Harogo Docs

Introduction

Harogo is a multi-model subscription service for coding agents.

The discounted subscription price is $4.90/month for a limited time, then $9.90/month once the offer ends. 100 discounted subscription seats are currently open — remaining seats and the exact settlement price shown on the subscription page are authoritative.

After subscribing, you can access multiple coding models through a single entry point. All models share the same Harogo account, API key, and subscription quota — there's no need to register for or top up multiple model platforms separately.

Why Harogo

Different models excel at different coding tasks. Using several model providers directly usually means managing multiple accounts, API keys, balances, rate limits, and bills.

Harogo connects a set of models suited for coding agents behind a single service. You can switch models based on the task, and every call is billed against the same shared quota.

Current models fall roughly into these groups:

• Kimi K3, Claude Opus 5, GPT-5.6 Sol — suited for large codebases, long-context tasks, and high-difficulty reasoning.
• Qwen3.8 Max, GLM-5.2, Claude Sonnet 5, GPT-5.6 Terra, MiniMax M3, DeepSeek V4 Pro — suited for everyday agent coding, code generation, and debugging.
• DeepSeek V4 Flash, Claude Haiku 4.5, GPT-5.6 Luna — suited for high-frequency calls, fast responses, and cost-sensitive tasks.
• Claude Fable 5 — suited for longer-running, multi-step agent workflows.

These groupings are only a starting point for choosing a model. Actual results depend on the specific task, context, and how tools are used.

Getting Started

1. Log in to Harogo and complete your subscription.
2. Create and copy an API key in the console.
3. Set the Base URL to https://api.harogo.ai.
4. Paste the API key.
5. Start using it.

Supported Models

Currently supported models:

The names and model IDs above are the routing identifiers Harogo uses. The model list, versions, and availability may change — refer to what's actually shown in the console.

Usage & Limits

Harogo enforces the following usage limits:

All three limits are enforced at the same time. Every model call draws from the shared subscription quota, and some models have a lower maximum usage cap — see the pricing table below for details.

Limits are defined by dollar value, not a fixed number of requests. The actual number of requests you get depends on the model you use — lower-priced models generally allow more requests, and higher-priced models allow fewer.

You can check your current usage and remaining quota in the Harogo console.

Estimated Request Volume

The table below gives estimated request counts based on typical Harogo usage patterns:

Each row assumes the quota for that period is used mainly on the corresponding model. When mixing multiple models, they draw from the same shared quota together, so request counts across models can't simply be added up.

Actual request counts vary with context length, cache hit rate, reasoning effort, and output length. A coding agent may also issue multiple model requests for a single task, so one request does not equal one complete coding task.

Estimation Methodology

Estimates are based on the following typical request patterns:

• GLM-5.2 — about 700 input tokens, 52,000 cache-read tokens, and 150 output tokens per request.
• Qwen3.8 Max — about 8,731 input tokens, 31,831 cache-read tokens, and 2,029 output tokens per request.
• Kimi K3 — about 1,050 input tokens, 76,500 cache-read tokens, and 300 output tokens per request.
• MiniMax M3 — about 4,381 input tokens, 28,768 cache-read tokens, and 1,108 output tokens per request.
• DeepSeek V4 Pro — about 750 input tokens, 82,000 cache-read tokens, and 290 output tokens per request.
• DeepSeek V4 Flash — about 790 input tokens, 68,000 cache-read tokens, and 280 output tokens per request.
• Claude Opus 5 — about 580 input tokens, 386,230 cache-read tokens, 19,476 cache-write tokens, and 2,025 output tokens per request.
• Claude Sonnet 5 — about 159 input tokens, 129,804 cache-read tokens, 6,260 cache-write tokens, and 850 output tokens per request.
• Claude Fable 5 — about 399 input tokens, 315,393 cache-read tokens, 12,907 cache-write tokens, and 1,016 output tokens per request.
• Claude Haiku 4.5 — about 700 input tokens, 52,000 cache-read tokens, and 150 output tokens per request.
• GPT-5.6 Luna — about 45,726 input tokens, 10,720 cache-read tokens, and 1,905 output tokens per request.
• GPT-5.6 Sol — about 6,032 input tokens, 112,138 cache-read tokens, and 678 output tokens per request.
• GPT-5.6 Terra — about 6,825 input tokens, 69,476 cache-read tokens, and 734 output tokens per request.

Qwen3.8 Max, MiniMax M3, and the Claude and GPT-5.6 families use observed average request structures. The remaining models use a typical coding-request structure. Claude Opus 5 currently references actual usage data from Opus-tier models at the same price point.

These figures are used to estimate typical usage and don't constitute a commitment to a fixed number of requests.

Model Pricing & Usage Caps

The table below lists each model's API reference price per 1 million tokens, along with the usage cap Harogo applies to that model.

"—" means the item isn't billed separately at this time, or doesn't apply to that model.

Estimated request counts are calculated from a combination of model pricing, typical request structure, real usage data, and each model's usage cap. Preview model pricing and availability may change.

Model pricing, usage caps, and estimated request counts may change as the service evolves — refer to the Harogo console and the current version of this page for the latest details.