LLM Broker

One OpenAI-compatible key for every model: measured benchmark scores and live prices, each request routed to the cheapest offer that meets your criteria.

Hosted MCP Server

npx add-mcp 'https://api.llm-broker.net/api/v1/broker/mcp'

Installs into Claude Code, Codex, Cursor and more

Documentation

Grimoire llm-broker 0.39 / 1.95 Hexblade llm-broker 0.39 / 1.95 Oracle llm-broker 0.39 / 1.95 Wraith llm-broker 0.39 / 1.95 GLM 5.3 Flash Morph 0.11 / 0.31 ▼ GLM 5.3 Decart 0.65 / 2.09 ▲ DeepSeek V4.1 Flash 1017 Wafer 0.070 / 0.50 GLM 5.3 Flash DigitalOcean 0.098 / 0.33 DeepSeek V4.1 Flash 1017 DigitalOcean 0.15 / 0.60 ▼ Kimi K3.1 Pro DigitalOcean 1.90 / 9.50 ▼ Kimi K3.1 Pro Morph 1.31 / 8.55 ▼ Kimi K3.1 Pro InferenceNet 1.80 / 9.00 Kimi K3.1 Pro InferenceNet 1.80 / 9.00 DeepSeek V4.1 Flash 1017 Particle AI 0.030 / 0.28 ▲ DeepSeek V4.1 Flash 1017 Particle AI 0.060 / 0.55 ▲ Minimax M3.1 Flash Minimax 0.075 / 0.30 ▼ DeepSeek V4.1 Flash 1017 Reka 0.30 / 1.20 ▲ GLM 5.3 Flash Particle AI 0.13 / 0.65 ▲ GLM 5.3 Flash Particle AI 0.066 / 0.33 ▲ DeepSeek V4.1 Flash 1017 Alibaba 0.091 / 0.36 ▼ GLM 5.3 Doubleword 0.70 / 2.20 ▼ GLM 5.3 Inceptron 1.05 / 3.50 ▲ DeepSeek V4.1 Flash 1017 Baidu 0.15 / 0.60 ▼ GLM 5.3 Featherless 0.98 / 3.08 ▲ GLM 5 Novita 0.90 / 2.88 ▼ GLM 5.3 Flash Inceptron 0.16 / 0.50 Ling 3.1 Flash Novita 0.060 / 0.18 ▲ GLM 5.3 Morph 1.05 / 4.88 ▲ GLM 5.3 Flash DeepInfra 0.10 / 0.33 ▼ GLM 5.3 Flash DeepInfra 0.10 / 0.33 ▼ DeepSeek V4.1 Flash 1017 OpenInference 0.060 / 0.80 ▼ GLM 5.3 Flash Cohere 0.10 / 0.34 ▼ DeepSeek V4.1 Flash 1017 Cohere 0.12 / 0.40 ▼ DeepSeek V4.1 Flash 1017 Ambient 0.045 / 0.25 ▲ GLM 5.3 Fireworks 0.70 / 2.20 ▼ DeepSeek V4.1 Flash 1017 Mistral 0.30 / 1.20 ▲ DeepSeek V4.1 Flash 1017 Fireworks 0.15 / 0.60 ▼ DeepSeek V4.1 Flash 1017 InferenceNet 0.080 / 0.67 ▲ DeepSeek V4.1 Flash 1017 InferenceNet 0.080 / 0.67 ▲ Mimo V2.6 Pro PrimaLabs 2.64 / 5.28 ▼ DeepSeek V4.1 Flash 1017 Nebius 0.18 / 0.72 ▼ DeepSeek V4.1 Flash 1017 Novita 0.16 / 0.66 ▲ DeepSeek V4.1 Flash 1017 Inceptron 0.10 / 0.60 ▼ GLM 5.3 Flash OpenInference 0.050 / 0.45 DeepSeek V4.1 Flash 1017 DekaLLM 0.10 / 0.60 ▼ Minimax M3.1 Flash SambaNova 0.15 / 0.60 ▲ DeepSeek V4.1 Flash 1017 Makora 0.15 / 0.80 ▲ Minimax M3.1 Flash Makora 0.15 / 0.50 ▼ Mimo V2.6 Flash Makora 0.20 / 0.40 ▲ GLM 5.3 SiliconFlow 0.70 / 1.90 ▼ DeepSeek V4.1 Flash 1017 SiliconFlow 0.15 / 0.60 ▲ GLM 5.3 Makora 1.50 / 4.40 ▲ GLM 5.3 Flash InferenceNet 0.11 / 0.35 GLM 5.3 Flash InferenceNet 0.11 / 0.35 Kimi K3.1 Pro Decart 1.20 / 6.00 ▼ Minimax M3.1 Flash Inferact vLLM 0.15 / 0.60 ▲ Kimi K3.1 Pro PrimaLabs 2.00 / 10.00 ▼ Mimo V2.6 Pro StreamLake 0.43 / 0.87 ▼ GLM 5.3 Flash River 0.12 / 0.40 ▲ DeepSeek V4.1 Flash 1017 River 0.12 / 0.48 ▲ DeepSeek V4.1 Flash 1017 Morph 0.060 / 0.52 ▼ GLM 5.3 Cohere 0.50 / 3.50 GLM 5.3 Tenstorrent 0.60 / 2.50 ▲ Qwen4 Max Alibaba 1.43 / 4.28 ▼ GLM 5.3 Alibaba 1.68 / 5.28 ▼ Kimi K3.1 Pro Alibaba 2.00 / 10.00 ▼ DeepSeek V4.1 Flash 1017 Morph 0.60 / 2.40 ▲ GLM 5.3 Baidu 1.40 / 4.40 ▲ DeepSeek V4.1 Flash 1017 Wafer 0.036 / 0.50 Kimi K3.1 Pro Relace 0.24 / 8.40 ▼ Kimi K3.1 Pro Wafer 0.25 / 8.80 Minimax M3.1 Flash Wafer 0.13 / 0.54 ▼ DeepSeek V4.1 Flash 1017 Ionstream 0.21 / 0.62 ▲ GLM 5.3 Flash Relace 0.039 / 0.079 GPT 6 Luna OpenAI 0.050 / 0.25 Mimo V2.6 Flash GMICloud 0.11 / 0.22 Qwen3.8 Flash DeepInfra 0.11 / 0.38 Claude Haiku 5.5 Amazon Bedrock 0.10 / 0.50 Muse Ember Meta 0.10 / 0.50 Inkling 2 Small PrimeIntellect 0.15 / 0.50 Mimo V2.6 Pro GMICloud 0.33 / 0.65 GLM 5.3 Wafer 0.070 / 1.40 Hy4 Tencent 0.42 / 1.25 GLM 5 StreamLake 0.60 / 1.92 GPT 6.1 Terra OpenAI 0.50 / 2.50 Step 5 Preview Novita 1.00 / 2.70 Gemini 4 Hydrogen Google AI Studio 0.75 / 3.75 GLM 5 Turbo Z.AI 1.20 / 4.00 Qwen4 Max Novita 1.41 / 4.24 GPT 6.1 Sol OpenAI 1.00 / 