## Provider Tiers Overview

ModelStack categorizes all models into three pricing tiers. Understanding provider tiers — and which wallet pays for them — helps you choose the right model for your budget and use case.

## Tier System

Each tier has two separate rates: one for your **bonus wallet**, one for your **prepaid wallet**. These are genuinely different numbers, not two views on the same discount — see [How a Request Can Split Across Wallets](#how-a-request-can-split-across-wallets) below for why that matters.

| Tier         | Bonus Wallet Rate                              | Prepaid Wallet Rate                             | Wallet Eligibility             | Use Cases                                          |
| ------------ | ----------------------------------------------- | ------------------------------------------------ | ------------------------------- | --------------------------------------------------- |
| **Economy**  | 1.0x (no discount)                              | 0.2x by default (80% off provider cost)          | Bonus wallet or prepaid wallet  | Ultra high-volume, cost-sensitive batch work         |
| **Standard** | 1.5x (50% markup)                               | 0.5x by default (50% off provider cost)          | Bonus wallet or prepaid wallet  | Everyday production workloads, high-volume tasks     |
| **Premium**  | Not eligible — bonus can never pay Premium      | 1.0x (full provider cost, never discounted)      | Prepaid wallet only             | Complex reasoning, highest-quality output            |

<Note>
  Bonus credit is never discounted below face value, and on Standard it's actually
  worth *less* than face value — a 1.5x rate means $1 of bonus balance only covers
  about $0.67 of raw provider cost. Bonus exists to give you free usage, not a
  discount; the real discount lives on the prepaid side.
</Note>

A few model families get a deeper **prepaid** discount than the defaults above. These overrides only ever apply to the prepaid rate — bonus rates and Premium are never overridden:

- **Standard tier**: Anthropic models default to 0.35x, OpenAI models default to 0.2x, and two specific models (`gpt-5.4`, `gpt-5.6-terra`) are pinned to 0.2x directly.
- **Economy tier**: no overrides exist yet — every model currently uses the 0.2x default.

## How a Request Can Split Across Wallets

Because bonus and prepaid rate the same tier differently, a single Economy or Standard request can be billed partly from each wallet — at each wallet's own rate — within the same transaction.

Here's the mechanism: the bonus wallet covers as much of the request's raw provider cost as your available bonus balance can afford **at bonus's rate**. The prepaid wallet covers whatever raw cost is left over, **at prepaid's rate**. The two portions add up to the total charge, so the "discount" you see on that transaction is a blended, effective rate — not simply one of the two table values above.

**Worked example — Standard tier, $100 raw provider cost, $30 available bonus balance, prepaid funded:**

1. Bonus can afford $30 ÷ 1.5x = $20 worth of raw cost, and is charged the full $30 for that slice.
2. The remaining $80 of raw cost goes to prepaid at 0.5x = $40.
3. **Total charged: $70** for $100 of raw cost — a blended 0.7x effective rate, even though neither wallet's own rate is 0.7x.

If that same user had no bonus balance at all, the full $100 would go to prepaid at 0.5x = **$50** — the tier's plain default rate, with no blending.

<Tip>
  The "discount rate" shown on a transaction in your dashboard is this blended
  rate, not a fixed per-tier number. Two transactions on the same model and
  tier can show different rates depending on how much bonus balance was
  available at the time each one ran.
</Tip>

<Info>
  If you've never funded your prepaid balance (it's at or below $0), Economy
  and Standard requests are billed **entirely from bonus**, at bonus's rate —
  even the portion that would normally go to prepaid. This is intentional: an
  unfunded free account should never accidentally rack up real-money debt.
  Bonus balance can dip slightly negative to absorb this (see the overdraft
  note in [Balance & Wallets](/docs/billing/balance)), but it never spills
  over into an actual prepaid charge.
</Info>

## Providers by Tier

### Anthropic

| Model               | Available Tiers               | Best For                                             |
| ------------------- | ------------------------------ | ----------------------------------------------------- |
| `claude-opus-4-6`   | Premium only                    | Complex reasoning, code generation, creative writing  |
| `claude-sonnet-4-6` | Economy + Standard + Premium    | Balanced performance and cost, production workloads   |
| `claude-sonnet-5`   | Economy + Standard + Premium    | Advanced coding, complex reasoning, production use     |
| `claude-haiku-4-5`  | Economy + Standard + Premium    | Fast responses, high-volume, simple tasks              |

**Key Features:**

- Extended context windows (up to 1M tokens on newer models)
- Strong reasoning and instruction following
- Excellent code generation
- Tool use and function calling
- Vision capabilities (Opus/Sonnet)

