Skip to content

Answers

Models

Per-model questions: the rate, the licence, the context window, how to call it, and whether it is servable today.

180 questions

Kimi K2.6

In context →
What does Kimi K2.6 cost on AI Token Router, and what is the official rate?
Input is $0.55/M here against $0.95/M official, and output is $2.30/M against $4.00/M official — $0.40/M off every million input tokens and $1.70/M off every million output tokens. The headline 42% blends the two, weighted three to one towards output, because real workloads write far more than they read and a flat average of the two percentages would overstate what you actually save. Both figures are printed side by side on this page: the official rate is the only benchmark that makes ours mean anything.
Can I call Kimi K2.6 right now?
Yes. A request naming `moonshotai/kimi-k2.6` reaches a live upstream today and is billed at the rates above. 7 of the 24 models in this catalogue are callable right now; the rest are listed at their intended price and marked unavailable on their own pages, so you find that out here rather than from an error at call time.
Does Kimi K2.6 support prompt caching, and what does a cached token cost?
Yes. A repeated prefix — a fixed system prompt, a tool schema, a document you keep querying — is served from the model's KV cache and billed at $0.11/M instead of the standard $0.55/M, which is 80% below the input rate. Moonshot AI's own cached rate is $0.19/M, so the discount applies on top of a rate that is already below official. Nothing has to be set on the request; the prefix is matched automatically. On an agent loop, where most of every request is unchanged from the last one, this is usually the largest single line on the bill.
How do I call Kimi K2.6 from my code?
Point an OpenAI SDK at `https://router.xark.io/api/v1` and pass `model="moonshotai/kimi-k2.6"` — the chat completions endpoint is `https://router.xark.io/api/v1/chat/completions`. Request shape, streaming frames, the usage block and the error envelope all match what your OpenAI client already handles, so the base URL and the model id are the only two lines that change. The bare name `kimi-k2.6` and the URL slug `kimi-k2-6` resolve to the same model. An integration written before the ids were namespaced therefore keeps working.
How large is Kimi K2.6's context window?
256K — 262,144 tokens in a single request, shared between everything you send and everything the model generates back. For comparison, the widest context in this catalogue is Llama 4 Maverick at 1M. Context is billed like any other input: filling the window costs the input rate on every token in it, every turn, which is why the cached-input rate matters more than the window size on a workload that resends the same prefix.
What licence is Kimi K2.6 under, and can I use it commercially?
Kimi K2.6 is published by Moonshot AI under the Modified MIT licence. That is a standard permissive text with publisher-specific modifications, and the modifications are the entire reason for the qualifier — so we will not summarise them from the name. The base terms allow commercial use, modification and redistribution with the notices preserved; whether Moonshot AI's changes narrow any of that for your deployment is a question the licence text answers and this page cannot. Read it before you ship anything that depends on the answer.
How does Kimi K2.6 pricing here compare with other providers carrying it?
Of the 3 other providers we track carrying Kimi K2.6, the lowest is DeepInfra at $0.68/M input and $2.75/M output, against our $0.55/M input and $2.30/M output. That is where the rates stood when we last reconciled them, and it can change without notice — we are not claiming to be the cheapest source for this model, only that our rate sits below Moonshot AI's official $0.95/M. If you find one of these figures stale, tell us and we will correct it.
What is Kimi K2.6 actually good at?
Moonshot's agentic flagship. Holds its reasoning together across very long tool-use chains, which is where most cheaper models quietly fall apart. In practice that points at long-horizon coding tasks and multi-file refactors, agent workflows that chain dozens of tool calls and repo-scale code review and migration passes. The specifics that actually decide a fit — the rate, the licence and what it will and will not serve today — are all on this page rather than behind a sales conversation.