5.00 GPT 6 Sol OpenAI 1.00 / 5.00 Gemini 4 Argon Google AI Studio 1.00 / 5.00 Gemini 4 Carbon Google AI Studio 1.00 / 5.00 GPT 5.6 Terra OpenAI 1.00 / 6.00 Grimoire llm-broker 0.39 / 1.95 Hexblade llm-broker 0.39 / 1.95 Oracle llm-broker 0.39 / 1.95 Wraith llm-broker 0.39 / 1.95 GLM 5.3 Flash Morph 0.11 / 0.31 ▼ GLM 5.3 Decart 0.65 / 2.09 ▲ DeepSeek V4.1 Flash 1017 Wafer 0.070 / 0.50 GLM 5.3 Flash DigitalOcean 0.098 / 0.33 DeepSeek V4.1 Flash 1017 DigitalOcean 0.15 / 0.60 ▼ Kimi K3.1 Pro DigitalOcean 1.90 / 9.50 ▼ Kimi K3.1 Pro Morph 1.31 / 8.55 ▼ Kimi K3.1 Pro InferenceNet 1.80 / 9.00 Kimi K3.1 Pro InferenceNet 1.80 / 9.00 DeepSeek V4.1 Flash 1017 Particle AI 0.030 / 0.28 ▲ DeepSeek V4.1 Flash 1017 Particle AI 0.060 / 0.55 ▲ Minimax M3.1 Flash Minimax 0.075 / 0.30 ▼ DeepSeek V4.1 Flash 1017 Reka 0.30 / 1.20 ▲ GLM 5.3 Flash Particle AI 0.13 / 0.65 ▲ GLM 5.3 Flash Particle AI 0.066 / 0.33 ▲ DeepSeek V4.1 Flash 1017 Alibaba 0.091 / 0.36 ▼ GLM 5.3 Doubleword 0.70 / 2.20 ▼ GLM 5.3 Inceptron 1.05 / 3.50 ▲ DeepSeek V4.1 Flash 1017 Baidu 0.15 / 0.60 ▼ GLM 5.3 Featherless 0.98 / 3.08 ▲ GLM 5 Novita 0.90 / 2.88 ▼ GLM 5.3 Flash Inceptron 0.16 / 0.50 Ling 3.1 Flash Novita 0.060 / 0.18 ▲ GLM 5.3 Morph 1.05 / 4.88 ▲ GLM 5.3 Flash DeepInfra 0.10 / 0.33 ▼ GLM 5.3 Flash DeepInfra 0.10 / 0.33 ▼ DeepSeek V4.1 Flash 1017 OpenInference 0.060 / 0.80 ▼ GLM 5.3 Flash Cohere 0.10 / 0.34 ▼ DeepSeek V4.1 Flash 1017 Cohere 0.12 / 0.40 ▼ DeepSeek V4.1 Flash 1017 Ambient 0.045 / 0.25 ▲ GLM 5.3 Fireworks 0.70 / 2.20 ▼ DeepSeek V4.1 Flash 1017 Mistral 0.30 / 1.20 ▲ DeepSeek V4.1 Flash 1017 Fireworks 0.15 / 0.60 ▼ DeepSeek V4.1 Flash 1017 InferenceNet 0.080 / 0.67 ▲ DeepSeek V4.1 Flash 1017 InferenceNet 0.080 / 0.67 ▲ Mimo V2.6 Pro PrimaLabs 2.64 / 5.28 ▼ DeepSeek V4.1 Flash 1017 Nebius 0.18 / 0.72 ▼ DeepSeek V4.1 Flash 1017 Novita 0.16 / 0.66 ▲ DeepSeek V4.1 Flash 1017 Inceptron 0.10 / 0.60 ▼ GLM 5.3 Flash OpenInference 0.050 / 0.45 DeepSeek V4.1 Flash 1017 DekaLLM 0.10 / 0.60 ▼ Minimax M3.1 Flash SambaNova 0.15 / 0.60 ▲ DeepSeek V4.1 Flash 1017 Makora 0.15 / 0.80 ▲ Minimax M3.1 Flash Makora 0.15 / 0.50 ▼ Mimo V2.6 Flash Makora 0.20 / 0.40 ▲ GLM 5.3 SiliconFlow 0.70 / 1.90 ▼ DeepSeek V4.1 Flash 1017 SiliconFlow 0.15 / 0.60 ▲ GLM 5.3 Makora 1.50 / 4.40 ▲ GLM 5.3 Flash InferenceNet 0.11 / 0.35 GLM 5.3 Flash InferenceNet 0.11 / 0.35 Kimi K3.1 Pro Decart 1.20 / 6.00 ▼ Minimax M3.1 Flash Inferact vLLM 0.15 / 0.60 ▲ Kimi K3.1 Pro PrimaLabs 2.00 / 10.00 ▼ Mimo V2.6 Pro StreamLake 0.43 / 0.87 ▼ GLM 5.3 Flash River 0.12 / 0.40 ▲ DeepSeek V4.1 Flash 1017 River 0.12 / 0.48 ▲ DeepSeek V4.1 Flash 1017 Morph 0.060 / 0.52 ▼ GLM 5.3 Cohere 0.50 / 3.50 GLM 5.3 Tenstorrent 0.60 / 2.50 ▲ Qwen4 Max Alibaba 1.43 / 4.28 ▼ GLM 5.3 Alibaba 1.68 / 5.28 ▼ Kimi K3.1 Pro Alibaba 2.00 / 10.00 ▼ DeepSeek V4.1 Flash 1017 Morph 0.60 / 2.40 ▲ GLM 5.3 Baidu 1.40 / 4.40 ▲ DeepSeek V4.1 Flash 1017 Wafer 0.036 / 0.50 Kimi K3.1 Pro Relace 0.24 / 8.40 ▼ Kimi K3.1 Pro Wafer 0.25 / 8.80 Minimax M3.1 Flash Wafer 0.13 / 0.54 ▼ DeepSeek V4.1 Flash 1017 Ionstream 0.21 / 0.62 ▲ GLM 5.3 Flash Relace 0.039 / 0.079 GPT 6 Luna OpenAI 0.050 / 0.25 Mimo V2.6 Flash GMICloud 0.11 / 0.22 Qwen3.8 Flash DeepInfra 0.11 / 0.38 Claude Haiku 5.5 Amazon Bedrock 0.10 / 0.50 Muse Ember Meta 0.10 / 0.50 Inkling 2 Small PrimeIntellect 0.15 / 0.50 Mimo V2.6 Pro GMICloud 0.33 / 0.65 GLM 5.3 Wafer 0.070 / 1.40 Hy4 Tencent 0.42 / 1.25 GLM 5 StreamLake 0.60 / 1.92 GPT 6.1 Terra OpenAI 0.50 / 2.50 Step 5 Preview Novita 1.00 / 2.70 Gemini 4 Hydrogen Google AI Studio 0.75 / 3.75 GLM 5 Turbo Z.AI 1.20 / 4.00 Qwen4 Max Novita 1.41 / 4.24 GPT 6.1 Sol OpenAI 1.00 / 5.00 GPT 6 Sol OpenAI 1.00 / 5.00 Gemini 4 Argon Google AI Studio 1.00 / 5.00 Gemini 4 Carbon Google AI Studio 1.00 / 5.00 GPT 5.6 Terra OpenAI 1.00 / 6.00

