Claude vs GPT vs Gemini: Which AI API Is Actually Cheapest (September 2026)

"Which AI is cheapest" sounds like a simple question until you actually try to answer it. Claude, GPT and Gemini each publish per-million-token prices across a handful of tiers, the cheapest model on paper is not always the cheapest model for your prompts, and every vendor has quietly changed at least one price this year. Here is the real comparison, using each provider's own published numbers as of 18 September 2026.

Why "which AI is cheapest" doesn't have one answer

Every major provider prices per million tokens, split into input (what you send) and output (what the model generates) — and output almost always costs more than input, sometimes five times more. A model that looks cheap on its input price can end up expensive if your use case generates long responses, and a model with a high sticker price can be cheap in practice if you cache repeated context.

On top of that, each provider now ships multiple tiers — a flagship model for hard reasoning, a mid-tier model most apps actually run in production, and a budget model for high-volume, low-stakes tasks. Comparing a budget model from one provider to a flagship model from another tells you nothing useful. The only fair comparison is tier-to-tier.

Flagship models: for the hardest tasks

These are the models each company points to for their most demanding reasoning, coding and agentic work.

Gemini 3.1 Pro's price doubles past the 200k-token context mark — a detail that's easy to miss when comparing sticker prices.

Mid-tier models: what most production apps actually run on

This is the tier that matters most for most builders — strong enough for real product features, priced for sustained usage.

Gemini's Flash tier undercuts both Claude Sonnet and GPT Sol on paper, but it's a lighter-weight model — check it handles your task quality before switching purely on price.

Budget models: for high-volume, low-stakes jobs

For classification, tagging, simple extraction and other high-volume jobs where every model is "good enough," these are the tiers to compare.

GPT-5.6 Luna is the cheapest budget model of the three by a wide margin — but budget models vary a lot in quality, so test before committing at scale.

The output token trap

Because output tokens cost several times more than input tokens across every provider, the models that "talk more" — longer explanations, more reasoning tokens, more verbose formatting — end up costing more than their sticker price suggests. Two models with identical input pricing can produce very different bills if one tends to write twice as much per response. When comparing cost, always test with your actual prompts and count the tokens your specific use case produces, not just the published per-token rate.

Cached input changes the math

All three providers now offer a discount for repeated/cached input — content like a long system prompt or reused document context that appears identically across many requests. Cached input can run as low as 10% of the standard input price. If your app sends the same large context on every call (a support bot with a long knowledge base, for example), prompt caching can cut your effective cost dramatically regardless of which model you pick — so it's worth checking whether your provider and SDK support it before assuming a model is too expensive.

How to actually compare your own use case

The published per-token prices are a starting point, not the answer. To find the real cheapest option for your app:

• Estimate your typical input and output token counts per request, not just a guess — actually count them.
• Multiply by each candidate model's input/output rates and compare the per-request cost, not the per-token headline price.
• Factor in prompt caching if your requests share repeated context.
• Weigh cost against quality — the cheapest model that fails your task and needs a retry (or a human fix) isn't actually cheaper.

If you want to skip the manual math, plug your own prompt and expected output length into our free LLM Cost Calculator and it will work out the per-request and monthly cost across Claude, GPT and Gemini tiers automatically.