Kimi K3

In context →
What does Kimi K3 cost on AI Token Router, and what is the official rate?
Input is $1.85/M here against $3.00/M official, and output is $9.00/M against $15.00/M official — $1.15/M off every million input tokens and $6.00/M off every million output tokens. The headline 40% blends the two, weighted three to one towards output, because real workloads write far more than they read and a flat average of the two percentages would overstate what you actually save. Both figures are printed side by side on this page: the official rate is the only benchmark that makes ours mean anything.
Can I call Kimi K3 right now?
Yes. A request naming `moonshotai/kimi-k3` reaches a live upstream today and is billed at the rates above. 7 of the 24 models in this catalogue are callable right now; the rest are listed at their intended price and marked unavailable on their own pages, so you find that out here rather than from an error at call time.
Does Kimi K3 support prompt caching, and what does a cached token cost?
Yes. A repeated prefix — a fixed system prompt, a tool schema, a document you keep querying — is served from the model's KV cache and billed at $0.37/M instead of the standard $1.85/M, which is 80% below the input rate. Moonshot AI's own cached rate is $0.60/M, so the discount applies on top of a rate that is already below official. Nothing has to be set on the request; the prefix is matched automatically. On an agent loop, where most of every request is unchanged from the last one, this is usually the largest single line on the bill.
How do I call Kimi K3 from my code?
Point an OpenAI SDK at `https://router.xark.io/api/v1` and pass `model="moonshotai/kimi-k3"` — the chat completions endpoint is `https://router.xark.io/api/v1/chat/completions`. Request shape, streaming frames, the usage block and the error envelope all match what your OpenAI client already handles, so the base URL and the model id are the only two lines that change. The bare name `kimi-k3`, without the provider prefix, resolves to the same model. An integration written before the ids were namespaced therefore keeps working.
How large is Kimi K3's context window?
512K — 524,288 tokens in a single request, shared between everything you send and everything the model generates back. For comparison, the widest context in this catalogue is Llama 4 Maverick at 1M. Context is billed like any other input: filling the window costs the input rate on every token in it, every turn, which is why the cached-input rate matters more than the window size on a workload that resends the same prefix.
What licence is Kimi K3 under, and can I use it commercially?
Kimi K3 is published by Moonshot AI under the Modified MIT licence. That is a standard permissive text with publisher-specific modifications, and the modifications are the entire reason for the qualifier — so we will not summarise them from the name. The base terms allow commercial use, modification and redistribution with the notices preserved; whether Moonshot AI's changes narrow any of that for your deployment is a question the licence text answers and this page cannot. Read it before you ship anything that depends on the answer.
How does Kimi K3 pricing here compare with other providers carrying it?
Of the 2 other providers we track carrying Kimi K3, the lowest is DeepInfra at $2.40/M input and $11.50/M output, against our $1.85/M input and $9.00/M output. That is where the rates stood when we last reconciled them, and it can change without notice — we are not claiming to be the cheapest source for this model, only that our rate sits below Moonshot AI's official $3.00/M. If you find one of these figures stale, tell us and we will correct it.
What is Kimi K3 actually good at?
The current open-weight frontier. Half-million-token context and frontier-class reasoning, at roughly a fifth of what closed models charge for the same work. In practice that points at whole-codebase reasoning without a retrieval layer, research synthesis over hundreds of documents and complex multi-step planning agents. The specifics that actually decide a fit — the rate, the licence and what it will and will not serve today — are all on this page rather than behind a sales conversation.

