ChatGPT API vs Claude API pricing: within a cent, nine points apart

Price per answer with the score beside it, which is the only fair table

The vendors publish per million token rates, and that's the wrong unit for a decision. A model that's cheap on input and dear on output can cost more per reply than one priced the other way round. The Whizi cost index fixes the unit: one standard answer of 1,000 input tokens and 500 output tokens, priced at the provider rates it read on 20 August 2026. The scores are the Artificial Analysis Intelligence Index, read on 22 September 2026, at each model's maximum reasoning setting.

The frontier runs Luna, Sonnet, Terra, Sol, Opus. Sonnet 5 makes it by a single point over Luna at nine times the price, Haiku 4.5 sits well inside it, and from $0.008 to a cent the best score at any spend is OpenAI's. Above a cent it's Anthropic's. That's the chart's one sentence: the OpenAI row scores higher for the money until you can afford Opus.

Two cautions on reading it. The Intelligence Index is one benchmark family. A nine point gap on it doesn't mean Sol writes better prose than Sonnet, and on drafting and editing the score has never matched what editors report. And the index prices list rates: prompt caching, batch discounts and long context surcharges move real bills in both directions. What are tokens in AI covers why output is priced five to six times input on every row.

The full rate card, OpenAI beside Anthropic

Three readings from the table. Anthropic prices its tiers on a clean ladder, $1, $2 and $5 input, with output at exactly five times input on every row. OpenAI's ladder is steeper at the bottom and flatter in the middle: Luna is a tenth of Terra, but Terra and Sol are within 25 percent of each other. And the newest Anthropic flagship, Fable, is priced at twice Opus, $10 in and $50 out, the same rates the index lists for the fast variants of Opus 5 and Opus 4.8.

The GPT-5.5 row is a warning, not a recommendation. It costs twice Sol per answer, and the leaderboard lists its Instant variant at 26. At list price its own successor dominates it. The AI model cost index carries all 100 priced rows, including the rows from Google, xAI, Moonshot and DeepSeek that undercut both vendors here.

Which vendor is cheaper depends on which tier you are buying

At the bottom, OpenAI, by a wide margin. Luna is $0.0008 an answer and scores 37. Haiku 4.5 is $0.0035 and scores 17. For classification, extraction, rewrites and everyday questions, the OpenAI row is a quarter of the price and twice the score. There's no honest way to read that differently.

In the middle, a wash on price and an OpenAI lead on the index. Sonnet 5 at $0.007 against Sol at $0.01 is a 30 percent gap in Anthropic's favour. Sol's 47 against Sonnet's 38 is a nine point gap in OpenAI's. Which one matters depends on the task. On code, maths and multi step reasoning the index tracks outcomes reasonably well. On prose it doesn't, and Sonnet is the row editors keep picking.

At the top, Anthropic, if you need the top. Opus 5 at $0.0175 scores 51, four points above Sol for 75 percent more per answer. Above it, Fable 5 at $0.035 is the most expensive row in the cost index short of the fast variants. Whether four points is worth 75 percent depends on the task. You'll find what each extra point costs on the reasoning effort dial in what is a reasoning model. Opus 5.5 came out on 22 September 2026 at $4 in and $20 out, so the same answer is now $0.014. The cheapest way to use Claude Opus halves that on the Batch API.

The cheapest way to use Claude and GPT together runs light, medium and heavy months on these rates for someone who wants both APIs at once. ChatGPT API vs subscription is the break even against a flat $20 plan.

What the same rows cost per message on a credit plan

This site publishes a per message price for every one of these models, so here they are next to the API rates, not instead of them. On Whizi GPT-5.6 Luna is 1 credit a message, Terra is 4, Sol is 20, Claude Haiku 4.5 is 4, Sonnet 5 is 10 and Opus 5 is 20. The monthly allowances are 800 credits on Starter at $15.99, 2,500 on Pro at $29.99 and 10,000 on Powerhouse at $49.99. Of those six rows Starter reaches only Luna, and its Claude row is the newer Haiku 5.5 at 1 credit. Terra, Sol, Haiku 4.5, Sonnet and Opus start on Pro.

Read against the API, the credit ladder prices Sonnet below Sol, 10 credits to 20, where the API prices them within a cent of each other. It also prices Opus level with Sol at 20, where the API charges 75 percent more for Opus. That's a product decision. Whizi charges both the Sonnet family and Sol below the rung their token cost would put them on, and the credits reference has the full ladder. Per flagship answer a credit plan costs more than the API. Per everyday answer on Luna it's close. The plan also adds the app, memory, uploads and a bill with a ceiling, which the API doesn't.

If you're buying API access, buy OpenAI for anything under a cent an answer. Buy Anthropic when the task needs the top of the ladder. Price the middle by the task, not by the index.