API cost calculator · prices checked 2026-10-11 · runs measured 2026-10-10 to 2026-10-11

Same price, different bill Claude Haiku 5.5 vs GPT-6 Luna

Claude Haiku 5.5 and GPT-6 Luna have the same list price, but their tokenizers split text differently: for the same English text, Haiku counts ×1.60 as many tokens as Luna. Enter your task to see what each would actually bill.

List price, both models: $0.10 in · $0.50 outper million tokens
Haiku tokens per Luna token, same text: EN ×1.60 · JA ×1.26code and JSON ×1.56–×1.62
Higher price tier starts at: 100K · 272Kinput tokens per request (Haiku · Luna)

Your task

Example scenarios

Approximates our measured run. Measured: $0.237 (Claude Haiku 5.5) and $0.159 (GPT-6 Luna) per 1,000 requests.

System prompt language

System prompt characters

Each input is

Input characters per request

≈ Haiku 217 · Luna 172 input tokens per request

JSON structured outputNumber of fields

Both APIs add schema instructions to the input: about 162 + 26 per field for Haiku, 14 + 9 per field for Luna (string fields, measured).

Output is

Output characters per request (the two models may write different lengths)

Claude Haiku 5.5

GPT-6 Luna

≈ Haiku 402 · Luna 215 output tokens per request, before reasoning

Reasoning tokens per request (billed as output)

Claude Haiku 5.5

GPT-6 Luna

The scenarios use reasoning tokens from our medium-effort runs: Luna reports them, Haiku's are estimated (assumption). Your own logs are the best source.

Requests per month

Share of input read from cache: 0%

Batch API

Your monthly bill

Claude Haiku 5.5 costs 49% more than GPT-6 Luna: $7.29 vs $4.91 a month

The main reason is reply length: Claude Haiku 5.5 writes about 442 characters per reply, GPT-6 Luna about 297, and output costs more ($0.50 per million output tokens against $0.10 for input).

Claude Haiku 5.5

Input$0.10/M	217 tok	$0.650
Cache reads$0.01/M	0 tok	$0
Output$0.50/M	402 tok	$6.02
Reasoning$0.50/M	41 tok	$0.620

$7.29 a month $0.243 per 1,000 requests

GPT-6 Luna

Input$0.10/M	172 tok	$0.515
Cache reads$0.01/M	0 tok	$0
Output$0.50/M	215 tok	$3.23
Reasoning$0.50/M	77 tok	$1.16

$4.91 a month $0.164 per 1,000 requests

How the gap changes as each input gets longerClaude Haiku 5.5's bill divided by GPT-6 Luna's, with every other setting fixed. Dashed lines mark where each model's higher price tier starts; the solid line is your current input length. The curve stops where a model's input limit is reached.With your other settings, the higher price tier starts for Claude Haiku 5.5 from about 109,898 characters and for GPT-6 Luna from about 375,260 characters of input per request.

Show the curve as a table

Input characters per request	Haiku ÷ Luna
1,000	×1.42
10,000	×1.30
50,000	×1.26
100,000	×1.26
200,000	×6.29
500,000	×3.14
1,000,000	×3.14

What three real tasks cost

Same prompts, medium effort on both models, costs recomputed from the billed tokens at current list prices. The scenarios above approximate these runs. Why the gaps differ

Cost per 1,000 requests; accuracy and latency listed Haiku / Luna
Task	Requests	Claude Haiku 5.5	GPT-6 Luna	Haiku ÷ Luna	Accuracy / blind judging	Median latency	Measured
Ticket classification (English and Japanese)	50	$0.033	$0.033	×0.98	100% / 100% correct	0.90 s / 1.45 s	2026-10-11
Invoice extraction (structured output, 6 fields)	50	$0.147	$0.090	×1.64	100% / 100% correct	1.72 s / 2.23 s	2026-10-10
Japanese business email	25	$0.237	$0.159	×1.49	GPT-6 Luna preferred: 22/25, 25/25	2.44 s / 3.56 s	2026-10-10

Email judging: Claude Opus 5.5 and GPT-6.1 Sol each compared the two replies blind, in both orders.

How the calculator works

Characters become tokens through the characters-per-token ratios below, measured with each vendor's token-counting endpoint. Corpus and method
Every request also carries message-structure tokens (9 for Haiku, 6 for Luna). Structured output adds the schema instructions shown under that input.
List prices come from each vendor's official pricing page, checked 2026-10-11. Once a single request's input passes a price tier, the whole request is billed at the higher rate. Prices and sources
Cache and Batch discounts multiply. Not included: cache-write charges, retries, timeouts.
assumption Claude Haiku 5.5: The batch table lists only input and output prices; we assume the 50% batch discount also applies to cache reads ("Batch API requests are 50% off").

Characters per token by content type
Content	Claude Haiku 5.5	GPT-6 Luna	Haiku ÷ Luna tokens	95% interval	Samples
English	2.60	4.17	×1.604	1.552–1.683	24
Japanese	1.10	1.38	×1.261	1.235–1.286	24
Chinese	1.00	1.39	×1.393	1.375–1.413	24
JavaScript / TypeScript	2.20	3.56	×1.616	1.559–1.681	24
Python	2.65	4.14	×1.561	1.527–1.592	24
JSON	2.11	3.39	×1.608	1.517–1.690	15