The best performance at the best price, automatically.

Get an API key Try the API

Authorization: Bearer ast_try · try without an account: model elliot or elliot-unchained, 10 million output tokens per IP in 30 days for free

Tell us once what you need: score or model, model family, region, speed, capabilities, price. Your API key gets exactly that — today and with every new model generation, without changing a line. LLMs, images, speech and transcription with one key; prices compared live, performance measured by us, account and billing via API too. You pay the provider's price plus 5% — no subscription, no base fee.

41 models compared

66 platforms

0.079 USD per 1M output, from

You set the criteria. We keep them current.

Decide once what matters — your API key gets exactly that, re-decided on every request. All fields are part of the same OpenAI-compatible API.

Score or model

The best measured model per use case, a level (solid, strong, top), your own minimum score — say 0.80 in coding — or a fixed model name. Otherwise we take the cheapest one that meets it.

category + level | min_score · model

Model family

For example claude, qwen3 or deepseek/deepseek-v4. When a new generation ships, it is included as soon as it is on the market — with min_score, as soon as we have measured it.

family

Region

Switzerland, EU, USA or Asia: only providers whose execution region we have recorded (eu includes ch).

region

Speed

Rank by measured throughput when response time matters.

sort: "speed"

Capabilities

Tools, structured output, reasoning, image input, minimum context — every chosen provider must support it.

capabilities · min_context

Price

Cheapest first, best score per franc, or a hard maximum price per million tokens.

sort: "price" | "value" · max_price

Integrate once. Never chase models again.

If you build a platform or a tool, today you commit to a model — and tomorrow you switch when a better or cheaper one appears. Here you commit to your requirements, not to a model. The interface stays the same; we keep re-making the choice behind it.

  • New model generations arrive without code changes.
  • If a provider fails, the next one that meets your rules takes over.
  • Market price cuts reach you automatically.
  • If nothing meets your rules, we say so clearly (409) — no silent fallback.
{"model": "james-direct",
 "routing": {"family": "deepseek", "category": "coding", "min_score": 0.8,
             "sort": "value", "region": "eu", "capabilities": ["tools"]},
 "messages": [{"role": "user", "content": "…"}]}

Market prices

Platform list prices in USD per 1 million tokens (daily rate, without our margin), checked every 15 minutes. Bars make price and speed comparable; index = Artificial Analysis via OpenRouter (pre-selection only — we commit only to what we measure ourselves). Region: as recorded by us — otherwise "global" (provider-dependent). Our own models (origin "llm-broker") show our selling price and only our own measurements.

Performance map

Each dot is a model: cheap on the left, strong at the top. The glowing line is the efficiency frontier — no other model is both cheaper and better. Bubble size = context window.

↖ strong & cheap 20 30 40 50 60 70 0.01 0.1 1 10 100 USD per 1M output tokens (log scale) → Reasoning index ↑ Gemini 4 Carbon GPT 6 Astra Gemini 4 Hydrogen GPT 6.1 Terra DeepSeek V4.1 Flash 1017 DeepSeek V4.1 Flash 1017 GLM 5.3 Flash GLM 5.3 Flash GLM 5.3 Flash Mimo V2.6 Flash Claude Haiku 5.5 Ling 3.1 Flash Grok 4.6 Hexblade

Efficiency frontier openaianthropicgooglellm-brokerz-aimetaqwenx-aixiaomideepseekinclusionaiminimaxmoonshotaistepfuntencentthinkingmachines

Reasoning and coding are external indices (Artificial Analysis). Speed and HumanEval+ we measure ourselves; our own models appear there. ×2.2 = most expensive ÷ cheapest platform price of a model, −55 % = saving against the most expensive provider. Price ↓ 90 % = output price over the period (first to last day); ↓ means cheaper.

Price range per platform

The same model costs different amounts on each platform. The glowing dot is the cheapest — that's where we route.

0.1110100

Gemini 4 Neon

×2.0 · −50 %

Gemini 4 Carbon

×2.0 · −50 %

GPT 6 Astra

×2.0 · −50 %

GPT 6.1 Sol

×2.0 · −50 %

Gemini 4 Hydrogen

×2.0 · −50 %

GPT 6.1 Terra

×2.0 · −50 %

Gemini 4 Argon

×2.0 · −50 %

Kimi K3.1 Pro

×3.3 · −70 %

GPT 6 Sol

×2.0 · −50 %

Qwen4 Max

×1.0 · −1 %

Hy4

×2.0 · −50 %

DeepSeek V4.1 Flash 1017

×5.1 · −80 %

Price movement

Cheapest output price (USD per 1M tokens) across all platforms — candles per interval, red = more expensive, teal = cheaper.

