OpenAI vs Anthropic pricing in 2026

OpenAI and Anthropic publish per-million-token prices on dedicated pricing pages, and both move them around two or three times a year. Rather than memorise current numbers (they'll be out of date by the time you read this), the framework that matters is: pick a tier, then pick whichever provider is cheaper inside that tier today.

The three tiers

Both providers organize their lineup into three tiers, even though the names differ:

• Cheap and fast. OpenAI: gpt-4o-mini. Anthropic: claude-haiku-4-5. Penny-per-million-input-tokens territory. Use this as the default for high-throughput, latency-sensitive, simpler tasks.
• Workhorse. OpenAI: gpt-4o. Anthropic: claude-sonnet-4-6. The default "good model" for most production workloads. Order-of-magnitude more expensive than the cheap tier.
• Heavy reasoning. OpenAI: o1 / o3 family. Anthropic: claude-opus-4-7. Reserve for genuinely hard problems where the latency and cost are worth it.

Compare like-for-like

When you do compare, normalize on input + output combined cost per million tokens for a representative ratio (1:1 is fine for chat; 5:1 input-heavy is fine for retrieval). Don't pick on input cost alone — providers periodically lower input and quietly raise output, or vice versa, to look better on superficial comparisons.

Caching is the bigger lever

Both providers ship prompt caching: Anthropic at ~10% of normal input cost, OpenAI at a similar discount tier. If your prompts have a stable prefix (system prompt + retrieval context + tool spec), caching cuts 60–80% off your bill in practice. This usually dwarfs the per-token gap between providers.

Batch APIs cut another 50%

If your workload tolerates a 24-hour SLA, both providers run a batch endpoint at a flat 50% discount on all tokens (cached or not). Useful for nightly summarization, classification, evals, fine-tuning data prep.

Hidden lines

• Image input: charged as token equivalents. A high-res image can be 1,500+ tokens.
• Tool use: function call parameters and outputs are billed as input/output tokens like any other content.
• Reasoning tokens: o1/o3-style thinking tokens count toward output cost even if you don't see them in the response.

What to do next

Run the same evaluation prompt through both providers using the playground to compare quality before chasing pennies. The provider that's slightly cheaper but worse at your task ends up costing more in retries.