### OpenAI

| Model          | Available Tiers               | Best For                                      |
| -------------- | ------------------------------ | ----------------------------------------------- |
| `gpt-5.5`      | Premium only                    | Advanced reasoning, complex problem solving     |
| `gpt-5.4`      | Economy + Standard + Premium    | Production applications, balanced performance   |
| `gpt-4.1-mini` | Economy + Standard + Premium    | Fast responses, lower cost                       |
| `gpt-4o-mini`  | Economy + Standard + Premium    | High-volume, simple completions                  |

**Key Features:**

- Strong general knowledge
- Fast response times
- Excellent at following instructions
- Function calling and tools
- Vision capabilities (GPT-4o family)
- Reasoning tokens (o1 models)

### Google (Gemini)

| Model                   | Available Tiers               | Best For                            |
| ----------------------- | ------------------------------ | ------------------------------------ |
| `gemini-3.1-pro`        | Economy + Standard + Premium    | Advanced reasoning, long context     |
| `gemini-2.5-pro`        | Economy + Standard + Premium    | Production workloads, balanced cost  |
| `gemini-2.5-flash`      | Economy + Standard + Premium    | Fast responses, cost-effective       |
| `gemini-3-flash`        | Economy + Standard + Premium    | High-volume, simple tasks            |
| `gemini-3.1-flash-lite` | Economy + Standard + Premium    | Ultra-fast, lowest cost              |
| `gemma-4-31b-it`        | Economy + Standard + Premium    | Open weights, privacy-focused        |

**Key Features:**

- Massive context windows (up to 2M tokens)
- Native multimodal (text, image, audio, video)
- Strong reasoning and analysis
- Fast inference (Flash models)
- Open weights available (Gemma)

### DeepSeek

| Model               | Available Tiers               | Best For                         |
| ------------------- | ------------------------------ | ----------------------------------- |
| `deepseek-v4-pro`   | Economy + Standard + Premium    | Advanced reasoning, complex code  |
| `deepseek-v3.2`     | Economy + Standard + Premium    | Cost-effective reasoning          |
| `deepseek-v4-flash` | Economy + Standard + Premium    | Ultra-cheap, high-volume          |

**Key Features:**

- Exceptional value for reasoning
- Strong code understanding
- Chinese language support
- Very competitive pricing

### MiniMax

| Model          | Available Tiers               | Best For                      |
| -------------- | ------------------------------ | -------------------------------- |
| `minimax-m2.7` | Economy + Standard + Premium    | General tasks, cost-effective  |
| `minimax-m2.5` | Economy + Standard + Premium    | Conversational AI               |
| `minimax-m2.1` | Economy + Standard + Premium    | High-volume applications        |

**Key Features:**

- Chinese language excellence
- Conversational capabilities
- Competitive pricing
- Low latency in Asia

### ZAI (GLM)

| Model     | Available Tiers               | Best For                   |
| --------- | ------------------------------ | ----------------------------- |
| `glm-5`   | Economy + Standard + Premium    | General-purpose tasks        |
| `glm-4.7` | Economy + Standard + Premium    | Cost-effective completions   |

**Key Features:**

- Chinese language support
- Fast inference
- Regional availability (Asia)

### Qwen

| Model               | Available Tiers               | Best For              |
| ------------------- | ------------------------------ | ------------------------ |
| `qwen3-coder-next`  | Economy + Standard + Premium    | Code generation         |
| `qwen3-coder-flash` | Economy + Standard + Premium    | Fast code completions   |

**Key Features:**

- Specialized for code
- Multi-language support
- Fast generation

## Choosing the Right Tier

### Use Economy When:

- You're running **ultra high-volume, cost-sensitive workloads** — batch jobs, background pipelines, bulk classification
- Raw quality bar is low and cost per request matters more than anything else
- You're paying from **prepaid** — Economy's 0.2x default prepaid rate is the cheapest way to run a model, deeper than Standard's discount
- You're paying from **bonus** and want to stretch it as far as possible — Economy is the only tier where bonus buys full face value (1.0x, no markup)

**Examples:**

- Bulk data labeling and tagging
- Large-scale embeddings-adjacent preprocessing
- Background summarization jobs with no latency/quality pressure

### Use Standard When:

- You need **balanced performance and cost**
- Workloads range from **development to high-volume production**
- Tasks require **good reasoning** but not exceptional
- You want **reliable, consistent quality** at a discount
- You're paying from **prepaid** — Standard's default 0.5x rate (less for some model families) beats Premium's full price