Candle = 1 day · since 21.10.26 23:47 · 990 price changes

0.250 ▲ 138.6 %

0.0869 0.432 0.777 1.123 21.10.26 02:00 05.11.26 01:00 20.11.26 01:00 0.250

Value for money

Index points per USD of output price · price movement over the last 30 days.

1. Z GLM 5.3 Flash z-ai 353.5 P/USD Price ↑ 209 %

2. I Ling 3.1 Flash inclusionai 164.5 P/USD Price ↓ 20 %

3. D DeepSeek V4.1 Flash 1017 deepseek 130.3 P/USD Price ↑ 139 %

4. X Mimo V2.6 Flash xiaomi 119.2 P/USD Price ↑ 68 %

5. O GPT 6 Luna openai 111.5 P/USD Price ± 0 %

6. M Minimax M3.1 Flash minimax 96.9 P/USD Price ↓ 50 %

7. Q Qwen3.8 Flash qwen 74.8 P/USD new

8. A Claude Haiku 5.5 anthropic 72.9 P/USD new

9. M Muse Ember meta 71.9 P/USD new

10. T Inkling 2 Small thinkingmachines 61.1 P/USD new

ModelReasoningCodePrice in/outSpeedRegionTrendPlatformContext
Claude Fable 5.2 anthropic658.00 40.00185 tok/s Uptime 100 %globalnewAnthropic · 3 ▾ - Anthropic: 8.00 / 40.00 - Azure: 8.00 / 40.00 - Google: 8.00 / 40.001 M
GPT 6.2 Astra openai628.00 40.0083 tok/s Uptime 100 %globalnewOpenAI · 3 ▾ - OpenAI: 8.00 / 40.00 - Azure: 8.00 / 40.00 - Amazon Bedrock: 8.00 / 40.001 M
Gemini 4 Neon google611.50 7.50108 tok/s Uptime 100 %globalGoogle AI Studio · 2 ▾ - Google AI Studio: 1.50 / 7.50 - Google: 3.00 / 15.001 M
Gemini 4 Carbon google601.00 5.0051 tok/s Uptime 100 %globalnewGoogle AI Studio · 2 ▾ - Google AI Studio: 1.00 / 5.00 - Google: 2.00 / 10.001 M
Claude Opus 5.5 anthropic59HumanEval+ 94 %4.00 20.00205 tok/s Uptime 100 %globalnewGoogle · 6 ▾ - Google: 4.00 / 20.00 - Azure: 4.00 / 20.00 - DeepInfra: 4.00 / 20.00 - Amazon Bedrock: 4.00 / 20.00 - Claude Platform on AWS: 4.00 / 20.00 - Anthropic: 4.00 / 20.001 M
GPT 6.2 Sol openai572.00 10.0082 tok/s Uptime 100 %globalnewOpenAI · 3 ▾ - OpenAI: 2.00 / 10.00 - Azure: 2.00 / 10.00 - Amazon Bedrock: 2.00 / 10.001 M
GPT 6 Astra openai57HumanEval+ 95 %5.00 25.0061 tok/s Uptime 100 %globalOpenAI · 3 ▾ - OpenAI: 5.00 / 25.00 - Azure: 10.00 / 50.00 - Amazon Bedrock: 10.00 / 50.001 M
GPT 6.1 Sol openai551.00 5.00135 tok/s Uptime 100 %globalOpenAI · 3 ▾ - OpenAI: 1.00 / 5.00 - Azure: 2.00 / 10.00 - Amazon Bedrock: 2.00 / 10.001 M
Claude Sonnet 5.5 anthropic55HumanEval+ 94 %2.00 10.00119 tok/s Uptime 100 %globalnewAmazon Bedrock · 6 ▾ - Amazon Bedrock: 2.00 / 10.00 - Google: 2.00 / 10.00 - Azure: 2.00 / 10.00 - DeepInfra: 2.00 / 10.00 - Claude Platform on AWS: 2.00 / 10.00 - Anthropic: 2.00 / 10.001 M
Gemini 4 Hydrogen google540.75 3.75138 tok/s Uptime 100 %globalnewGoogle AI Studio · 2 ▾ - Google AI Studio: 0.75 / 3.75 - Google: 1.50 / 7.501 M
GPT 6.1 Terra openai510.50 2.5071 tok/s Uptime 100 %globalOpenAI · 2 ▾ - OpenAI: 0.50 / 2.50 - Azure: 1.00 / 5.001 M
Gemini 4 Argon google501.00 5.0072 tok/s Uptime 100 %globalnewGoogle AI Studio · 2 ▾ - Google AI Studio: 1.00 / 5.00 - Google: 2.00 / 10.001 M
Claude Fable 5.1 anthropic5010.00 50.0065 tok/s Uptime 100 %globalnewGoogle · 3 ▾ - Google: 10.00 / 50.00 - Azure: 10.00 / 50.00 - Anthropic: 10.00 / 50.001 M
Kimi K3.1 Pro moonshotai471.20 6.00143 tok/s Uptime 100 %globalDecart · 21 ▾ - Decart: 1.20 / 6.00 - Relace: 0.24 / 8.40 - Wafer: 0.25 / 8.80 - Morph: 1.31 / 8.55 - Sail Research: 1.04 / 9.35 - InferenceNet: 1.80 / 9.00 - DigitalOcean: 1.90 / 9.50 - DeepInfra: 1.90 / 9.50 - Makora: 2.00 / 9.49 - AkashML: 1.75 / 10.00 - River: 2.00 / 10.00 - Fireworks: 2.00 / 10.00 - Alibaba: 2.00 / 10.00 - Moonshot AI: 2.00 / 10.00 - PrimaLabs: 2.00 / 10.00 - Parasail: 2.00 / 10.00 - Novita: 2.00 / 10.00 - Modal: 2.00 / 10.00 - Together: 2.00 / 10.00 - Nebius: 2.00 / 10.00 - Amazon Bedrock: 4.00 / 20.001 M
