AI Cost Calc

Claude vs GPT vs Gemini: API Pricing Compared (2026)

2026-07-05

Comparing LLM providers on price is harder than it looks: each one meters input, output and cache differently, and a "cheap" model can lose on a real workload. Here we run one identical workload through the current price lists (verified 2026-07-05 on each provider's official page) and look at where each provider actually wins.

The test workload

A typical assistant feature: 150,000 requests/month, ~400 input tokens and ~300 output tokens per request → 60M input + 45M output tokens monthly. No caching, no extras. Estimated token bill per model:

  • Gemini 2.5 Flash ($0.30 in / $2.50 out): ~$131/month
  • GPT-5.4 mini ($0.75 / $4.50): ~$248/month
  • Claude Haiku 4.5 ($1 / $5): ~$285/month
  • Claude Sonnet 5 ($2 / $10, introductory through 2026-08-31): ~$570/month
  • Gemini 3.1 Pro Preview ($2 / $12): ~$660/month
  • GPT-5.4 ($2.50 / $15): ~$825/month

Two orders of magnitude separate the cheapest and priciest tiers across providers — which is why model choice, not provider loyalty, is the biggest cost decision. Run your own workload in the calculator to see your ranking.

Update (2026-07-09): OpenAI's new GPT-5.6 family reshuffles this table: Luna ($1 / $6) lands at ~$330/month on this workload — between Claude Haiku 4.5 and Claude Sonnet 5 — while Terra matches GPT-5.4's price with stronger claimed performance. See the fresh head-to-heads: Terra vs Sonnet 5 and Luna vs Haiku 4.5.

Caching changes the ranking

If your requests share a long stable prefix (system prompt, tool definitions, knowledge snippets), cached-input pricing matters as much as the sticker price:

  • OpenAI: cached input at ~10% of the input rate (e.g. GPT-5.4: $0.25 vs $2.50), applied automatically to repeated prefixes.
  • Anthropic: cache reads at 10% of input, but cache writes cost 1.25× input — caching pays off after a single reuse within the 5-minute window.
  • Google Gemini: explicit context caching at roughly 10% of input, with storage-time billing on some tiers.

An input-heavy RAG app with a 60–80% cache-hit rate can flip the comparison table entirely. Set the cache slider in the calculator to your expected hit rate and watch the ranking change.

When each provider makes sense

  • Claude (Anthropic) — long-context work, agent workflows, careful writing; Sonnet 5's introductory pricing (until 2026-08-31) makes it unusually competitive in the mid tier.
  • GPT (OpenAI) — broadest ecosystem and tooling; the mini tier is a strong price/quality default for product features.
  • Gemini (Google) — the Flash tiers are among the cheapest capable models anywhere, and integrate naturally with a Google Cloud stack.

The honest answer

There is no universally cheapest provider — there's a cheapest provider for your token shape (input-heavy vs output-heavy, cacheable vs not). Model it in two minutes with the AI API Cost Calculator, and re-check quarterly: providers reprice aggressively, and Claude Sonnet 5's intro pricing already has an expiry date. Estimates are for planning only — always verify current prices on each provider's official page.