Everyone talks about which AI model is best. Nobody talks about what it actually costs when you use four different providers daily for real work. I tracked every API call for 30 days, and the results changed how I think about model selection.
My total bill: $127 across all providers. That includes coding, content generation, data analysis, and experimentation. Here is the full breakdown.
My setup
I run a small SaaS product and a content site. My AI usage breaks down into:
- Coding: 40% of usage (code review, bug fixes, feature development)
- Content: 30% of usage (writing, editing, summarization)
- Data analysis: 20% of usage (log analysis, user behavior, metrics)
- Experimentation: 10% of usage (testing new models, comparing outputs)
I use multiple providers because different models excel at different tasks. I am not loyal to any single provider.
The 30-day breakdown
| Provider | Model | Spend | % of Total | Primary Use |
|---|---|---|---|---|
| Anthropic | Claude Sonnet 5 | $42 | 33% | Coding, content |
| OpenAI | GPT-5.6 Luna | $31 | 24% | Data analysis, bulk tasks |
| Gemini 3.6 Flash | $28 | 22% | Multimodal, speed tasks | |
| DeepSeek | V4 Pro | $18 | 14% | Budget coding, experimentation |
| Other | Various | $8 | 6% | Testing new models |
Total: $127/month
What surprised me
1. Luna is my workhorse, not the frontier models
I expected Claude Opus 5 or GPT-5.6 Sol to dominate my spending. Instead, GPT-5.6 Luna at $0.20/$1.20 handles 60% of my API calls. Most tasks do not need frontier intelligence. They need good-enough responses at the lowest cost.
2. Sonnet 5 is my “important work” model
When I need high-quality output (client-facing content, complex code review, architecture decisions), I switch to Sonnet 5 at $2/$10. It costs 10x more than Luna, but the quality difference matters for 20% of my work.
3. Gemini 3.6 Flash is my multimodal workhorse
Any task involving images, screenshots, or visual content goes to Gemini 3.6 Flash. At $1.50/$7.50 with built-in computer use, it is the only model that handles visual workflows natively.
4. DeepSeek is my experimentation sandbox
I use DeepSeek V4 Pro for testing ideas, running batch experiments, and tasks where I need open-source flexibility. At $2.19/$8.76, it is affordable for experimentation.
Cost per task type
| Task Type | Avg Tokens | Avg Cost | Model Used |
|---|---|---|---|
| Code review (1 file) | 3K in / 1K out | $0.006 | Luna |
| Blog post draft | 2K in / 3K out | $0.036 | Sonnet 5 |
| Data analysis query | 5K in / 1K out | $0.012 | Luna |
| Image description | 1K in / 500 out | $0.005 | Gemini Flash |
| Complex refactoring | 10K in / 5K out | $0.08 | Sonnet 5 |
| Bulk content generation | 1K in / 2K out | $0.004 | Luna |
The model routing strategy
I use a simple routing strategy based on task complexity:
Low complexity (use Luna):
- Simple code completions
- Data formatting and transformation
- Bulk content generation
- Log analysis and summarization
Medium complexity (use Sonnet 5 or Gemini Flash):
- Code review and refactoring
- Blog post drafts and editing
- Architecture discussions
- Multimodal tasks
High complexity (use Opus 5 or Sol):
- Complex debugging
- Client-facing content
- Strategic decisions
- Research and analysis
What I would change
If I wanted to cut costs to $50/month:
- Use Luna for 80% of tasks (up from 60%)
- Use Gemini 3.5 Flash-Lite ($0.30/$2.50) instead of 3.6 Flash
- Reserve Sonnet 5 for truly important work only
- Stop experimenting with new models
If I wanted the best quality regardless of cost:
- Use Opus 5 ($5/$25) for everything
- Monthly bill would be ~$400-500
- Quality improvement is marginal for most tasks
My take
The biggest insight from 30 days of tracking: most tasks do not need frontier intelligence. Luna at $0.20/$1.20 handles the majority of my work at 10-25x lower cost than frontier models. The key is knowing when to escalate.
For developers starting out, begin with Luna and Gemini 3.6 Flash. That covers 90% of use cases for under $30/month. Add Sonnet 5 for important work. Only use Opus 5 or Sol when the task genuinely justifies the cost.
The model routing strategy matters more than the model choice. Using the right model for the right task saves more money than picking the cheapest model for everything.
FAQ
How much should I budget for AI API costs?
For a solo developer, $50-100/month covers most use cases. For a small team (5 people), budget $200-500/month. For enterprise use, costs scale with usage but model routing can keep per-task costs under $0.10.
Is it worth using multiple providers?
Yes. Different models excel at different tasks. Luna for bulk work, Sonnet 5 for quality, Gemini for multimodal, DeepSeek for experimentation. The overhead of managing multiple APIs is worth the cost savings.
How do I track my AI API costs?
Most providers have dashboards (OpenAI, Anthropic, Google). For multi-provider tracking, use a tool like Helicone or LangSmith, or simply log costs in a spreadsheet.
Which model is the best value?
GPT-5.6 Luna at $0.20/$1.20 is the best value for most tasks. It scores 84.3% on Terminal-Bench, which is good enough for 80% of development work. Use Sonnet 5 or Opus 5 only when quality matters.
Should I use the cheapest model for everything?
No. Cheap models struggle with complex reasoning, nuanced writing, and multi-step problems. Use cheap models for bulk work and expensive models for important work. The cost difference between a $0.004 task and a $0.08 task is negligible if the expensive model produces better results.
Related: GPT-5.6 Luna Price Drop | AI API Pricing Compared | Claude Sonnet 5 Complete Guide | Best Flash AI Models Compared | Cost of Running AI Coding Agent 24/7