GPT 6 Sol openai441.00 5.00119 tok/s Uptime 100 %globalOpenAI · 3 ▾ - OpenAI: 1.00 / 5.00 - Azure: 2.00 / 10.00 - Amazon Bedrock: 2.00 / 10.001 M
Claude Opus 5 anthropic44HumanEval+ 93 %5.00 25.00193 tok/s Uptime 100 %globalnewAmazon Bedrock · 6 ▾ - Amazon Bedrock: 5.00 / 25.00 - Azure: 5.00 / 25.00 - Claude Platform on AWS: 5.00 / 25.00 - DeepInfra: 5.00 / 25.00 - Anthropic: 5.00 / 25.00 - Google: 5.00 / 25.001 M
Muse Wonder meta444.00 20.00100 tok/s Uptime 100 %globalnewMeta · 2 ▾ - Meta: 4.00 / 20.00 - Google: 4.00 / 20.001 M
Qwen4 Max qwen431.41 4.2448 tok/s Uptime 100 %globalnewNovita · 3 ▾ - Novita: 1.41 / 4.24 - DeepInfra: 1.41 / 4.24 - Alibaba: 1.43 / 4.281 M
Claude Haiku 5.5 anthropic370.10 0.50181 tok/s Uptime 100 %globalnewAmazon Bedrock · 5 ▾ - Amazon Bedrock: 0.10 / 0.50 - Claude Platform on AWS: 0.10 / 0.50 - Google: 0.10 / 0.50 - Anthropic: 0.10 / 0.50 - Azure: 0.10 / 0.501 M
Muse Ember meta360.10 0.50360 tok/s Uptime 100 %globalMeta · 2 ▾ - Meta: 0.10 / 0.50 - Google: 0.10 / 0.501 M
Grok 4.7 x-ai362.00 6.0091 tok/s Uptime 100 %globalAmazon Bedrock · 2 ▾ - Amazon Bedrock: 2.00 / 6.00 - xAI: 2.00 / 6.00500 k
Hy4 tencent350.42 1.2598 tok/s Uptime 100 %globalTencent · 3 ▾ - Tencent: 0.42 / 1.25 - Novita: 0.42 / 1.25 - Together: 0.83 / 2.501 M
DeepSeek V4.1 Flash 1017 deepseek330.045 0.25462 tok/s Uptime 100 %globalAmbient · 41 ▾ - Ambient: 0.045 / 0.25 - Particle AI: 0.030 / 0.28 - DeepInfra: 0.079 / 0.23 - Pearl Research: 0.050 / 0.30 - Alibaba: 0.091 / 0.36 - Cohere: 0.12 / 0.40 - StepFun: 0.10 / 0.42 - Wafer: 0.036 / 0.50 - StreamLake: 0.11 / 0.44 - Morph: 0.060 / 0.52 - Ant Digital: 0.12 / 0.48 - River: 0.12 / 0.48 - Relace: 0.12 / 0.50 - DekaLLM: 0.10 / 0.60 - Inceptron: 0.10 / 0.60 - Cloudflare: 0.10 / 0.60 - PrimeIntellect: 0.044 / 0.70 - InferenceNet: 0.080 / 0.67 - Fireworks: 0.15 / 0.60 - DigitalOcean: 0.15 / 0.60 - SiliconFlow: 0.15 / 0.60 - Baidu: 0.15 / 0.60 - Novita: 0.16 / 0.66 - Ionstream: 0.21 / 0.62 - ModelRun: 0.17 / 0.67 - OpenInference: 0.060 / 0.80 - GMICloud: 0.18 / 0.72 - Nebius: 0.18 / 0.72 - Makora: 0.15 / 0.80 - AtlasCloud: 0.20 / 0.78 - PrimaLabs: 0.24 / 0.96 - Modal: 0.25 / 1.00 - Crusoe: 0.29 / 1.20 - CoreWeave: 0.30 / 1.20 - Near AI: 0.30 / 1.20 - BaseTen: 0.30 / 1.20 - Venice: 0.30 / 1.20 - Mistral: 0.30 / 1.20 - Reka: 0.30 / 1.20 - Together: 0.30 / 1.20 - Parasail: 0.30 / 1.201 M
GPT 5.6 Terra openai321.00 6.00112 tok/s Uptime 100 %globalnewOpenAI · 2 ▾ - OpenAI: 1.00 / 6.00 - Azure: 2.00 / 12.001 M
GLM 5.3 z-ai32HumanEval+ 67 %0.070 1.40152 tok/s Uptime 100 %globalWafer · 37 ▾ - Wafer: 0.070 / 1.40 - DeepInfra: 0.45 / 2.00 - SiliconFlow: 0.70 / 1.90 - Novita: 0.63 / 1.98 - Decart: 0.65 / 2.09 - Cloudflare: 0.70 / 2.20 - GMICloud: 0.70 / 2.20 - Doubleword: 0.70 / 2.20 - Fireworks: 0.70 / 2.20 - Tenstorrent: 0.60 / 2.50 - InferenceNet: 0.083 / 3.60 - Sail Research: 0.75 / 3.00 - CoreWeave: 0.75 / 3.00 - DigitalOcean: 0.91 / 2.86 - Cohere: 0.50 / 3.50 - Featherless: 0.98 / 3.08 - Relace: 0.070 / 4.38 - Inceptron: 1.05 / 3.50 - AkashML: 0.19 / 4.40 - Alibaba: 1.19 / 3.74 - Friendli: 1.26 / 3.96 - SambaNova: 1.40 / 4.40 - Baidu: 1.40 / 4.40 - Modal: 1.40 / 4.40 - BaseTen: 1.40 / 4.40 - AtlasCloud: 1.40 / 4.40 - Mistral: 1.40 / 4.40 - Together: 1.40 / 4.40 - Venice: 1.40 / 4.40 - Parasail: 1.40 / 4.40 - Z.AI: 1.40 / 4.40 - Crusoe: 1.40 / 4.40 - Nebius: 1.40 / 4.40 - Makora: 1.50 / 4.40 - Morph: 1.05 / 4.88 - Amazon Bedrock: 1.68 / 5.28 - ModelRun: 2.10 / 6.601 M
Grok 4.6 x-ai31HumanEval+ 94 %2.00 6.0072 tok/s Uptime 100 %globalnewxAI · 1 ▾ - xAI: 2.00 / 6.00500 k
