OpenAI vs Anthropic API pricing

If you are choosing between OpenAI and Anthropic on price, the honest answer to "which is cheaper" is: it depends on the tier you need, how your text tokenizes, and how much you cache — and at least once this year the cheaper option flips. This is a claim-and-evidence walk through OpenAI vs Anthropic API pricing at every tier, with the numbers that actually decide your bill rather than the ones on the pricing page.

Two rules before the table. First, price is quoted per million tokens, but you pay per token your text actually becomes — and the two providers tokenize the same text differently. Second, list prices move; the figures below are dated, and a couple of them expire on a date we will name.

The sticker prices, tier by tier

Here are the current list prices, tier-matched — flagship against flagship, workhorse against workhorse, budget against budget, plus OpenAI's dedicated reasoning model. Cached input is the discounted rate for reading a previously cached prompt prefix.

Claude Sonnet 5's $2 / $10 was introduced as a time-limited rate, but in August 2026 Anthropic made it the standard list price and cancelled the scheduled increase to $3 / $15. GPT-5.4 mini's context window is not published in the source dataset, shown as —.

Read the flagship row carefully

At the very top the two providers are now level on sticker: GPT-6 Astra and Claude Fable 5.1 both list $10.00 / $50.00, which puts the same representative 10K-input / 1K-output request at $0.15 on either. That tie does not survive contact with real text, for the reason in the next section — but it is the first time the two top models have matched to the cent. It also does not survive caching: Fable 5.1 and Mythos 5.1 (launched 2026-09-01) price cache reads at 0.025× input, $0.25 per million against Astra’s $1.00, so a prefix-heavy workload pays roughly 35% less on the input side at a 90% hit rate despite the identical list price. On paper the flagship tier now favors OpenAI. For a representative request of 10K input and 1K output tokens, Claude Opus 5 costs $0.075 and GPT-5.6 Sol costs $0.060 — Sol about 20% cheaper, after OpenAI cut it to a promotional $4 / $20 in effect at least through 2026-11-21. But sticker price is not cost per task, and here the real gap is wider than the sticker suggests, not narrower.

The tokenizer trap: per-token price is not cost per task

A price of "$5 per million tokens" only tells you your bill if you know how many tokens your text becomes, and that depends on the tokenizer. Anthropic's newest-generation models — Opus 5 and the premium Fable 5 — use a tokenizer that produces roughly 30% more tokens for the same English text than the previous generation. Claude Haiku 4.5, the budget row in the table above, predates that tokenizer and is unaffected. Price the same actual prompt on both sides and the flagship gap widens: the request that is 10K input / 1K output on GPT-5.6 Sol is closer to 13K / 1.3K on Opus 5, which moves Opus 5 from $0.075 to about $0.098 — now roughly 60% more expensive than GPT-5.6 Sol for identical text. The lesson is general: compare cost per task on your own prompts, not headline price per million tokens.

This cuts both ways. Anthropic's mid and budget models — Sonnet 5 and Haiku 4.5 — do not carry the newer-tokenizer note, so their sticker prices are directly comparable to OpenAI's. Always check whether a per-token price is even denominated in the same tokens before you trust a comparison. The same applies at the top: Astra and Fable 5 list identically, but priced on the same actual prompt rather than the same token count, Fable 5 lands roughly 30% higher.

The mid-tier cliff that never came

The workhorse tier is where most production traffic lives, and it favors Anthropic. Claude Sonnet 5 at $2 / $10 runs the 10K/1K request at $0.030 against GPT-5.4's $0.040 — 25% cheaper, and text-for-text, because Sonnet 5 uses the standard tokenizer. When this post first ran, that came with an asterisk: $2 / $10 was an introductory price scheduled to rise to $3 / $15 on 2026-09-01, which would have flipped the tier to GPT-5.4. The cliff never came — in August 2026 Anthropic made $2 / $10 the permanent standard price and cancelled the increase. The mid-tier stays with Sonnet 5.

The budget tier goes to OpenAI

At the small-and-fast end, OpenAI is cheaper: GPT-5.4 mini runs the 10K/1K request at $0.012 versus Claude Haiku 4.5's $0.015, about 20% less. And OpenAI's line extends well below mini — GPT-5.6 Luna and GPT-5.4 nano list at $0.20 input, and legacy GPT-5 nano at $0.05 — territory Anthropic's public API does not price against today. If your workload is high-volume classification, extraction, or routing, that gap compounds; the cheapest-LLM guide linked below ranks the whole budget field.

