📖 3 min read
Pricing pages lie. Every AI company shows you the per-token rate but none of them tell you what it actually costs to get useful work done. So I ran 100 real-world tasks across GPT-5, Claude Opus 4.7, and Gemini 2.5 Ultra and tracked the total cost per completed task.
The cheapest model was not the one with the lowest token price.
The Test Methodology
100 tasks split across 5 categories (20 each):
- Code generation: Build functional components from specifications
- Long document analysis: Extract insights from 20-50 page documents
- Creative writing: Marketing copy, blog posts, email sequences
- Data transformation: Parse, clean, and restructure datasets
- Multi-step reasoning: Complex problem-solving requiring chains of logic
I measured: total tokens consumed (including retries), time to completion, accuracy of output (graded by a separate evaluator), and total API cost per successful task.
The Results
Code Generation (20 tasks)
| Model | Success Rate | Avg Cost/Task | Avg Time |
|---|---|---|---|
| Claude Opus 4.7 | 95% | $0.18 | 45s |
| GPT-5 | 90% | $0.31 | 38s |
| Gemini 2.5 Ultra | 85% | $0.24 | 52s |
Claude dominated code generation. Higher first-pass accuracy meant fewer retries, which drove the cost down despite a similar per-token rate. GPT-5 was faster but needed more correction cycles.
Long Document Analysis (20 tasks)
| Model | Success Rate | Avg Cost/Task | Avg Time |
|---|---|---|---|
| Gemini 2.5 Ultra | 95% | $0.08 | 30s |
| Claude Opus 4.7 | 90% | $0.22 | 55s |
| GPT-5 | 85% | $0.28 | 42s |
Gemini crushed this category. The 2M token context window meant zero chunking overhead, and Google priced input tokens aggressively. For pure document processing, Gemini is 2-3x cheaper than the competition.
Creative Writing (20 tasks)
| Model | Success Rate | Avg Cost/Task | Avg Time |
|---|---|---|---|
| Claude Opus 4.7 | 90% | $0.15 | 35s |
| GPT-5 | 90% | $0.19 | 28s |
| Gemini 2.5 Ultra | 80% | $0.12 | 40s |
Close race between Claude and GPT-5. Claude produced more natural, less “AI-sounding” copy. GPT-5 was faster. Gemini was cheapest but output quality was noticeably more generic.
Data Transformation (20 tasks)
| Model | Success Rate | Avg Cost/Task | Avg Time |
|---|---|---|---|
| GPT-5 | 95% | $0.14 | 25s |
| Claude Opus 4.7 | 90% | $0.16 | 32s |
| Gemini 2.5 Ultra | 90% | $0.10 | 35s |
GPT-5 edged ahead here. Structured output mode and function calling made data transformation tasks cleaner with fewer parsing errors.
Multi-Step Reasoning (20 tasks)
| Model | Success Rate | Avg Cost/Task | Avg Time |
|---|---|---|---|
| Claude Opus 4.7 | 90% | $0.42 | 90s |
| GPT-5 | 85% | $0.55 | 75s |
| Gemini 2.5 Ultra | 80% | $0.35 | 85s |
The most expensive category across the board. Extended thinking tokens add up fast. Claude had the highest accuracy, Gemini was cheapest, GPT-5 sat in the middle.
Total Cost Across All 100 Tasks
- Gemini 2.5 Ultra: $17.80 (cheapest overall)
- Claude Opus 4.7: $22.60 (best quality-adjusted value)
- GPT-5: $29.40 (most expensive when factoring retries)
The Real Takeaway
There is no single “best” model. The smart play in April 2026 is routing:
- Documents and analysis – Gemini 2.5 Ultra
- Code and reasoning – Claude Opus 4.7
- Data transformation – GPT-5
- Creative writing – Claude or GPT-5 depending on voice preference
Tools like OpenRouter and LiteLLM make multi-model routing dead simple. If you are still sending everything to one model, you are overpaying by 30-50%.
Get the Full Cost Comparison Spreadsheet
Download our updated April 2026 AI pricing tracker with all 100 task results, cost breakdowns, and routing recommendations.