OpenAI vs Anthropic API: which costs less per task?

OpenAI is cheaper per task than Anthropic at the budget and mid tiers in my measurements. GPT-6 Luna undercuts Claude Haiku 5.5, and GPT-6.1 Sol undercuts Claude Sonnet 5.5, on chat, RAG and invoice extraction. At the top end the pairing breaks down: Claude Opus 5.5 costs less per task than GPT-6 Astra, because its list price is far lower than Astra's.

How I paired the models

I matched each OpenAI tier to the Anthropic model with the closest list price (per 1M tokens, input / output):

• Budget: GPT-6 Luna vs Claude Haiku 5.5. Both list at $0.10 / $0.50.
• Mid: GPT-6.1 Sol vs Claude Sonnet 5.5. Both list at $2.00 / $10.00.
• Premium: GPT-6 Astra vs Claude Fable 5.1. Both list at $10.00 / $50.00.
• Reference: Claude Opus 5.5 at $4.00 / $20.00 sits between Sonnet and Fable, so I treat it as a separate point. GPT-5.4 Nano and Mini are older OpenAI tiers that I include as cheaper alternatives.

Context windows are close. GPT-6 Luna, Sol and Astra list 1,050,000 tokens. The Claude models list 1,000,000. GPT-5.4 Nano and Mini list 400,000. For the tasks here, context size did not decide the cost.

Cost per task

These are cents per single task. Luna, Sol, Haiku 5.5 and Sonnet 5.5 are measured. Astra, Fable 5.1 and Opus 5.5 are estimated. Runs used default settings via OpenRouter, with output capped at 16,000 tokens. Reasoning tokens are billed as output.

The median across all 332 models is 0.25¢ for a chat answer, 0.49¢ for RAG and 0.23¢ for invoice extraction. See the task pages for chat answers, RAG and invoice extraction.

Budget tier: OpenAI wins

GPT-6 Luna is below Haiku 5.5 on all three tasks. The gap is small on chat (0.06¢ vs 0.07¢) and largest on RAG (0.05¢ vs 0.14¢).

The list prices are identical, so the difference comes from token counts. On RAG, Haiku used 5,184 input tokens against Luna's 3,690. Part of that is the tokenizer: the same English document comes to 1.39 tokens on Anthropic's tokenizer against 1.00 on OpenAI's. Haiku also produced more output on RAG, 1,096 tokens against 234.

GPT-5.4 Nano is cheap too, at 0.15¢ on chat and 0.12¢ on RAG, but Luna beats it on both. Luna is the cheapest option of those I compared.

Mid tier: OpenAI wins again

Sol costs about half of Sonnet 5.5 on chat and invoices: 0.67¢ vs 1.3¢ on chat, and 0.52¢ vs 1.0¢ on invoices. On RAG it is 1.2¢ vs 2.7¢.

The list price is the same, so this is token behavior. On chat, Sonnet produced 1,256 output tokens against Sol's 560. On RAG, Sonnet used 5,184 input tokens against Sol's 3,690, and produced 1,079 output tokens plus 615 reasoning tokens against Sol's 442 output and none for reasoning.

GPT-5.4 Mini sits below Sol on chat (0.53¢) and RAG (0.43¢), and below Sonnet on both. It is a cheaper option if you can accept an older model.

Premium tier: Astra beats Fable, Opus beats Astra

GPT-6 Astra is cheaper than Claude Fable 5.1 on every task: 3.4¢ vs 6.3¢ on chat, 5.9¢ vs 14¢ on RAG, and 2.6¢ vs 5.2¢ on invoices.

Claude Opus 5.5 is the surprise. Both Opus and Astra figures are estimates, built from the measured token counts of Sonnet 5.5 and GPT-6.1 Sol. Opus lists at $4.00 / $20.00, well below Astra's $10.00 / $50.00, and it costs less per task on all three: 2.5¢ vs 3.4¢ on chat, 5.5¢ vs 5.9¢ on RAG, and 2.1¢ vs 2.6¢ on invoices. On RAG the gap is narrow. Opus has an LMArena Text score of 1507 against Astra's 1475, so on these numbers it is cheaper and rated higher. Astra's lead over Fable is real, but Opus is the better pick at this level. See the Opus 5.5 page for details.

Where reasoning mode changes the result

List price tells you little when reasoning tokens are billed as output. Reasoning behavior differs a lot between these models:

• GPT-6 Luna reasons by default. On chat it used 639 reasoning tokens and still came out cheapest, because its list price is low and its visible output was short, at 540 tokens.
• GPT-6.1 Sol always reasons, yet used only 103 reasoning tokens on chat and none on RAG. That keeps its bill low.
• Claude Sonnet 5.5 also always reasons. It used 0 reasoning tokens on chat, but its visible output was 1,256 tokens, which is what drove its cost.
• Claude Haiku 5.5 used 244 reasoning tokens on chat and 753 on RAG, on top of long visible answers.

Reasoning tokens vary by task and by run, so check them on your own prompts. These are single runs, and reasoning behavior can shift the ranking on harder tasks. On code review, for example, Luna used 4,921 reasoning tokens and still cost 0.31¢, but results there depend heavily on the input. See the code review page.

Which provider wins at each budget

Claude Fable 5.1 was the most expensive model on every task here, so I would only pick it for reasons other than cost.

These results come from one run per task with default settings. Some figures are estimates, and the task pages show which. To model your own workload, use the calculator, and read the methodology for how tokens and reasoning are counted.

FAQ

Is GPT-6 Luna really cheaper than Claude Haiku 5.5?

Yes, on all three tasks here, though the chat gap is small: 0.06¢ vs 0.07¢. Both list at $0.10 input and $0.50 output, so the gap comes from token counts and tokenization.

Does the tokenizer difference matter for cost?

It can. The same English document is 1.39 tokens on Anthropic's tokenizer against 1.00 on OpenAI's, which raises input token counts on Claude even when list prices match.

Should I always pick the cheapest model?

No. Cost per task is one number. Check quality on your own inputs, and compare LMArena scores on the model pages where available.

Sources

• OpenRouter model list and provider endpoints (list prices, read daily)
• TaskCents methodology: how tokens per task are measured and estimated