GLM-5.2

In context →
What does GLM-5.2 cost on AI Token Router, and what is the official rate?
Input is $0.82/M here against $1.40/M official, and output is $2.55/M against $4.40/M official — $0.58/M off every million input tokens and $1.85/M off every million output tokens. The headline 42% blends the two, weighted three to one towards output, because real workloads write far more than they read and a flat average of the two percentages would overstate what you actually save. Both figures are printed side by side on this page: the official rate is the only benchmark that makes ours mean anything.
Can I call GLM-5.2 right now?
Yes. A request naming `z-ai/glm-5.2` reaches a live upstream today and is billed at the rates above. 7 of the 24 models in this catalogue are callable right now; the rest are listed at their intended price and marked unavailable on their own pages, so you find that out here rather than from an error at call time.
Does GLM-5.2 support prompt caching, and what does a cached token cost?
Yes. A repeated prefix — a fixed system prompt, a tool schema, a document you keep querying — is served from the model's KV cache and billed at $0.16/M instead of the standard $0.82/M, which is 80% below the input rate. Z.ai's own cached rate is $0.28/M, so the discount applies on top of a rate that is already below official. Nothing has to be set on the request; the prefix is matched automatically. On an agent loop, where most of every request is unchanged from the last one, this is usually the largest single line on the bill.
How do I call GLM-5.2 from my code?
Point an OpenAI SDK at `https://router.xark.io/api/v1` and pass `model="z-ai/glm-5.2"` — the chat completions endpoint is `https://router.xark.io/api/v1/chat/completions`. Request shape, streaming frames, the usage block and the error envelope all match what your OpenAI client already handles, so the base URL and the model id are the only two lines that change. The bare name `glm-5.2` and the URL slug `glm-5-2` resolve to the same model. An integration written before the ids were namespaced therefore keeps working.
How large is GLM-5.2's context window?
200K — 204,800 tokens in a single request, shared between everything you send and everything the model generates back. For comparison, the widest context in this catalogue is Llama 4 Maverick at 1M. Context is billed like any other input: filling the window costs the input rate on every token in it, every turn, which is why the cached-input rate matters more than the window size on a workload that resends the same prefix.
What licence is GLM-5.2 under, and can I use it commercially?
GLM-5.2 is published by Z.ai under the MIT licence. MIT is a standard permissive licence: commercial use, modification and redistribution are all allowed, and the only obligation is to keep the copyright and licence notice with the software. It sets no user threshold, no field-of-use restriction and no separate acceptable-use policy, so there is nothing in it that limits what you build commercially on top of the output.
How does GLM-5.2 pricing here compare with other providers carrying it?
Of the 3 other providers we track carrying GLM-5.2, the lowest is DeepInfra at $0.95/M input and $3.10/M output, against our $0.82/M input and $2.55/M output. That is where the rates stood when we last reconciled them, and it can change without notice — we are not claiming to be the cheapest source for this model, only that our rate sits below Z.ai's official $1.40/M. If you find one of these figures stale, tell us and we will correct it.
What is GLM-5.2 actually good at?
The strongest all-rounder in the catalog for the money. Genuinely competitive with closed frontier models on coding and tool use, under a plain MIT license. In practice that points at general-purpose production workloads, structured output and function calling at scale and bilingual EN/ZH applications. The specifics that actually decide a fit — the rate, the licence and what it will and will not serve today — are all on this page rather than behind a sales conversation.

GLM-5.2 Air

In context →
What does GLM-5.2 Air cost on AI Token Router, and what is the official rate?
Input is $0.17/M here against $0.28/M official, and output is $0.62/M against $1.10/M official — $0.11/M off every million input tokens and $0.48/M off every million output tokens. The headline 43% blends the two, weighted three to one towards output, because real workloads write far more than they read and a flat average of the two percentages would overstate what you actually save. Both figures are printed side by side on this page: the official rate is the only benchmark that makes ours mean anything.
Can I call GLM-5.2 Air right now?
No. GLM-5.2 Air is catalogued but not currently servable — no configured upstream provider carries it, so a request naming `z-ai/glm-5.2-air` returns `no available channel` instead of a result. The rate on this page is what it will cost once an upstream serves it, not something you can spend today. The text models you can call today are Kimi K2.6, Kimi K3, GLM-5.2, DeepSeek V4 Pro and DeepSeek V4 Flash. 7 of the 24 catalogued models are callable, and every page says which it is.
Does GLM-5.2 Air support prompt caching, and what does a cached token cost?
Yes. A repeated prefix — a fixed system prompt, a tool schema, a document you keep querying — is served from the model's KV cache and billed at $0.034/M instead of the standard $0.17/M, which is 80% below the input rate. Z.ai's own cached rate is $0.056/M, so the discount applies on top of a rate that is already below official. Nothing has to be set on the request; the prefix is matched automatically. On an agent loop, where most of every request is unchanged from the last one, this is usually the largest single line on the bill.
How do I call GLM-5.2 Air from my code?
Point an OpenAI SDK at `https://router.xark.io/api/v1` and pass `model="z-ai/glm-5.2-air"` — the chat completions endpoint is `https://router.xark.io/api/v1/chat/completions`. Request shape, streaming frames, the usage block and the error envelope all match what your OpenAI client already handles, so the base URL and the model id are the only two lines that change. The bare name `glm-5.2-air` and the URL slug `glm-5-2-air` resolve to the same model. An integration written before the ids were namespaced therefore keeps working.
How large is GLM-5.2 Air's context window?
128K — 131,072 tokens in a single request, shared between everything you send and everything the model generates back. For comparison, the widest context in this catalogue is Llama 4 Maverick at 1M. Context is billed like any other input: filling the window costs the input rate on every token in it, every turn, which is why the cached-input rate matters more than the window size on a workload that resends the same prefix.
What licence is GLM-5.2 Air under, and can I use it commercially?
GLM-5.2 Air is published by Z.ai under the MIT licence. MIT is a standard permissive licence: commercial use, modification and redistribution are all allowed, and the only obligation is to keep the copyright and licence notice with the software. It sets no user threshold, no field-of-use restriction and no separate acceptable-use policy, so there is nothing in it that limits what you build commercially on top of the output.
How does GLM-5.2 Air pricing here compare with other providers carrying it?
Of the 3 other providers we track carrying GLM-5.2 Air, the lowest is DeepInfra at $0.20/M input and $0.75/M output, against our $0.17/M input and $0.62/M output. That is where the rates stood when we last reconciled them, and it can change without notice — we are not claiming to be the cheapest source for this model, only that our rate sits below Z.ai's official $0.28/M. If you find one of these figures stale, tell us and we will correct it.
What is GLM-5.2 Air actually good at?
GLM-5.2 distilled for latency and cost. The right default for high-volume classification, extraction and routing steps inside a larger pipeline. In practice that points at high-throughput classification and extraction, cheap router/planner step in front of a larger model and real-time chat where latency beats depth. The specifics that actually decide a fit — the rate, the licence and what it will and will not serve today — are all on this page rather than behind a sales conversation.

