Reasoning Model Cost Guide
Updated 12 July 2026 · first published 4 July 2026
Reasoning models like o3, o4-mini, DeepSeek-R1, QwQ, and MiMo generate internal "thinking" tokens before producing output. This hidden cost can inflate your bill 5-30x. This guide breaks down when reasoning models are worth the premium and when a standard model is the smarter spend.
- All prices in USD per 1 million tokens
- Reasoning token counts vary by task complexity
- Thinking tokens are billed as output tokens
What are "thinking" tokens?
Reasoning models spend tokens on internal chain-of-thought before answering. A simple question might trigger 2,000 thinking tokens; a complex coding task can produce 30,000+. These tokens are invisible in the API response but billed at output rates - often the most expensive tier.
Methodology
Standard pricing
Output token price for normal (non-reasoning) completion, from official provider pages.
Reasoning pricing
Effective cost including thinking tokens. Measured as standard output price × average thinking multiplier for each model.
Thinking multiplier
Average ratio of total billed tokens to visible output tokens across coding, math, and analysis benchmarks.
Best use cases
Tasks where reasoning models demonstrably outperform standard models by enough to justify the cost premium.
Reasoning Model Pricing Table
| Model ↕ | Provider ↕ | Standard Output $/1M ↕ | Reasoning Output $/1M ↕ | Multiplier ↕ | Best Use Cases |
|---|---|---|---|---|---|
| Claude Fable 5 | Anthropic | $50.00 | $200.00 | 4x | Frontier reasoning, complex software vulnerabilities, deep analysis |
| Claude Opus 4.8 | Anthropic | $25.00 | $100.00 | 4x | Frontier reasoning, complex analysis, agentic workflows |
| GPT-5.6-Sol | OpenAI | $30.00 | $120.00 | 4x | Frontier reasoning, complex code, math, multi-step analysis |
| GPT-5.6-Terra | OpenAI | $15.00 | $60.00 | 4x | Reasoning, code generation, structured analysis |
| Grok 4.5 | xAI | $6.00 | $24.00 | 4x | Math, reasoning, general intelligence |
| o3 | OpenAI | $8.00 | $40.00 | 5x | Complex coding, math proofs, deep reasoning |
| o4-mini | OpenAI | $4.40 | $17.60 | 4x | Cost-sensitive reasoning, coding, analysis |
| DeepSeek R1 | DeepSeek | $2.19 | $8.76 | 4x | Math, code, reasoning, Chinese-language tasks |
| Claude Sonnet 5 | Anthropic | $10.00 | $40.00 | 4x | General reasoning, coding, analysis (balanced) |
| GPT-5.5 (standard) | OpenAI | $30.00 | $30.00 | 1x | General chat, coding, complex analysis |
Reasoning Output $/1M includes thinking tokens billed at output rates. Multiplier is average across typical tasks; actual multiplier varies by prompt complexity.
DeepSeek R1 and QwQ output thinking tokens by default. OpenAI o3/o4-mini charge thinking tokens separately from visible output.
Gemini 2.5 Pro charges "thinking tokens" at a reduced rate vs visible output - multiplier is effective blended cost.
Some links may include referral codes.
When to Use Reasoning Models
1) Multi-step coding and debugging
Reasoning models excel at tasks requiring planning: refactoring across multiple files, debugging complex state, or writing algorithms with edge cases. The thinking tokens let the model "work through" the problem rather than pattern-match. Worth the 4-5x premium when the alternative is a human spending 30+ minutes.
2) Mathematical and logical proofs
Tasks requiring formal reasoning - proofs, optimisation problems, constraint satisfaction - see the biggest quality jump. Standard models often produce plausible-but-wrong answers; reasoning models self-verify. The cost premium is justified when correctness matters more than speed.
3) Complex data analysis with multiple steps
Analysing a dataset with conditional logic, cross-referencing multiple sources, or building multi-step pipelines benefits from chain-of-thought. A standard model might miss a step; a reasoning model works through each explicitly.
When to Skip Reasoning Models
1) Simple extraction and formatting
Pulling data from documents, reformatting JSON, summarising text, or translating content. These tasks need pattern matching, not reasoning. Using o3 for extraction is like hiring a mathematician to sort mail - expensive and unnecessary.
2) High-volume, low-complexity workloads
Content generation, FAQ answers, basic chatbot responses, and data labelling. At 10M+ tokens/month, the 4-5x reasoning multiplier adds up fast. A standard model at $10/1M output beats a reasoning model at $40-60/1M for tasks where quality is "good enough."
3) Latency-sensitive applications
Reasoning models take longer - thinking tokens add 2-10 seconds of latency. For real-time chat, interactive tools, or user-facing responses where speed matters, standard models deliver faster at lower cost.
Cost Optimisation Strategies
Route by complexity
Use a cheap standard model (GPT-4o-mini, Gemini Flash) as a router. If the task is simple, answer directly. If complex, escalate to a reasoning model. This "cascade" pattern can cut reasoning-model usage by 60-80% while preserving quality where it matters.
Limit thinking budget
Some providers (OpenAI, Google) let you set a maximum thinking token budget per request. Capping at 4,000-8,000 thinking tokens prevents runaway costs on tasks that don't need deep reasoning. Start tight and increase only when quality drops.
Cache your prompts
If your reasoning tasks share a common system prompt or context (codebase, documentation), prompt caching can reduce input costs by 50-90%. This doesn't reduce thinking tokens but lowers the overall per-request cost significantly.
Batch non-urgent reasoning
For evaluations, test generation, or analysis that tolerates latency, batch APIs offer 50% discounts. Combined with reasoning models, you get the quality benefit at half the output cost.
Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →