Claude 4.5 Haiku vs Gemini 3.1 Pro: pricing & cost comparison
On input tokens, Claude 4.5 Haiku is the cheaper of the two — 50% less per million ($1 vs $2). On output, Claude 4.5 Haiku is 58% cheaper ($5 vs $12) — and since output is usually the dominant cost driver, that gap matters more than it looks.
Side by side
| Claude 4.5 Haiku | Gemini 3.1 Pro | |
|---|---|---|
| Input / 1M tokens | $1 | $2 |
| Output / 1M tokens | $5 | $12 |
| Context window | 200,000 | 1,000,000 |
| Token-count accuracy | ±2% | ±3% |
| Cost — 10,000 input + 2,000 output tokens | $0.02 | $0.044 |
What a real request costs
Take a representative turn — 10,000 input + 2,000 output tokens. Claude 4.5 Haiku comes to $0.02, Gemini 3.1 Pro to $0.044. Across 100,000 requests that's a $2400 swing in favour of Claude 4.5 Haiku. To run the numbers on your actual prompt, paste it into the calculator and toggle Compare across all models.
Close call — decide on output cost
Claude 4.5 Haiku and Gemini 3.1 Pro are close enough on price that the tie-breakers matter more than the headline rate. Output tokens are usually the dominant cost driver, so weigh $5 vs $12 per 1M output first, then context window — Gemini 3.1 Pro carries the larger one at 1,000,000 tokens. They're different vendors, so factor in a small tokenizer calibration buffer on whichever side isn't exact.
See the full breakdown on the dedicated pages for Claude 4.5 Haiku and Gemini 3.1 Pro.
FAQ
- Is Claude 4.5 Haiku or Gemini 3.1 Pro cheaper?
- For a typical request (10,000 input + 2,000 output tokens), Claude 4.5 Haiku is cheaper — about 55% less, or roughly $2400 saved per 100,000 requests. Claude 4.5 Haiku runs $1/$5 per 1M input/output tokens; Gemini 3.1 Pro runs $2/$12.
- Which has the larger context window?
- Gemini 3.1 Pro, at 1,000,000 tokens versus 200,000.
- How accurate are these token counts?
- Claude 4.5 Haiku: Approximated with cl100k_base — drift typically <2% on English and code. Gemini 3.1 Pro: Approximated with o200k_base; drift typically ~3% on English and code. The dollar math itself is exact once the token count is known.
Both prices are computed from tokenmath's verified pricing table. Rates sourced from www.anthropic.com and ai.google.dev, verified 2026-07-06. Vendor pricing changes often — confirm before you commit.