DeepSeek V4 Pro

In context →
What does DeepSeek V4 Pro cost on AI Token Router, and what is the official rate?
Input is $0.28/M here against $0.43/M official, and output is $0.55/M against $0.87/M official — $0.15/M off every million input tokens and $0.32/M off every million output tokens. The headline 37% blends the two, weighted three to one towards output, because real workloads write far more than they read and a flat average of the two percentages would overstate what you actually save. Both figures are printed side by side on this page: the official rate is the only benchmark that makes ours mean anything.
Can I call DeepSeek V4 Pro right now?
Yes. A request naming `deepseek/deepseek-v4-pro` reaches a live upstream today and is billed at the rates above. 7 of the 24 models in this catalogue are callable right now; the rest are listed at their intended price and marked unavailable on their own pages, so you find that out here rather than from an error at call time.
Does DeepSeek V4 Pro support prompt caching, and what does a cached token cost?
Yes. A repeated prefix — a fixed system prompt, a tool schema, a document you keep querying — is served from the model's KV cache and billed at $0.055/M instead of the standard $0.28/M, which is 80% below the input rate. DeepSeek's own cached rate is $0.087/M, so the discount applies on top of a rate that is already below official. Nothing has to be set on the request; the prefix is matched automatically. On an agent loop, where most of every request is unchanged from the last one, this is usually the largest single line on the bill.
How do I call DeepSeek V4 Pro from my code?
Point an OpenAI SDK at `https://router.xark.io/api/v1` and pass `model="deepseek/deepseek-v4-pro"` — the chat completions endpoint is `https://router.xark.io/api/v1/chat/completions`. Request shape, streaming frames, the usage block and the error envelope all match what your OpenAI client already handles, so the base URL and the model id are the only two lines that change. The bare name `deepseek-v4-pro`, without the provider prefix, resolves to the same model. An integration written before the ids were namespaced therefore keeps working.
How large is DeepSeek V4 Pro's context window?
160K — 163,840 tokens in a single request, shared between everything you send and everything the model generates back. For comparison, the widest context in this catalogue is Llama 4 Maverick at 1M. Context is billed like any other input: filling the window costs the input rate on every token in it, every turn, which is why the cached-input rate matters more than the window size on a workload that resends the same prefix.
What licence is DeepSeek V4 Pro under, and can I use it commercially?
DeepSeek V4 Pro is published by DeepSeek under the MIT licence. MIT is a standard permissive licence: commercial use, modification and redistribution are all allowed, and the only obligation is to keep the copyright and licence notice with the software. It sets no user threshold, no field-of-use restriction and no separate acceptable-use policy, so there is nothing in it that limits what you build commercially on top of the output.
How does DeepSeek V4 Pro pricing here compare with other providers carrying it?
Of the 3 other providers we track carrying DeepSeek V4 Pro, the lowest is DeepInfra at $0.30/M input and $0.60/M output, against our $0.28/M input and $0.55/M output. That is where the rates stood when we last reconciled them, and it can change without notice — we are not claiming to be the cheapest source for this model, only that our rate sits below DeepSeek's official $0.43/M. If you find one of these figures stale, tell us and we will correct it.
What is DeepSeek V4 Pro actually good at?
Still the best reasoning-per-dollar in open weights. Aggressive cache discounts make it especially cheap for agent loops with a large fixed system prompt. In practice that points at cost-sensitive batch processing at scale, math and reasoning-heavy evaluation pipelines and agent loops with large repeated system prompts. The specifics that actually decide a fit — the rate, the licence and what it will and will not serve today — are all on this page rather than behind a sales conversation.

