input + output | ✓ / ✓ |
| `gpt-4.1-mini` | input + output | ✓ / ✓ |
| `gpt-4.1-nano` | input + output | ✓ / ✓ |
| `o3` | input + output | ✓ / ✓ |
| `o3-pro` | input + output | — / — |
| `o4-mini` | input + output | ✓ / ✓ |
| `o3-mini` | input + output | ✓ / ✓ |
| `o1` | input + output | ✓ / ✓ |
**Model notes**
* `gpt-6-astra` — 272K long-context tier; has a separate cache-write rate; max input 922K, max output 12.8K; tool calling is not supported; use `reasoning_effort: "low"`; `"none"` is unsupported, and `temperature`, `top_p` and `logprobs` are not supported
* `gpt-5.6-luna` — 272K long-context tier; has a separate cache-write rate; max input 922K, max output 128K; supports automatic cache hits
* `gpt-5.6-terra` — 272K long-context tier; has a separate cache-write rate; max input 922K, max output 12.8K
* `gpt-5.6-sol` — 272K long-context tier; has a separate cache-write rate; max input 922K, max output 12.8K
* `gpt-5.5` — 272K long-context tier; max input 1M, max output 12.8K; no separate cache-write rate
* `gpt-5.4` — 272K long-context tier; max input 1M, max output 12.8K; no separate cache-write rate
* `gpt-5.4-pro` — 272K long-context tier; `/v1/responses` only
* `gpt-5.3-codex` — `/v1/responses` only
* `gpt-5.2-pro` — `/v1/responses` only
* `gpt-5-pro` — `/v1/responses` only
* `o3-pro` — `/v1/responses` only
GPT-5 and later use `max_completion_tokens` to cap output and `reasoning_effort` to pick a reasoning tier (GPT-6 Astra supports `"low"` only). `max_tokens` is rejected. `gpt-5-pro`, `gpt-5.2-pro`, `gpt-5.4-pro`, `gpt-5.3-codex` and `o3-pro` are served only on the [Responses API](/en/api-reference/text/openai-multimodal) and return 400 on Chat. The 272K tier is judged on **total input**: exactly 272,000 stays on the base rate, anything above moves the whole request (cache and output included) to the long-context rate. GPT-5.6 cache writes cost 1.25× the tier's input rate and replace the ordinary input charge rather than adding to it; GPT-5.4 / 5.5 have no separate cache-write rate.
### Anthropic
| Model ID | Billing | JSON / Tools |
| ------------------- | --------------- | :----------: |
| `claude-fable-5-1` | + 3 cache rates | — / ✓ |
| `claude-fable-5` | + 3 cache rates | — / ✓ |
| `claude-opus-5` | + 3 cache rates | — / ✓ |
| `claude-sonnet-5` | + 3 cache rates | — / ✓ |
| `claude-opus-4-8` | input + output | — / ✓ |
| `claude-opus-4-7` | input + output | — / ✓ |
| `claude-opus-4-6` | input + output | — / ✓ |
| `claude-opus-4-5` | input + output | — / ✓ |
| `claude-sonnet-4-6` | input + output | — / ✓ |
| `claude-sonnet-4-5` | input + output | — / ✓ |
| `claude-haiku-4-5` | input + output | — / ✓ |
**Model notes**
* `claude-fable-5-1` — cache writes are priced at two TTLs, 5-minute and 1-hour; max input 1M, max output 128K; JSON Schema structured output is not supported; `tool_choice` accepts only `"auto"`; `"required"` and named functions return 400
* `claude-fable-5` — cache writes are priced at two TTLs, 5-minute and 1-hour; max input 1M, max output 128K; JSON Schema structured output is not supported; `tool_choice` accepts only `"auto"`
* `claude-opus-5` — cache writes are priced at two TTLs, 5-minute and 1-hour; max input 128K, max output 16K; JSON Schema and `json_object` are not supported; the native `/v1/messages` endpoint is available
* `claude-sonnet-5` — cache writes are priced at two TTLs, 5-minute and 1-hour; max input 1M, max output 128K; JSON Schema and `json_object` are not supported; the native `/v1/messages` endpoint is available
Use `max_tokens` on the Chat endpoint to cap output; the GPT-only `reasoning_effort` and `max_completion_tokens` do not apply to Claude models. The four Claude 5 / Fable models carry three cache rates (read / 5-minute write / 1-hour write) and are the only models that currently **allow explicit `cache_control`** — see [Caching](#caching) for the request shape. These four do **not support structured output**: `response_format` with `json_schema` or `json_object` returns 400. On the Fable models `tool_choice` accepts only `"auto"`; `"required"` or a named function returns 400. The 4.x line bills input and output only and does not support `cache_control`. Apps built on the Anthropic SDK can use the native [Messages endpoint](/en/api-reference/text/claude-messages), but **only `claude-sonnet-5` and `claude-opus-5` support it**; any other Claude model returns 400 there — use the Chat endpoint instead.