**Examples:**

- Code autocomplete suggestions
- Simple text classification and data extraction
- Chat responses for FAQs and customer support
- Content generation and document summarization
- Code review and general-purpose AI features

### Use Premium When:

- Tasks require **complex reasoning**
- **Quality is critical** over cost
- You need **best-in-class performance**
- **Low volume**, high-impact requests
- **Difficult problems** that simpler models struggle with

**Examples:**

- Legal document analysis
- Complex code generation
- Research and analysis
- Creative writing (novels, scripts)
- Multi-step problem solving
- Advanced reasoning tasks

## Provider-Specific Considerations

### Context Window Sizes

| Provider | Max Context   | Notes                                             |
| -------- | ------------- | -------------------------------------------------- |
| Gemini   | 2M tokens     | Largest available, perfect for long documents      |
| Claude   | 200K tokens   | Extended context, strong at utilizing full window  |
| OpenAI   | 128K tokens   | Standard long context                              |
| DeepSeek | 64K tokens    | Moderate context                                    |
| Others   | 8K-32K tokens | Varies by model                                    |

### Multimodal Capabilities

**Vision (Image Understanding):**

- ✅ Claude Opus/Sonnet
- ✅ GPT-4o models
- ✅ Gemini all models

**Tool Use/Function Calling:**

- ✅ All Anthropic models
- ✅ All OpenAI GPT models
- ✅ Gemini Pro models

### Reasoning Tokens

Some models expose internal reasoning:

- **o1 models** (OpenAI): Reasoning tokens separate from output
- **Claude Opus**: Extended thinking mode
- **DeepSeek v4**: Internal reasoning chain

<Warning>
  Reasoning tokens are charged at the same rate as input tokens but can
  significantly increase costs. Monitor usage carefully when using
  reasoning-heavy models.
</Warning>

## Cost Optimization Tips

### 1. Start Economy or Standard, Upgrade as Needed

Begin with Economy or Standard tier models for testing and development. Economy is cheapest on prepaid (0.2x default) but carries no bonus discount (1.0x); Standard is a smaller prepaid discount (0.5x default) but still beats Premium. Upgrade to Premium only when quality requirements justify paying full price.

### 2. Mix and Match Tiers

Use different tiers for different tasks:

- **Economy**: Bulk, low-stakes, background work where cost per request matters most
- **Standard**: Autocomplete, drafts, customer-facing generation, most production traffic
- **Premium**: Complex analysis, critical decisions

### 3. Leverage Model Stacks

Create model stacks with:

1. Economy or Standard model as primary (discounted on prepaid, covers most tasks)
2. Premium model as fallback (guaranteed quality when the cheaper tier isn't enough)

### 4. Monitor Provider Performance

Track which providers work best for your use case:

- **Response time**: Some providers are faster in certain regions
- **Quality**: Test different providers for your specific task
- **Reliability**: Monitor error rates and availability

### 5. Use Premium Deliberately

Premium requests are billed from your **prepaid wallet only**, at full price (1.0x — no discount). Bonus balance can never pay for Premium, even if you have plenty of it. Your dashboard shows a monthly premium budget figure for cost awareness, but it's informational only — it doesn't gate or block Premium usage. Reserve Premium for requests that genuinely need it, since it's always more expensive than Economy or Standard's discounted prepaid rates.

## Regional Considerations

### North America & Europe

**Best latency:** OpenAI, Anthropic, Google
**Recommended:** Claude Sonnet, GPT-4o, Gemini Pro

### Asia-Pacific

**Best latency:** DeepSeek, MiniMax, Qwen, GLM
**Recommended:** DeepSeek v4-flash, MiniMax models

### Chinese Language

**Best support:** DeepSeek, MiniMax, GLM, Qwen
**Recommended:** GLM-5, DeepSeek v4-pro for complex tasks

## Provider Reliability

ModelStack automatically handles provider failures through model stacks:

- **Automatic fallback**: If primary provider fails, try next model
- **Health monitoring**: Real-time provider status tracking
- **Retry logic**: Automatic retries with exponential backoff

Check current provider status in your [Dashboard](https://modelstack.cc/dashboard/analytics).

## Next Steps

- [View all supported models](/models/supported-models)
- [Create model stacks](/dashboard/model-stacks) for automatic fallback
- [Check pricing](/billing/model-pricing) for detailed cost comparison
- [Monitor usage](/dashboard/analytics) by provider and tier