Hidden reasoning costs
OpenAI’s o-series models charge for internal reasoning tokens at the standard output rate. This hidden computation means a model might generate thousands of reasoning tokens for a single request, yet the final response shows only a few words. For o3, input tokens cost $2.00 per million and output tokens cost $8.00 per million. For the higher tier o3-pro, input costs $20.00 per million and output reaches $80.00 per million. For o4-mini, the input cost is $1.10 per million and output is $4.40 per million. The effective cost for a complex task runs 3x to 10x the base rate because these internal traces consume significant GPU time and power. Users see only the final answer, but the bill reflects every hidden thought the model produced to reach that conclusion. A single thoughtful answer can burn thousands of extra forward passes. This mechanism makes reasoning models expensive for tasks where the logic is trivial. The o1 models use reasoning tokens to break down understanding of the prompt, consider multiple approaches, and work through problems step-by-step. These tokens also count as part of the context window usage, which can impede performance. Because reasoning tokens are billed as output, they make up a significant portion of the total cost for complex logic tasks. The internal reasoning process involves breaking down understanding and considering multiple approaches to generating a response.
DeepSeek savings and model tiers
DeepSeek R1 provides massive savings when compared to OpenAI’s o1 pricing. While o1 tasks cost $15.00 for input and $60.00 for output per million tokens, DeepSeek R1 costs as low as $0.55 for input and $2.19 for output. This price gap allows enterprises to process large datasets without the massive expense of frontier reasoning models. OpenAI’s own hierarchy creates a 100-fold cost spread between its lowest and highest models. GPT-6 Luna costs $0.20 for input and $1.20 for output per million tokens, whereas the flagship GPT-6 Astra charges $10.00 for input and $50.00 for output. You should evaluate your workload before selecting a model tier.
| Model | Input $/1M | Output $/1M |
| o4 Mini | $1.10 | $4.40 |
| o3 | $2.00 | $8.00 |
| o3-pro | $20.00 | $80.00 |
| GPT-6 Luna | $0.20 | $1.20 |
| GPT-6 Astra | $10.00 | $50.00 |
| DeepSeek R1 | $0.55 | $2.19 |
For developers using the GPT-5.6 family, output costs are exactly six times the input cost, so long responses dominate most bills, which means an assistant that rambles costs far more than one that answers tightly. The GPT-5.6 Terra tier costs $2.00 for input and $12.00 for output, while the GPT-5.6 Sol tier costs $5.00 for input and $30.00 for output. On July 30, 2026, OpenAI cut GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20%. If you run 10,000 summaries a month using GPT-6 Astra, the cost is $155, but using GPT-6 Luna reduces that monthly expense to $1.55. Model selection is the primary cost driver.
Context thresholds and competitive pricing
OpenAI applies a long-context surcharge for requests exceeding 272,000 input tokens. When users pass prompts beyond this limit, OpenAI reprices the entire request, which doubles the input rate and increases the output rate by 50% for every token in that call. This pricing structure differs from Anthropic, where Claude models from version 4.6 onward include their full 1M context window at standard rates. At the flagship tier, OpenAI and Anthropic charge the same, but Google charges roughly a fifth to a quarter as much. Specifically, a 1,000-token reply costs five cents on GPT-6 Astra, but the same reply on Gemini 3.1 Pro costs about 1.2 cents.
How do you verify that reasoning tokens do not inflate your bill?
Users often encounter difficulties when trying to interpret their usage because OpenAI usage graphs by model still only show input tokens. You must apply the model costs to the token count yourself to see the true output consumption. For the GPT-5.6 family, the context window is 1.05M tokens across all three tiers. If you use the Batch API, OpenAI halves both input and output rates for asynchronous jobs. The GPT-6 Astra model was released on September 3, 2026. For embeddings, text-embedding-3-small costs $0.02 per million tokens and text-embedding-3-large costs $0.13. Whisper API pricing bills at $0.006 per minute for whisper-1. For classification, a label is about 20 output tokens, so the bill is mostly input. On the earlier GPT-6 Sol, the same batch of 1,000 classifications with 400 input and 20 output tokens cost $1.00, while Astra’s version costs $5.00. For the GPT-6 Astra model, any prompt past 272,000 input tokens reprices the entire request, instead of only the overflow.