### Google Gemini
| Model ID | Billing | JSON / Tools |
| ------------------------ | -------------- | :----------: |
| `gemini-3.8-flash` | + cache read | ✓ / ✓ |
| `gemini-3.7-flash` | + cache read | ✓ / ✓ |
| `gemini-3.6-flash` | + cache read | ✓ / ✓ |
| `gemini-3.5-flash` | input + output | ✓ / ✓ |
| `gemini-3.5-flash-lite` | input + output | ✓ / ✓ |
| `gemini-3.1-pro-preview` | input + output | ✓ / — |
| `gemini-3-flash-preview` | input + output | ✓ / ✓ |
| `gemini-2.5-pro` | input + output | ✓ / ✓ |
| `gemini-2.5-flash-lite` | input + output | ✓ / ✓ |
**Model notes**
* `gemini-3.8-flash` — max input 1M, max output 65K; the Gemini native endpoint is available
* `gemini-3.7-flash` — max input 1M, max output 65K; the Gemini native endpoint is available
* `gemini-3.6-flash` — max input 1M, max output 65K; the Gemini native endpoint is available
* `gemini-3.5-flash` — the Gemini native endpoint is available
* `gemini-3.1-pro-preview` — 200K long-context tier; the Gemini native endpoint is available
* `gemini-3-flash-preview` — the Gemini native endpoint is available
* `gemini-2.5-pro` — 200K long-context tier; the Gemini native endpoint is available
* `gemini-2.5-flash-lite` — the Gemini native endpoint is available
Besides the OpenAI-compatible endpoint, these models can also be called on the Gemini native endpoint: `gemini-2.5-flash-lite`, `gemini-2.5-pro`, `gemini-3-flash-preview`, `gemini-3.1-pro-preview`, `gemini-3.5-flash`, `gemini-3.6-flash`, `gemini-3.7-flash`, `gemini-3.8-flash` (`gemini-3.5-flash-lite` is Chat-only). The native entry is `POST /v1beta/models/{model}:generateContent` (streaming: `:streamGenerateContent?alt=sse`), see [Gemini native API](/en/api-reference/text/gemini-native). Native responses report usage in `usageMetadata` without `usage.cost`; `candidatesTokenCount` and `thoughtsTokenCount` are reported separately and billed output is their sum. The Pro models (`gemini-2.5-pro` / `gemini-3.1-pro-preview`) have a 200K long-context tier that switches the whole request above 200,000 total input. 3.6 / 3.7 / 3.8 Flash have an implicit cache-read rate whose hits are decided upstream; the other models bill two rates. The whole Gemini line supports automatic caching only; explicit caching (`cache_control` / `cachedContent`) returns 400, as do built-in tools, `web_search_options` and any `service_tier` other than `standard` / `default`.
### DeepSeek
| Model ID | Billing | JSON / Tools |
| ------------------------ | -------------- | :----------: |
| `deepseek-v4-pro` | time-of-day | ✓ / ✓ |
| `deepseek-v4-flash` | time-of-day | ✓ / ✓ |
| `deepseek-r1-0528` | input + output | ✓ / ✓ |
| `deepseek-v3.2` | input + output | ✓ / ✓ |
| `deepseek-v3.2-exp` | input + output | — / ✓ |
| `deepseek-v3.1-terminus` | input + output | — / ✓ |
**Model notes**
* `deepseek-v4-pro` — max input 1M, max output 393K; supports automatic cache hits; time-of-day pricing, see the table at the end of this section
* `deepseek-v4-flash` — max input 128K, max output 16K; supports automatic cache hits; time-of-day pricing, see the table at the end of this section
* `deepseek-v3.2-exp` — no strict JSON Schema; separate model from `deepseek-v3.2` (2025-09-29 experimental build)
* `deepseek-v3.1-terminus` — no strict JSON Schema; no longer listed on DeepSeek's main API
V4 Pro / Flash are priced by the **UTC time slot at request start** (Beijing time = UTC+8). The rate is fixed when the request begins and is used for both the response cost and the final settlement, even if the response crosses a slot boundary. Pass `thinking: {"type": "disabled"}` to turn thinking off. V4 cache-hit rates replace the matching input charge; there is no separate cache write. V3.x / R1 bill two rates; `deepseek-v3.2-exp` and `deepseek-v3.1-terminus` offer no strict JSON Schema — use `deepseek-v3.2` or `deepseek-r1-0528` when you need schema guarantees.
| Slot (UTC, start inclusive, end exclusive) | Tier |
| ------------------------------------------ | ------------------- |
| Mon–Fri 01:00–04:00, 06:00–10:00 | peak |
| Daily 00:00–14:00 outside the slots above | standard |
| Daily 14:00–24:00 | off-peak (cheapest) |
Beijing time is UTC+8. The per-slot input / output / cache-read rates for Pro and Flash are in
`price_config.text_schedule` on `GET /v1/models` — that is the complete table and the basis for settlement.
### Qwen (Alibaba Cloud)
| Model ID | Billing | JSON / Tools |
| --------------- | -------------- | :----------: |
| `qwen3.8-max` | + cache read | ✓ / ✓ |
| `qwen3.7-max` | input + output | ✓ / ✓ |
| `qwen3.7-plus` | input + output | ✓ / ✓ |
| `qwen3.6-plus` | input + output | — / ✓ |
| `qwen3.6-flash` | input + output | — / ✓ |
**Model notes**
* `qwen3.8-max` — max input 983K, max output 131K; supports automatic cache hits; explicit `cache_control` is not supported; use `tool_choice: "auto"`
* `qwen3.7-plus` — 256K long-context tier
* `qwen3.6-plus` — 256K long-context tier
* `qwen3.6-flash` — 256K long-context tier
Qwen3.8 Max supports automatic caching only, and whether a request hits is decided upstream; explicit `cache_control` returns 400. Plus / Flash have a 256K tier switched on total input for the whole request. Reasoning tokens come out of the output budget and cannot be turned off.
### Kimi (Moonshot)
| Model ID | Billing | JSON / Tools |
| -------------------------- | -------------- | :----------: |
| `kimi-k3` | + cache read | ✓ / ✓ |
| `kimi-k2.7-code` | input + output | ✓ / ✓ |
| `kimi-k2.7-code-highspeed` | input + output | ✓ / ✓ |
| `kimi-k2.6` | input + output | — / ✓ |
**Model notes**
* `kimi-k3` — max input 1M, max output 105K; supports automatic cache hits; use `tool_choice: "auto"`; give function calls an output budget of 512 or more
Kimi K3 supports automatic caching. Use `tool_choice: "auto"`, and give function calls enough output budget — 512 or more. K2.x bill input and output only.
### Grok (xAI)
| Model ID | Billing | JSON / Tools |
| ------------------------------ | -------------- | :----------: |
| `grok-4.6` | + cache read | ✓ / — |
| `grok-4.5` | input + output | ✓ / — |
| `grok-4.3` | input + output | ✓ / ✓ |
| `grok-4.20-0309-reasoning` | input + output | ✓ / ✓ |
| `grok-4.20-0309-non-reasoning` | input + output | ✓ / ✓ |
| `grok-build-0.1` | input + output | ✓ / ✓ |
**Model notes**
* `grok-4.6` — 200K long-context tier; max input 500K, max output 12.8K; supports automatic cache hits; tool calling is not supported; `max_tokens` is not a hard cap; reasoning can push completion\_tokens past it
* `grok-4.5` — 200K long-context tier; tool calling is not supported
* `grok-4.3` — 200K long-context tier
* `grok-4.20-0309-reasoning` — 200K long-context tier
* `grok-4.20-0309-non-reasoning` — 200K long-context tier
* `grok-build-0.1` — 200K long-context tier
The whole Grok line has a 200K tier: once total input (cache included) reaches 200,000, input, cache read and output for the entire request move to the higher rate — not just the excess. Grok 4.6 supports automatic caching. The whole line does **not support tool calling** — sending `tools` returns 400. `max_tokens` is not a hard cap on reasoning models: `completion_tokens` includes reasoning tokens and can exceed it, so budget with key quotas.
### GLM (Z.ai)
| Model ID | Billing | JSON / Tools |
| --------- | -------------- | :----------: |
| `glm-5.2` | input + output | — / ✓ |
| `glm-5.1` | input + output | — / ✓ |
| `glm-5` | input + output | ✓ / ✓ |
| `glm-4.7` | input + output | ✓ / ✓ |
| `glm-4.6` | input + output | — / ✓ |
Input and output are the only two rates; `cache_control` is not supported.
### MiniMax
| Model ID | Billing | JSON / Tools |
| -------------- | -------------- | :----------: |
| `minimax-m3` | input + output | — / ✓ |
| `minimax-m2.7` | input + output | — / ✓ |
| `minimax-m2.5` | input + output | — / ✓ |
| `minimax-m2.1` | input + output | — / ✓ |
Input and output are the only two rates; `cache_control` is not supported.
### Others
| Model ID | Billing | JSON / Tools |
| ---------------- | -------------- | :----------: |
| `mimo-v2.5-pro` | input + output | ✓ / ✓ |
| `step-3.7-flash` | input + output | — / ✓ |
**Model notes**
* `mimo-v2.5-pro` — Xiaomi MiMo
* `step-3.7-flash` — StepFun Step
Input and output are the only two rates; `cache_control` is not supported.
## Catalog API
Live availability, price configuration and capability flags come from the catalog API:
```bash theme={"system"}
curl https://www.qingbo.dev/v1/models \
-H "Authorization: Bearer $WAVE_API_KEY"
```
Single model: `GET /v1/models/{model}`. The `price_config` it returns is where this page's tables come from: `input` / `output` are the base rates, `cache_read` / `cache_write` / `cache_write_1h` show how far caching is opened up for that model, `input_tier_threshold` with the `*_above_price` fields is the long-context tier, and `cache_billing: "input_output"` means input and output are the only two rates.
The catalog returns every text model currently live. A model that does not appear in the response cannot be called.