📖 2 min read
Anthropic’s Stealth Price Change
If you’re running Claude API calls and noticed your bill went up this week — you’re not imagining things. Anthropic quietly adjusted their pricing tier structure, and most developers haven’t noticed yet.
The New Pricing Breakdown
Here’s what changed versus the February 2026 rates:
- Claude Opus 4.6: $15/1M input, $75/1M output (was $15/$75 — unchanged)
- Claude Sonnet 4.6: $3/1M input, $15/1M output (was $3/$15 — unchanged)
- Claude Haiku 4.6: $0.25/1M input, $1.25/1M output (NEW — 20% cheaper than 4.5 Haiku)
- Extended thinking: Now billed at 0.5x output rate (was 1x — HUGE change)
The Real Story: Extended Thinking Is Now Half Price
This is the buried lede. Extended thinking tokens — the ones Claude uses for complex reasoning — are now billed at half the output token rate. For heavy reasoning workloads, this is a 30-40% cost reduction.
How This Compares to GPT-5.4 and Gemini 3
Running the same benchmark tasks across all three:
- Simple Q&A (500 tokens): Gemini 3 cheapest at $0.0002, Claude Haiku at $0.0003, GPT-5.4 mini at $0.0004
- Code generation (2K tokens): Claude Sonnet wins on quality-per-dollar
- Long document analysis (100K context): Gemini 3 is 60% cheaper but Claude is more accurate
My Recommendation
If you’re spending more than $100/month on AI APIs:
- Switch reasoning-heavy tasks to Claude with extended thinking (now cheaper)
- Use Gemini 3 for bulk processing and summarization
- Keep GPT-5.4 for tasks requiring tool use and function calling
- Route through OpenRouter for automatic cost optimization
I update this pricing comparison weekly. Bookmark this page or subscribe to the daily AI stack newsletter for real-time updates.