Inkling 2 Small thinkingmachines310.15 0.50257 tok/s Uptime 100 %globalPrimeIntellect · 6 ▾ - PrimeIntellect: 0.15 / 0.50 - DeepInfra: 0.28 / 0.95 - Together: 0.30 / 1.00 - Cloudflare: 0.30 / 1.00 - Nebius: 0.30 / 1.00 - Modal: 0.30 / 1.001 M
Mimo V2.6 Pro xiaomi300.33 0.65100 tok/s Uptime 100 %globalGMICloud · 6 ▾ - GMICloud: 0.33 / 0.65 - DeepInfra: 0.39 / 0.78 - Novita: 0.43 / 0.87 - Xiaomi: 0.43 / 0.87 - StreamLake: 0.43 / 0.87 - PrimaLabs: 2.64 / 5.281 M
Ling 3.1 Flash inclusionai300.060 0.1871 tok/s Uptime 100 %globalNovita · 1 ▾ - Novita: 0.060 / 0.181 M
Minimax M3.1 Flash minimax290.075 0.30212 tok/s Uptime 100 %globalMinimax · 14 ▾ - Minimax: 0.075 / 0.30 - Makora: 0.15 / 0.50 - Wafer: 0.13 / 0.54 - StreamLake: 0.13 / 0.54 - DigitalOcean: 0.15 / 0.60 - Fireworks: 0.15 / 0.60 - Nebius: 0.15 / 0.60 - GMICloud: 0.15 / 0.60 - Parasail: 0.15 / 0.60 - Inferact vLLM: 0.15 / 0.60 - ModelRun: 0.15 / 0.60 - SambaNova: 0.15 / 0.60 - InferenceNet: 0.15 / 0.60 - Together: 0.15 / 0.601 M
Qwen3.8 Flash qwen290.11 0.3846 tok/s Uptime 100 %globalnewDeepInfra · 3 ▾ - DeepInfra: 0.11 / 0.38 - Alibaba: 0.15 / 0.47 - Novita: 0.15 / 0.471 M
Step 5 Preview stepfun291.00 2.70122 tok/s Uptime 99 %globalnewNovita · 3 ▾ - Novita: 1.00 / 2.70 - StepFun: 1.00 / 2.70 - Featherless: 1.00 / 2.701 M
GPT 6 Luna openai280.050 0.25125 tok/s Uptime 100 %globalOpenAI · 3 ▾ - OpenAI: 0.050 / 0.25 - Azure: 0.10 / 0.50 - Amazon Bedrock: 0.10 / 0.501 M
GLM 5 z-ai280.60 1.9264 tok/s Uptime 100 %globalStreamLake · 8 ▾ - StreamLake: 0.60 / 1.92 - GMICloud: 0.60 / 1.92 - Baidu: 0.70 / 2.24 - SiliconFlow: 0.95 / 2.55 - Novita: 0.90 / 2.88 - Amazon Bedrock: 1.00 / 3.20 - Venice: 1.00 / 3.20 - Z.AI: 1.00 / 3.20204 k
GLM 5.3 Flash z-ai28HumanEval+ 80 %0.039 0.079221 tok/s Uptime 100 %globalRelace · 36 ▾ - Relace: 0.039 / 0.079 - Wafer: 0.042 / 0.18 - Darkbloom: 0.066 / 0.18 - GMICloud: 0.075 / 0.25 - Tenstorrent: 0.075 / 0.25 - Novita: 0.083 / 0.28 - Makora: 0.14 / 0.24 - StreamLake: 0.088 / 0.30 - Particle AI: 0.066 / 0.33 - Reka: 0.12 / 0.30 - Morph: 0.11 / 0.31 - DigitalOcean: 0.098 / 0.33 - DeepInfra: 0.10 / 0.33 - Cohere: 0.10 / 0.34 - InferenceNet: 0.11 / 0.35 - OpenInference: 0.050 / 0.45 - SiliconFlow: 0.10 / 0.40 - CoreWeave: 0.12 / 0.38 - River: 0.12 / 0.40 - Near AI: 0.080 / 0.50 - DekaLLM: 0.10 / 0.50 - Z.AI: 0.15 / 0.50 - Friendli: 0.15 / 0.50 - Fireworks: 0.15 / 0.50 - Mistral: 0.15 / 0.50 - BaseTen: 0.15 / 0.50 - Modal: 0.15 / 0.50 - Baidu: 0.15 / 0.50 - Crusoe: 0.15 / 0.50 - Nebius: 0.15 / 0.50 - Venice: 0.15 / 0.50 - KT: 0.15 / 0.50 - Together: 0.15 / 0.50 - Parasail: 0.15 / 0.50 - Inceptron: 0.16 / 0.50 - Sail Research: 0.18 / 0.631 M
Mimo V2.6 Flash xiaomi270.11 0.22248 tok/s Uptime 100 %globalGMICloud · 9 ▾ - GMICloud: 0.11 / 0.22 - Darkbloom: 0.10 / 0.28 - Novita: 0.14 / 0.28 - DeepInfra: 0.14 / 0.28 - Xiaomi: 0.14 / 0.28 - Cohere: 0.14 / 0.28 - Tenstorrent: 0.16 / 0.33 - Venice: 0.18 / 0.35 - Makora: 0.20 / 0.401 M
GLM 5 Turbo z-ai271.20 4.0052 tok/s Uptime 100 %globalnewZ.AI · 1 ▾ - Z.AI: 1.20 / 4.00202 k
Grimoire llm-broker · uncensoredKompakt 93 %HumanEval+ 90 %0.39 1.9532 tok/s measuredglobalnewllm-broker · 1 ▾ - llm-broker: 0.39 / 1.95–
Hexblade llm-broker · uncensoredKompakt 93 %HumanEval+ 93 %0.39 1.9551 tok/s measuredglobalnewllm-broker · 1 ▾ - llm-broker: 0.39 / 1.95–
Oracle llm-broker · uncensoredKompakt 93 %HumanEval+ 91 %0.39 1.9546 tok/s measuredglobalnewllm-broker · 1 ▾ - llm-broker: 0.39 / 1.95–
Wraith llm-broker · uncensoredKompakt 93 %HumanEval+ 91 %0.39 1.9520 tok/s measuredglobalnewllm-broker · 1 ▾ - llm-broker: 0.39 / 1.95–