DeepSeek V4 Flash

In context →
What does DeepSeek V4 Flash cost on AI Token Router, and what is the official rate?
Input is $0.090/M here against $0.14/M official, and output is $0.18/M against $0.28/M official — $0.050/M off every million input tokens and $0.10/M off every million output tokens. The headline 36% blends the two, weighted three to one towards output, because real workloads write far more than they read and a flat average of the two percentages would overstate what you actually save. Both figures are printed side by side on this page: the official rate is the only benchmark that makes ours mean anything.
Can I call DeepSeek V4 Flash right now?
Yes. A request naming `deepseek/deepseek-v4-flash` reaches a live upstream today and is billed at the rates above. 7 of the 24 models in this catalogue are callable right now; the rest are listed at their intended price and marked unavailable on their own pages, so you find that out here rather than from an error at call time.
Does DeepSeek V4 Flash support prompt caching, and what does a cached token cost?
Yes. A repeated prefix — a fixed system prompt, a tool schema, a document you keep querying — is served from the model's KV cache and billed at $0.018/M instead of the standard $0.090/M, which is 80% below the input rate. DeepSeek's own cached rate is $0.028/M, so the discount applies on top of a rate that is already below official. Nothing has to be set on the request; the prefix is matched automatically. On an agent loop, where most of every request is unchanged from the last one, this is usually the largest single line on the bill.
How do I call DeepSeek V4 Flash from my code?
Point an OpenAI SDK at `https://router.xark.io/api/v1` and pass `model="deepseek/deepseek-v4-flash"` — the chat completions endpoint is `https://router.xark.io/api/v1/chat/completions`. Request shape, streaming frames, the usage block and the error envelope all match what your OpenAI client already handles, so the base URL and the model id are the only two lines that change. The bare name `deepseek-v4-flash`, without the provider prefix, resolves to the same model. An integration written before the ids were namespaced therefore keeps working.
How large is DeepSeek V4 Flash's context window?
160K — 163,840 tokens in a single request, shared between everything you send and everything the model generates back. For comparison, the widest context in this catalogue is Llama 4 Maverick at 1M. Context is billed like any other input: filling the window costs the input rate on every token in it, every turn, which is why the cached-input rate matters more than the window size on a workload that resends the same prefix.
What licence is DeepSeek V4 Flash under, and can I use it commercially?
DeepSeek V4 Flash is published by DeepSeek under the MIT licence. MIT is a standard permissive licence: commercial use, modification and redistribution are all allowed, and the only obligation is to keep the copyright and licence notice with the software. It sets no user threshold, no field-of-use restriction and no separate acceptable-use policy, so there is nothing in it that limits what you build commercially on top of the output.
How does DeepSeek V4 Flash pricing here compare with other providers carrying it?
Of the 3 other providers we track carrying DeepSeek V4 Flash, the lowest is DeepInfra at $0.10/M input and $0.20/M output, against our $0.090/M input and $0.18/M output. That is where the rates stood when we last reconciled them, and it can change without notice — we are not claiming to be the cheapest source for this model, only that our rate sits below DeepSeek's official $0.14/M. If you find one of these figures stale, tell us and we will correct it.
What is DeepSeek V4 Flash actually good at?
The cheapest capable model on the platform. Fast enough for interactive use and priced for workloads measured in billions of tokens per month. In practice that points at billion-token-scale batch jobs, summarisation and reranking and cheap first-pass filtering before an expensive model. The specifics that actually decide a fit — the rate, the licence and what it will and will not serve today — are all on this page rather than behind a sales conversation.