Cache and batch: where the real savings hide

Both providers discount reading a cached prompt prefix by 90% on their GA flagship and workhorse models — Opus 5 drops cached input from $5.00 to $0.50, GPT-5.6 Sol from $4.00 to $0.40, and Sonnet 5 from $2.00 to $0.20 — and Anthropic goes further on Fable 5.1 and Mythos 5.1, at 97.5%. On the write side the two have converged since we first published this comparison. Anthropic charges a cache-write premium: populating the cache costs 1.25× the base input rate for the five-minute TTL and 2× for the one-hour TTL (Opus 5 writes at $6.25 against $5.00 input on the five-minute cache), so the cache pays off only once a cached prefix is reused enough to amortize that write. OpenAI's cache is still automatic and used to charge nothing to write, but as of GPT-5.6 it bills cache writes at 1.25× the uncached input rate — matching Anthropic's five-minute premium; on models through GPT-5.5 the write is still free. The difference that remains is on reads: OpenAI's dedicated reasoning models (o3, o4-mini) discount cached reads by 75%, not 90%. For asynchronous work, both offer a Batch API at 50% off. If you run a long, stable system prompt across many calls, model the cache explicitly — it moves the answer more than the sticker price does.

Which should you pick

There is no single winner; there is a winner per workload:

• ◦High-volume budget work (classification, extraction, routing): OpenAI. GPT-5.4 mini undercuts Haiku 4.5, and the nano tier goes lower still.
• ◦Mid-tier production traffic: Claude Sonnet 5 — 25% cheaper text-for-text than GPT-5.4, and its $2 / $10 rate is now the permanent standard price, not an intro.
• ◦Flagship quality: GPT-5.6 Sol, cut to a promotional $4 / $20, is now cheaper than Claude Opus 5 on sticker (about 20%), and Anthropic's ~30% higher tokenization on Opus 5 widens the gap on real text. Sol's cached input is lower too ($0.40 vs $0.50), so a heavily reused cached prompt does not flip it back — and with GPT-5.6 now billing cache writes at 1.25× input, the same multiple as Anthropic's five-minute cache, the write premium is no longer a point of difference.
• ◦Either way: decide on cost per task for your real input:output ratio, cache-hit rate, and tokenizer — not on price per million tokens.

Is OpenAI or Anthropic cheaper?

It depends on the tier. For budget, high-volume work OpenAI is cheaper — GPT-5.4 mini at $0.75 / $4.50 undercuts Claude Haiku 4.5 at $1.00 / $5.00. For mid-tier work Claude Sonnet 5 is cheaper — and its $2 / $10 rate, once an intro price, was made the permanent standard in August 2026. At the flagship tier GPT-5.6 Sol is now cheaper on sticker — about 20% below Claude Opus 5 after OpenAI's promotional cut to $4 / $20 — and pulls further ahead once you adjust for Anthropic's higher tokenization.

How much does prompt caching save with each provider?

Both cut the price of reading a cached prompt prefix by 90% on their GA flagship and workhorse models; Anthropic goes further on Fable 5.1 and Mythos 5.1, at 97.5%. Anthropic charges a 25% premium to write into the cache (1.25× base input for the five-minute TTL, 2× for the one-hour TTL), so reuse has to be high enough to earn it back. OpenAI used to charge nothing to write — but as of GPT-5.6 it matches Anthropic almost exactly, billing cache writes at 1.25× the uncached input rate; on models through GPT-5.5 the write is still free. The two providers have converged here since we first published this comparison. OpenAI's o-series reasoning models discount cached reads by 75% rather than 90%.

Why don't the sticker prices predict my bill?

Because you are billed per token, and the same text becomes a different number of tokens on each provider's tokenizer — Anthropic's newest models, Opus 5 and Fable 5, run about 30% higher than the prior generation, while Claude Haiku 4.5 predates that tokenizer and is unaffected. Your input:output ratio and cache-hit rate matter just as much. Two models with identical per-million prices can produce very different monthly bills, so price your own token mix.

Do these prices change often?

Yes — in both directions. In August 2026 alone, Anthropic made Claude Sonnet 5's intro price permanent (cancelling a planned increase), and OpenAI cut its GPT-5.6 Sol flagship to a promotional $4 / $20 in effect at least through 2026-11-21 (having already cut GPT-5.6 Luna sharply weeks earlier). Treat any published price as dated — the figures here are list prices as of 2026-09-15 — and re-verify against the provider's pricing page before you commit a budget.