Image & speech — same key

Not just LLMs: image generation, speech and transcription from the market, with the same OpenAI-compatible API and the same billing. You name a market id, we pick the cheapest provider. Prices in USD, including our margin, live from /api/v1/broker/models?type=….

A beam of light paints a glowing mountain landscape above a desk

Image generation

From quick drafts to print quality: FLUX, Qwen-Image, Seedream and more — or iris, our own image model.

  • iris 0.0077 USD per image
  • stabilityai/sdxl-turbo 0.00021 USD per image
  • black-forest-labs/FLUX-1-schnell 0.000525 USD per image
  • PrunaAI/p-image 0.0053 USD per image
  • black-forest-labs/FLUX-1-dev 0.0095 USD per image
  • black-forest-labs/FLUX-2-dev 0.0105 USD per image
curl https://llm-broker.net/api/v1/images/generations \
  -H "Authorization: Bearer $BROKER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"black-forest-labs/FLUX-1-schnell","prompt":"a lighthouse at dawn","size":"1024x1024"}'

Glowing sound waves flow out of a studio microphone

Text to speech

Text to speech in many voices and languages — Kokoro, Qwen3-TTS, Chatterbox and more — or quill with cloned voices.

  • quill 0.119 USD per 1,000 characters
  • hexgrad/Kokoro-82M 0.001 USD per 1,000 characters
  • ResembleAI/chatterbox-turbo 0.0011 USD per 1,000 characters
  • ResembleAI/chatterbox-multilingual 0.0011 USD per 1,000 characters
  • sesame/csm-1b 0.0074 USD per 1,000 characters
  • Qwen/Qwen3-TTS 0.021 USD per 1,000 characters
curl https://llm-broker.net/api/v1/audio/speech \
  -H "Authorization: Bearer $BROKER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"hexgrad/Kokoro-82M","input":"Hello from LLM Broker","voice":"af_bella"}' -o hello.mp3

Light trails flow across a city grid at night towards a bright centre

Transcription

Speech to text: Whisper, Qwen3-ASR, Voxtral and more — billed by the duration the provider measures.

  • quill 0.0119 USD per minute
  • nvidia/Nemotron-3.5-ASR-Streaming-Multilingual-0.6b 0.00021 USD per minute
  • openai/whisper-large-v3-turbo 0.00021 USD per minute
  • Qwen/Qwen3-ASR-0.6B 0.00021 USD per minute
  • Qwen/Qwen3-ASR-1.7B 0.000473 USD per minute
  • openai/whisper-large-v3 0.000473 USD per minute
curl https://llm-broker.net/api/v1/audio/transcriptions \
  -H "Authorization: Bearer $BROKER_API_KEY" \
  -F model=openai/whisper-large-v3-turbo -F [email protected]

One prompt — your AI assistant sets everything up

Paste the prompt into Claude Code, Cursor, Codex, OpenCode or any other coding assistant. It reads our docs, opens the account via API, stores the key safely, wires up the API and checks it with a smoke test. You only confirm the payment.

A glass robot hand puts a glowing key into a lock

Claude Code: paste the prompt into the session · Cursor/Windsurf: into the agent chat · OpenCode: opencode run "…" · Codex: codex "…"

Set up LLM Broker in this project (https://llm-broker.net) — one OpenAI-compatible API for LLMs, images, speech
and transcription with a single key. Work on your own; only ask me when a payment must be confirmed or a secret
must be entered.

1. Read the descriptions first, don't guess:
   - https://llm-broker.net/llms.txt (entry point for agents)
   - https://llm-broker.net/api/openapi (all endpoints and schemas)
   - https://astrioncore.wiki/api and https://astrioncore.wiki/api#konto (account and billing via API)
2. Key: if BROKER_API_KEY (starts with ast_sk_) is not set, open an account via the API:
   POST https://llm-broker.net/api/v1/broker/accounts
   {"billing": "topup", "amount": 25, "currency": "USD"}   (every request picks its level via "routing")
   Show me the checkout_url (the card is entered once at Stripe), then poll status_url until it returns 200.
   The api_key is shown EXACTLY ONCE: store it right away in .env as BROKER_API_KEY, add .env to .gitignore,
   never log or commit it. base_url is in the same response.
3. Integrate: the OpenAI SDK of the project's language with base_url and BROKER_API_KEY. Models are listed at
   GET /api/v1/broker/models?type=chat|image|speech|transcription (no key needed).
   - LLM: POST /api/v1/chat/completions, model "james-direct" with "routing" (category, level|min_score, sort,
     region, family, capabilities, max_price) or a market id such as "deepseek/deepseek-v4.1-flash".
   - Images: POST /api/v1/images/generations, speech: POST /api/v1/audio/speech,
     transcription: POST /api/v1/audio/transcriptions — each with a market id of the matching type.
4. Add a small client layer (configuration from the environment, no hard-coded model names — prefer routing),
   write a smoke test with one short chat call and run it.
5. Billing: GET /api/v1/broker/billing shows balance and spend; top up or change the monthly limit with
   PATCH /api/v1/broker/billing, the card with POST|DELETE /api/v1/broker/card.
   Errors: 402 insufficient_balance (top up), 402 outstanding_payment (an earlier request is unpaid — top up,
   it is collected automatically), 409 no_matching_offer (relax routing),
   404 model_not_found (check the market list), 429 rate_limited (retry later).
Finish with a summary of what is set up and where the key is stored.
Richte in diesem Projekt den LLM Broker ein (https://llm-broker.net) — eine OpenAI-kompatible API für LLM, Bilder,
Sprachausgabe und Transkription mit einem Schlüssel. Arbeite selbstständig; frag mich nur, wenn eine Zahlung
bestätigt oder ein Geheimnis eingegeben werden muss.

