Key takeaways
- Anthropic vs OpenAI API pricing turns on your input/output token ratio, not the sticker price on either pricing page.
- Claude lists lower flagship and balanced rates, OpenAI lists a cheaper fast tier as of September 2026. Prices change, so recheck both sources.
- Both vendors bill usage-based in US dollars, so how you pay the invoice matters as much as which API you choose.
Most teams pick an API by scanning two pricing pages and comparing the top number. That number rarely predicts your bill. Output tokens cost several times more than input tokens on both platforms, so the cheaper API for a summarizer that reads a lot and writes little is not the cheaper API for an agent that reads little and writes a lot. Your token mix decides. Below is a like-for-like Anthropic vs OpenAI API pricing comparison, the math that shifts it, and one detail both vendors share.
How does Anthropic vs OpenAI API pricing compare?
Both charge per million tokens, split into input and output, and bill in US dollars. Claude lists lower flagship and balanced rates, OpenAI lists a cheaper fast tier. That split is the whole story: a single blended price hides the fact that output is where most bills are won or lost, which is why an Anthropic vs OpenAI API pricing comparison has to run at the tier and token level, not on headline numbers.
Here are the current published rates by tier, per million tokens, as of September 2026.
| Tier | Claude (input / output) | OpenAI (input / output) |
|---|---|---|
| Flagship | Opus 5.5: $4 / $20 | GPT-5.6 Sol: $5 / $30 |
| Balanced | Sonnet 5: $2 / $10 | GPT-5.6 Terra: $2 / $12 |
| Fast and affordable | Haiku 4.5: $1 / $5 | GPT-5.6 Luna: $0.20 / $1.20 |
Claude also lists Fable 5.1 at $10 / $50 as a specialist tier with no direct OpenAI counterpart in this Anthropic vs OpenAI API pricing comparison, so we leave it out of the like-for-like rows. For the full Claude lineup and the levers that cut a Claude bill, see Anthropic API pricing. Both pages move, so treat this as a September 2026 snapshot and verify before you budget.
Which API is cheaper for high-output workloads?
The answer changes with your ratio, so run your own numbers. Here is an illustrative example on the flagship tier, using a fixed one million tokens split two ways.
Input-heavy job (900k input, 100k output):
- Opus 5.5: (0.9 x $4) + (0.1 x $20) = $5.60
- GPT-5.6 Sol: (0.9 x $5) + (0.1 x $30) = $7.50
Output-heavy job (100k input, 900k output):
- Opus 5.5: (0.1 x $4) + (0.9 x $20) = $18.40
- GPT-5.6 Sol: (0.1 x $5) + (0.9 x $30) = $27.50
Claude is cheaper in both cases here, but notice the gap swings from roughly 25 percent on input-heavy work to roughly 33 percent on output-heavy work. Output rates drive it. The comparison also flips by tier, not just by mix: Claude's flagship and balanced tiers come in lower, while OpenAI's fast Luna tier lists the lowest input rate on this table at $0.20. So a retrieval pipeline that ships huge prompts and short answers can land on a different vendor than an agent that writes long. Model your actual input/output split against these Anthropic vs OpenAI API pricing rates before you assume a winner.
How do caching and batch change the math?
The headline rate is the ceiling, not the bill. Both vendors offer the same two levers.
Batch on both platforms saves about 50 percent for work that does not need a real-time response. If you run nightly evaluations, bulk classification, or backfills, that is close to half off the numbers above with no code rewrite beyond the batch endpoint.
Caching cuts the cost of repeated context. Claude prompt caching makes cached reads much cheaper than fresh input, and OpenAI charges much less for cached input tokens too. For an agent that resends the same system prompt and tools on every turn, or a chatbot with a long fixed instruction, cached input can dominate your input line. A workload that looks expensive at list rates can be the cheaper option once caching is in place, so factor both levers into any Anthropic vs OpenAI API pricing comparison before you rank vendors.
Does the currency you pay in matter?
Yes, and this is the part the pricing pages do not show. Both invoices land in US dollars. If your company banks in euros, pounds, rupees, or Canadian dollars, every charge crosses a currency boundary, and your card issuer decides the rate and the markup, the same drag covered in how to pay for the OpenAI API. A 2 to 3 percent conversion fee on top of usage is common, and on an AI bill that grows every month, that surcharge compounds. You optimized the token rate to cents per million, then handed back single digit percentages at the border. The API you chose does not change that. No Anthropic vs OpenAI API pricing comparison does. The account you pay from does.
Like-for-like by tier
Fast tier
Haiku $1/$5 · Luna $0.20/$1.20OpenAI is cheapest at the fast tier.
Balanced tier
Sonnet $2/$10 · Terra $2/$12Near-identical; Claude edges output cost.
Flagship tier
Opus $4/$20 · Sol $5/$30Claude is cheaper at the top tier.
Per million tokens, input / output, as of September 2026. The cheaper API depends on your token mix, and both bill in USD.
How Endl fits
Both the Claude and OpenAI APIs bill usage-based in US dollars, so the cleanest way to pay either one is from a US dollar balance that does not convert on every charge.
Endl gives you a USD virtual card that spends at 0 percent on card transactions, with no hidden markup. Spend in other currencies settles at the Visa rate plus a flat 1 percent, so the cost is one predictable line rather than a variable bank spread. You can issue virtual and physical cards, set a per-card limit for each vendor, and freeze any card instantly if a key leaks or spend spikes. Your balance sits on self-custodial stablecoin rails, and moving money is a flat 0.5 percent.
A few things to be clear about. Endl is not a bank and balances are not insured. Endl operates as a registered VASP in the EU and an MSB in Canada. The cards are debit and spend from your balance, so they are not credit.
Ready to pay your AI bills in USD without the conversion drag? See the pricing page or start free.
