Wire flash
User test: GPT-6 Luna costs 40% less than Haiku 5.5 on summarization
Editorial responsibility
- No named human review is recorded for this page.
- Source reporting is collected, normalized, translated or condensed automatically when needed.
- Automatically published source-backed update
An X user reports that despite listing identical prices of $0.10 per million input tokens and $0.50 per million output tokens, GPT-6 Luna cost 40% less than Haiku 5.5 on a summarization job involving 20 podcast excerpts. Luna was cheaper on all 20 requests. The cost difference is attributed to token counts: for the same text, Haiku counted 119,723 input tokens versus Luna's 87,055, and Haiku charged for 2.1 times Luna's output tokens, mostly due to hidden thinking tokens, while visible words were only 1.28 times more. The user notes that the price list does not reflect these differences, and that the cost gap varies by task, being 12x on code generation and 40% on summarization.
Source report
Base Pricing
Both models list the same price: $0.10 per 1M input tokens and $0.50 per 1M output tokens.
Performance on Podcast Summarization
I ran 20 podcast excerpts through both models. Key findings:
- GPT-6 Luna cost 40% less than Haiku 5.5 on this job
- Luna was cheaper on 20 out of 20 requests
Task-Dependent Cost Variation
The cost gap varies significantly by task type:
| Task | Cost Difference | |------|----------------| | Code generation | 12x gap | | Summarization | 40% gap |
The cost difference is determined by the job type, driven by token counts for the same text, plus default "thinking" charged as output tokens.
Token Count Discrepancy
For the same input text:
| Metric | Haiku 5.5 | GPT-6 Luna | |--------|-----------|------------| | Input tokens | 119,723 | 87,055 | | Output token cost | 2.1x Luna's | Baseline | | Visible words | 1.28x Luna's | Baseline |
Haiku charged for 2.1x Luna's output tokens, mostly due to hidden "thinking" tokens, while visible words were only 1.28x more.
Conclusion
The published price list reveals none of these real-world cost differences. Actual costs depend heavily on task type, token counting methods, and hidden processing (thinking) tokens.
Source
aakashguptaNeutral / independent