OpenAI vs Claude vs Gemini: API Pricing Compared

Every major LLM provider bills the same way — per token, with input and output priced separately — but the actual rates, model tiers, and pricing quirks differ enough between OpenAI, Anthropic and Google that “which one is cheaper” doesn’t have a single answer. It depends on your task, your typical input/output split, and which tier within each provider’s lineup you’re actually comparing.

The shape of each lineup

All three providers structure their lineup around the same basic idea: a flagship tier for demanding reasoning tasks, a mid-tier for general-purpose work, and a lightweight tier for high-volume, latency-sensitive tasks — priced roughly an order of magnitude apart between the lightest and heaviest tier in each family.

• OpenAI: reasoning-focused o1-family models at the top, GPT-4o as the capable mid-tier default, GPT-4o mini as the high-volume budget option.
• Anthropic: Claude Opus at the top, Claude Sonnet as the popular general-purpose default, Claude Haiku as the lightweight option.
• Google: Gemini Pro for demanding tasks, Gemini Flash tuned specifically for strong performance at a much lower price point.

The GPT Token Calculator, Claude Token Calculator and Gemini Token Calculator on this site each default to that provider’s popular mid-tier rate, so you can price a specific task on each provider’s comparable tier before deciding.

Provider	Flagship tier	Mid-tier	Lightweight tier
OpenAI	o1-family (reasoning)	GPT-4o	GPT-4o mini
Anthropic	Claude Opus	Claude Sonnet	Claude Haiku
Google	Gemini Pro	—	Gemini Flash

Why comparing headline rates alone is misleading

A per-token price comparison across providers is a reasonable starting point, but it skips two things that meaningfully change a real workload’s actual cost.

The input/output split matters more than the total. Every provider prices output tokens higher than input — commonly 3-5x — so a task that generates a lot of text (long-form writing, code generation) costs disproportionately more relative to its total token count than a task that mostly reads and briefly responds (classification, short extraction). Comparing providers on a single blended rate hides this; comparing on your actual expected input/output split doesn’t.

Provider-specific pricing mechanics change the real number. Claude offers prompt caching — a substantially discounted rate for repeated, unchanged context like a long system prompt reused across many requests — which can meaningfully lower real cost for context-heavy workloads below the standard rate. Gemini’s pricing steps up for some models once a request’s context crosses a length threshold (commonly in the 128K–200K token range), meaning a long-context-heavy workload’s real average rate can be higher than the headline number suggests. Neither of these shows up in a simple side-by-side table of per-token prices.

”What does the Claude API cost for 1 million tokens?”

This is one of the most common searches around LLM pricing, and the honest answer is: it depends on the model tier and the input/output split, because a million tokens weighted toward output costs meaningfully more than the same million weighted toward input.

On Claude Sonnet 4, a million tokens split evenly (500K in / 500K out) costs about $9.00. The same split on Claude Opus 4 costs about $45.00 — five times more, reflecting Opus’s flagship-tier pricing. On the equivalent split for GPT-4o, the answer is about $6.25; for Gemini Pro, about $3.125. None of these numbers is “the” answer to “what does a million tokens cost” — they’re all correct for their specific model and split, which is exactly why a calculator that lets you set your own split and rate, rather than a single quoted figure, is the only way to get a number that actually applies to your workload.

Cost for 1 million tokens split evenly 500K in / 500K out, on each model's standard tier.

From a single request to a real budget

Pricing a single request is a useful sanity check, but the number that actually matters for planning is a monthly projection: expected requests per day, times average tokens per request, times your model’s rate, projected out across a month or year. That’s a different calculation from pricing one hypothetical request, and it’s the one that catches teams off guard when a feature that looked cheap in testing turns into a real line item at production volume. The OpenAI API Cost Calculator and Claude API Pricing Calculator on this site are built specifically for that projection, not just a single-call estimate.

The practical comparison workflow

1. Estimate your actual input/output token split for the specific task, not a generic average.
2. Price that split on each provider’s comparable tier using the token calculators above, rather than comparing headline per-token rates.
3. Check for provider-specific pricing mechanics that apply to your workload — prompt caching if you have repeated context, long-context tiers if your requests are large.
4. Project to your real expected volume before deciding, since the tier and provider choice compounds directly with scale.

None of this requires guessing — every number is directly computable, and the right choice between OpenAI, Claude and Gemini almost always comes down to your specific task and volume rather than a single universal “cheapest” answer.