Qwen3 Max Instruct

In context →
What does Qwen3 Max Instruct cost on AI Token Router, and what is the official rate?
Input is $0.49/M here against $0.85/M official, and output is $1.95/M against $3.40/M official — $0.36/M off every million input tokens and $1.45/M off every million output tokens. The headline 43% blends the two, weighted three to one towards output, because real workloads write far more than they read and a flat average of the two percentages would overstate what you actually save. Both figures are printed side by side on this page: the official rate is the only benchmark that makes ours mean anything.
Can I call Qwen3 Max Instruct right now?
No. Qwen3 Max Instruct is catalogued but not currently servable — no configured upstream provider carries it, so a request naming `qwen/qwen3-max-instruct` returns `no available channel` instead of a result. The rate on this page is what it will cost once an upstream serves it, not something you can spend today. The text models you can call today are Kimi K2.6, Kimi K3, GLM-5.2, DeepSeek V4 Pro and DeepSeek V4 Flash. 7 of the 24 catalogued models are callable, and every page says which it is.
Does Qwen3 Max Instruct support prompt caching, and what does a cached token cost?
Yes. A repeated prefix — a fixed system prompt, a tool schema, a document you keep querying — is served from the model's KV cache and billed at $0.10/M instead of the standard $0.49/M, which is 80% below the input rate. Alibaba's own cached rate is $0.17/M, so the discount applies on top of a rate that is already below official. Nothing has to be set on the request; the prefix is matched automatically. On an agent loop, where most of every request is unchanged from the last one, this is usually the largest single line on the bill.
How do I call Qwen3 Max Instruct from my code?
Point an OpenAI SDK at `https://router.xark.io/api/v1` and pass `model="qwen/qwen3-max-instruct"` — the chat completions endpoint is `https://router.xark.io/api/v1/chat/completions`. Request shape, streaming frames, the usage block and the error envelope all match what your OpenAI client already handles, so the base URL and the model id are the only two lines that change. The bare name `qwen3-max-instruct`, without the provider prefix, resolves to the same model. An integration written before the ids were namespaced therefore keeps working.
How large is Qwen3 Max Instruct's context window?
256K — 262,144 tokens in a single request, shared between everything you send and everything the model generates back. For comparison, the widest context in this catalogue is Llama 4 Maverick at 1M. Context is billed like any other input: filling the window costs the input rate on every token in it, every turn, which is why the cached-input rate matters more than the window size on a workload that resends the same prefix.
What licence is Qwen3 Max Instruct under, and can I use it commercially?
Qwen3 Max Instruct is published by Alibaba under the Apache 2.0 licence. Apache 2.0 is a standard permissive licence: commercial use, modification and redistribution are allowed, subject to keeping the notices and stating what you changed. It also carries an explicit patent grant from the contributors, which terminates if you bring a patent claim over the work — that grant is usually the reason a legal team prefers it to MIT. No user threshold and no field-of-use restriction apply.
How does Qwen3 Max Instruct pricing here compare with other providers carrying it?
Of the 3 other providers we track carrying Qwen3 Max Instruct, the lowest is DeepInfra at $0.60/M input and $2.40/M output, against our $0.49/M input and $1.95/M output. That is where the rates stood when we last reconciled them, and it can change without notice — we are not claiming to be the cheapest source for this model, only that our rate sits below Alibaba's official $0.85/M. If you find one of these figures stale, tell us and we will correct it.
What is Qwen3 Max Instruct actually good at?
Apache 2.0 with no strings attached, which matters if your legal team reads licenses. Strong multilingual coverage and reliable structured output. In practice that points at products shipping into regulated or licence-sensitive contexts, multilingual applications beyond EN/ZH and jSON-mode and schema-constrained generation. The specifics that actually decide a fit — the rate, the licence and what it will and will not serve today — are all on this page rather than behind a sales conversation.

