The $0.28 Question: How DeepSeek Broke AI Pricing
In January 2025, DeepSeek released R1—a reasoning model that matched OpenAI's o1 on benchmarks. The price? 27x cheaper than the competition.
We track pricing across 80 AI providers. Here's what DeepSeek did to the market.
The Numbers That Changed Everything
DeepSeek's official API pricing:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| DeepSeek Chat | $0.28 | $0.42 |
| DeepSeek Reasoner | $0.28 | $0.42 |
| Cached input | $0.028 | — |
For context, here's what the competition charges for comparable reasoning:
| Model | Input | Output | vs DeepSeek |
|---|---|---|---|
| OpenAI o1 | $15.00 | $60.00 | 143x more |
| OpenAI o3 | $2.00 | $8.00 | 19x more |
| OpenAI o3-mini | $1.10 | $4.40 | 10x more |
DeepSeek isn't just cheaper. It's a different category.
How Providers Responded
We track DeepSeek models across 25+ providers. The pricing varies wildly:
Budget Tier (Under $1/1M input)
| Provider | Model | Input | Output |
|---|---|---|---|
| DeepSeek (direct) | V3/Reasoner | $0.28 | $0.42 |
| Alibaba Cloud | DeepSeek V3 | $0.29 | $1.15 |
| Chutes | DeepSeek V3.2 | $0.25 | $0.38 |
| Baseten | DeepSeek V3.2 | $0.30 | $0.45 |
| AIHubMix | DeepSeek V3.2 | $0.30 | $0.45 |
Mid Tier ($1-2/1M input)
| Provider | Model | Input | Output |
|---|---|---|---|
| Azure OpenAI | DeepSeek V3.1 | $0.56 | $1.68 |
| Amazon Bedrock | DeepSeek V3.1 | $0.58 | $1.68 |
| Azure OpenAI | DeepSeek R1 | $1.35 | $5.40 |
| Amazon Bedrock | DeepSeek R1 | $1.35 | $5.40 |
Enterprise Tier ($3+/1M input)
| Provider | Model | Input | Output |
|---|---|---|---|
| Abacus | DeepSeek R1 | $3.00 | $7.00 |
The pattern: enterprise cloud providers (AWS, Azure) charge 2-5x the direct API price. Budget-focused providers match or undercut DeepSeek's rates.
Why DeepSeek Is So Cheap
Three technical decisions enabled this pricing:
1. Mixture of Experts (MoE) Architecture
DeepSeek V3 has 671B total parameters but only activates ~37B per inference. You pay for the compute you use, not the full model.
2. Multi-Token Prediction
Instead of predicting one token at a time, DeepSeek predicts multiple tokens simultaneously. This dramatically reduces the number of forward passes needed.
3. Aggressive Caching
Their cached input rate is $0.028/1M tokens—90% cheaper than uncached. Build your prompts right and costs drop to nearly nothing.
The Price War Effect
DeepSeek's pricing forced moves across the industry:
OpenAI's response:
- Released o3-mini at $1.10 input (down from o1's $15)
- Launched GPT-5-nano at $0.05 input—their cheapest model ever
Google's response:
- Gemini 2.0 Flash-Lite at $0.075 input
- Gemini 1.5 Flash-8B at $0.0375—even cheaper than DeepSeek
Anthropic's response:
- Claude 3 Haiku held at $0.25 input
- Focus shifted to enterprise features over price competition
What This Means For You
If you're building an MVP
Use DeepSeek direct. At $0.28/1M tokens, you can prototype for pennies. The Reasoner model handles complex logic at the same price as chat.
If you're running production workloads
Consider the hosted options on Bedrock or Azure. The 2-5x markup buys you SLAs, compliance, and enterprise support. Still dramatically cheaper than OpenAI.
If you need reasoning capabilities
DeepSeek Reasoner at $0.28 vs OpenAI o1 at $15 is not a typo. Test whether DeepSeek meets your quality bar—for most use cases, it will.
The Ralph Wiggum Loop
Here's the real unlock: you don't have to choose one model.
The smartest teams are running what's being called the "Ralph Wiggum Loop"—use a frontier model for planning, then let a cheap model do the work.
GPT-5.2 ($1.75 input) → generates detailed plan
↓
DeepSeek V3 ($0.28 input) → executes each step
Why this works
Frontier models like GPT-5.2 excel at:
- Breaking down ambiguous tasks
- Identifying edge cases
- Structuring complex workflows
- Making architectural decisions
But once you have a clear, specific plan? DeepSeek executes it just as well at 1/6th the cost.
The math
Traditional approach (all GPT-5.2):
- Planning: 2K tokens → $0.0035
- Execution: 50K tokens → $0.0875
- Total: $0.091 per task
Ralph Wiggum Loop:
- Planning with GPT-5.2: 2K tokens → $0.0035
- Execution with DeepSeek: 50K tokens → $0.014
- Total: $0.018 per task
That's 5x cheaper with identical output quality for well-defined tasks.
When to use it
The loop works best for:
- Agentic workflows with multiple steps
- Code generation from specifications
- Data transformation pipelines
- Any task where "what to do" is harder than "doing it"
Skip the loop when:
- Tasks require continuous reasoning
- Context from planning is essential for execution
- You need the frontier model's specific capabilities (vision, tool use)
The Bottom Line
DeepSeek proved that frontier AI capabilities don't require frontier pricing. Their models run on the same architecture principles as competitors but cost a fraction as much.
The industry noticed. Prices are falling across the board, and the models getting cheaper aren't getting worse—they're getting better.
We update pricing data across all 80 providers daily. Check current DeepSeek pricing →
Data sourced from Subquery's database of 2,000+ models across 80 providers. Prices accurate as of January 2026.