1. Lies zuerst die Beschreibungen, nicht raten:
   - https://llm-broker.net/llms.txt (Einstieg für Agenten)
   - https://llm-broker.net/api/openapi (alle Endpunkte und Schemata)
   - https://astrioncore.wiki/api und https://astrioncore.wiki/api#konto (Konto und Bezahlung per API)
2. Schlüssel: Ist BROKER_API_KEY (beginnt mit ast_sk_) nicht gesetzt, eröffne ein Konto per API:
   POST https://llm-broker.net/api/v1/broker/accounts
   {"billing": "topup", "amount": 25, "currency": "USD"}   (Leistung wählt jede Anfrage per "routing")
   Zeig mir die checkout_url (die Karte wird einmal bei Stripe eingegeben) und frag danach status_url ab,
   bis 200 kommt. Der api_key erscheint GENAU EINMAL: sofort in .env als BROKER_API_KEY speichern,
   .env in .gitignore, nie ins Log oder in einen Commit. base_url steht in derselben Antwort.
3. Einbinden: OpenAI-SDK der Projektsprache mit base_url und BROKER_API_KEY. Modelle findest du unter
   GET /api/v1/broker/models?type=chat|image|speech|transcription (ohne Schlüssel lesbar).
   - LLM: POST /api/v1/chat/completions, model "james-direct" mit "routing" (category, level|min_score, sort,
     region, family, capabilities, max_price) oder direkt eine Markt-ID wie "deepseek/deepseek-v4.1-flash".
   - Bilder: POST /api/v1/images/generations, Sprache: POST /api/v1/audio/speech,
     Transkription: POST /api/v1/audio/transcriptions — jeweils mit einer Markt-ID der passenden type.
4. Lege eine kleine Client-Schicht an (Konfiguration aus der Umgebung, keine festen Modellnamen im Code —
   lieber routing), schreibe einen Rauchtest mit einem kurzen Chat-Aufruf und führe ihn aus.
5. Abrechnung: GET /api/v1/broker/billing zeigt Guthaben und Verbrauch; Aufladen oder Monatsrahmen ändern
   mit PATCH /api/v1/broker/billing, Karte mit POST|DELETE /api/v1/broker/card.
   Fehler: 402 insufficient_balance (aufladen), 402 outstanding_payment (frühere Anfrage offen — aufladen,
   wird automatisch nachgebucht), 409 no_matching_offer (routing lockern),
   404 model_not_found (Marktliste prüfen), 429 rate_limited (später erneut).
Fasse am Ende zusammen, was eingerichtet ist und wo der Schlüssel liegt.

Performance per request

On every request you decide how strong the model has to be — via routing in the same OpenAI-compatible API. A level is a minimum score in OUR measurements per use case, best the highest-measured model; set your own threshold with "min_score": 0.85. The price is today's customer price (in the selected currency per 1M tokens in/out, including our margin) of the cheapest model that meets the level — works for taylor, james-direct and every market id.

Coding measured with HumanEval+ "category": "coding"

Solid ≥ 70 %

0.00 / 0.00

17 models qualify

"level": "solid"

Strong ≥ 80 %

0.00 / 0.00

16 models qualify

"level": "strong"

Top ≥ 88 %

0.00 / 0.00

12 models qualify

"level": "top"

Best top score

5.54 / 27.69

1 model qualifies

"level": "best"

Reasoning & knowledge measured with MMLU-Pro + GPQA Diamond "category": "reasoning"

Solid ≥ 55 %

0.00 / 0.00

17 models qualify

"level": "solid"

Strong ≥ 65 %

0.00 / 0.00

17 models qualify

"level": "strong"

Top ≥ 75 %

0.00 / 0.00

15 models qualify

"level": "top"

Best top score

1.11 / 5.54

1 model qualifies

"level": "best"

Math measured with MATH "category": "math"

Solid ≥ 60 %

0.00 / 0.00

20 models qualify

"level": "solid"

Strong ≥ 75 %

0.00 / 0.00

20 models qualify

"level": "strong"

Top ≥ 85 %

0.00 / 0.00

17 models qualify

"level": "top"

Best top score

11.08 / 55.39

1 model qualifies

"level": "best"

Instruction following measured with IFEval "category": "instruction_following"

Solid ≥ 75 %

0.00 / 0.00

17 models qualify

"level": "solid"

Strong ≥ 82 %

0.00 / 0.00

16 models qualify

"level": "strong"

Top ≥ 88 %

0.00 / 0.00

9 models qualify

"level": "top"

Best top score

1.11 / 5.54

1 model qualifies

"level": "best"

Get your key

Add a card once and you're set. Only Stripe sees the card number — we never do.

Sign up and pay — all via API

A program or agent opens the account, fetches the key and manages billing on its own. Money before key: no free credit, the key is created only after payment.

1 · ACCOUNT

Open an account

Choose billing — a level is optional, every request picks it via routing. The response contains checkout_url and status_url.

curl -X POST https://llm-broker.net/api/v1/broker/accounts \
  -H "Content-Type: application/json" \
  -d '{"billing":"topup",
       "amount":25}'

2 · CARD

Pay once

The card is entered once at Stripe — by you or by an agent with a browser. We never see the card number.

3 · KEY

Collect the key

202 until payment, then api_key (ast_sk_…) and base_url — exactly once.

curl https://llm-broker.net/api/v1/broker/accounts/$SESSION_ID

4 · BILLING

Self-service

Read balance and usage, change top-up or monthly limit, replace or remove the card.

GET|PATCH /api/v1/broker/billing
POST|DELETE /api/v1/broker/card

After that everything runs without humans: top-ups with the stored card, monthly billing, cancellation. Only the very first card entry still needs a browser at Stripe today. Details: Wiki: account & billing via API.

Which model do I get with taylor?

Without a level, the most capable one; with a level (per request or as the key's default), the cheapest of all models that reach it in our measurements. The table above shows the market; which model answers a single request is decided by the price at that moment. Want the strongest instead of the cheapest: "routing": {"category": "…", "level": "best"} — works for taylor, james-direct and every market id.

What does "measured" mean?

We test every model ourselves on open benchmarks (samples, identical settings) and re-measure regularly — all without thinking mode so the numbers are comparable; reasoning models often do better on hard tasks with thinking switched on. You control that per request with reasoning_effort (e.g. "none"). Third-party indices are used for pre-selection only.

And if no model meets the level?

Then the API answers with level_unavailable and charges nothing — never a silent fallback to a weaker model.

Are there image and speech models too?

Yes: image generation, speech and transcription from the market with the same key — market ids under /api/v1/broker/models?type=image|speech|transcription, plus our own models iris and quill.

How does billing work?

Pay as you go at the daily price — tokens, images, characters or minutes — by credit card only: prepaid credit, optionally topped up automatically. No subscription, no free credit.