Qwen3 235B A22B

In context →
What does Qwen3 235B A22B cost on AI Token Router, and what is the official rate?
Input is $0.13/M here against $0.22/M official, and output is $0.51/M against $0.88/M official — $0.090/M off every million input tokens and $0.37/M off every million output tokens. The headline 42% blends the two, weighted three to one towards output, because real workloads write far more than they read and a flat average of the two percentages would overstate what you actually save. Both figures are printed side by side on this page: the official rate is the only benchmark that makes ours mean anything.
Can I call Qwen3 235B A22B right now?
No. Qwen3 235B A22B is catalogued but not currently servable — no configured upstream provider carries it, so a request naming `qwen/qwen3-235b-a22b-instruct` returns `no available channel` instead of a result. The rate on this page is what it will cost once an upstream serves it, not something you can spend today. The text models you can call today are Kimi K2.6, Kimi K3, GLM-5.2, DeepSeek V4 Pro and DeepSeek V4 Flash. 7 of the 24 catalogued models are callable, and every page says which it is.
Does Qwen3 235B A22B support prompt caching, and what does a cached token cost?
Yes. A repeated prefix — a fixed system prompt, a tool schema, a document you keep querying — is served from the model's KV cache and billed at $0.026/M instead of the standard $0.13/M, which is 80% below the input rate. Alibaba's own cached rate is $0.044/M, so the discount applies on top of a rate that is already below official. Nothing has to be set on the request; the prefix is matched automatically. On an agent loop, where most of every request is unchanged from the last one, this is usually the largest single line on the bill.
How do I call Qwen3 235B A22B from my code?
Point an OpenAI SDK at `https://router.xark.io/api/v1` and pass `model="qwen/qwen3-235b-a22b-instruct"` — the chat completions endpoint is `https://router.xark.io/api/v1/chat/completions`. Request shape, streaming frames, the usage block and the error envelope all match what your OpenAI client already handles, so the base URL and the model id are the only two lines that change. The bare name `qwen3-235b-a22b-instruct` and the URL slug `qwen3-235b-a22b` resolve to the same model. An integration written before the ids were namespaced therefore keeps working.
How large is Qwen3 235B A22B's context window?
128K — 131,072 tokens in a single request, shared between everything you send and everything the model generates back. For comparison, the widest context in this catalogue is Llama 4 Maverick at 1M. Context is billed like any other input: filling the window costs the input rate on every token in it, every turn, which is why the cached-input rate matters more than the window size on a workload that resends the same prefix.
What licence is Qwen3 235B A22B under, and can I use it commercially?
Qwen3 235B A22B is published by Alibaba under the Apache 2.0 licence. Apache 2.0 is a standard permissive licence: commercial use, modification and redistribution are allowed, subject to keeping the notices and stating what you changed. It also carries an explicit patent grant from the contributors, which terminates if you bring a patent claim over the work — that grant is usually the reason a legal team prefers it to MIT. No user threshold and no field-of-use restriction apply.
How does Qwen3 235B A22B pricing here compare with other providers carrying it?
Of the 3 other providers we track carrying Qwen3 235B A22B, the lowest is DeepInfra at $0.15/M input and $0.60/M output, against our $0.13/M input and $0.51/M output. That is where the rates stood when we last reconciled them, and it can change without notice — we are not claiming to be the cheapest source for this model, only that our rate sits below Alibaba's official $0.22/M. If you find one of these figures stale, tell us and we will correct it.
What is Qwen3 235B A22B actually good at?
A sparse MoE that punches well above its price. The workhorse choice when you want Apache 2.0 and don't need frontier reasoning. In practice that points at general chat and assistant backends, rAG answer synthesis and self-hostable-later architectures. The specifics that actually decide a fit — the rate, the licence and what it will and will not serve today — are all on this page rather than behind a sales conversation.

The rest, answered on their own pages

Every question below is answered in full where it belongs, beside the rate table and the specification it refers to.

Check it against your own numbers

$5 in free credits, no card required.

Get API Key

Every rate quoted above is the rate the API bills.