mario0369/llm-cost-same-prompt

Measured per-call LLM cost — same prompt, every model

Vendors publish prices per million tokens. Nobody publishes what one call actually costs, because that depends on how many tokens the model chooses to emit — and on the same question models differ by more than an order of magnitude. One model finishes a JSON extraction in 23 tokens; another writes 300.

This dataset sends a fixed set of prompts to every model at temperature 0, every night, and records the cost computed from the token usage each provider actually reported. Because the question is identical, the difference is the model, not the workload.

Current results (window: 30 days, as of 2026-08-24)

Median spread across tasks: 43.8x. Clean runs in window: 425.

Two-sentence explanation of a concept (short_answer)

Cheapest: deepseek-v4-flash. The most expensive costs 43.8x more for the same question.

model	USD per call	tokens in / out	latency (ms)	runs
deepseek-v4-flash	0.000133	40 / 128	3399	23
gemini-3.5-flash	0.000625	97 / 71	6931	9
GLM5.2	0.000743	34 / 200	5712	15
gpt-5.6-luna	0.000771	77 / 148	4698	8
claude-sonnet-5	0.000968	41 / 103	13363	21
gpt-5.6-terra	0.001868	504 / 72	6165	3
claude-opus-4-8	0.002266	31 / 107	15776	15
claude-opus-5	0.002331	71 / 79	5181	7
Kimi-K3	0.002790	91 / 198	3842	21
gpt-5.5	0.003168	54 / 123	8967	3
grok-4.6	0.003747	228 / 549	17266	7
claude-fable-5	0.004079	30 / 96	11599	13
gpt-5.6-sol	0.005825	742 / 71	4576	2

Compress a technical passage into three bullets (summarize)

Cheapest: deepseek-v4-flash. The most expensive costs 31.9x more for the same question.

model	USD per call	tokens in / out	latency (ms)	runs
deepseek-v4-flash	0.000269	192 / 205	4611	22
gemini-3.5-flash	0.001092	176 / 122	6575	3
GLM5.2	0.001612	182 / 400	7736	16
gpt-5.6-luna	0.001721	233 / 320	7994	7
claude-sonnet-5	0.001830	243 / 161	13932	20
gpt-5.6-terra	0.003298	891 / 126	10765	3
claude-opus-4-8	0.004476	237 / 176	20408	15
claude-opus-5	0.004776	257 / 140	6243	7
grok-4.6	0.004895	374 / 691	15713	7
Kimi-K3	0.005446	240 / 374	6815	21
gpt-5.5	0.005964	203 / 215	10169	3
claude-fable-5	0.008582	234 / 168	14324	13

Extract fields, return JSON only (structured_json)

Cheapest: deepseek-v4-flash. The most expensive costs 104.1x more for the same question.

model	USD per call	tokens in / out	latency (ms)	runs
deepseek-v4-flash	0.000084	69 / 55	2636	23
gemini-3.5-flash	0.000302	54 / 33	5496	2
claude-sonnet-5	0.000376	81 / 27	11413	21
gpt-5.6-luna	0.000504	105 / 88	6628	8
claude-opus-4-8	0.000834	74 / 27	12785	14
GLM5.2	0.001096	62 / 292	5380	16
claude-opus-5	0.001097	112 / 22	16896	6
gpt-5.6-terra	0.001336	532 / 23	10363	3
claude-fable-5	0.001642	82 / 25	10845	15
Kimi-K3	0.001719	120 / 110	2925	21
gpt-5.5	0.002024	56 / 75	7042	2
gpt-5.6-sol	0.003570	532 / 30	12508	3
grok-4.6	0.008741	257 / 1371	25672	7

A routing artefact worth knowing about

Some requests arrive on a channel serving a large cached preamble: sending the single word hi comes back reporting ~4,400 prompt tokens, 3,840 of them cached. It is not stable — the same model and the same question sometimes hits it and sometimes does not, so it is a property of routing, not of the model. Those runs are excluded from the comparison above and reported here instead. They are still in probe_runs.csv (cached_prompt_tokens >= 1000), so you can check the exclusion rule yourself.

model	share of runs hitting it	USD when hit	USD when clean
gemini-3.5-flash	53%	0.005985	0.000679
gpt-5.5	72%	0.019670	0.003930
gpt-5.6-luna	32%	0.004081	0.000968
gpt-5.6-terra	57%	0.009686	0.002167
gpt-5.6-sol	76%	0.024795	0.004472

Method

• Identical prompt per task, temperature: 0, capped max_tokens. The prompts are in tasks.js and are deliberately timeless — they never change, so numbers stay comparable across months.
• Cost = tokens reported by the provider x the price actually paid. Responses without a usage field are discarded, not estimated — an estimate inside a "measured cost" dataset is a lie.
• One run per model per task per night; samples accumulate.
• probe_runs.csv contains every run, including failures and preamble hits, so the published averages can be recomputed from scratch.

What this does not tell you

Cost, not quality. Cheapest here says nothing about whether the answer is good. It also reflects these three prompts specifically — your prompts have a different input/output ratio, and that ratio is exactly what drives the difference. Use it as a starting point for bulk work, then measure your own.

Live version, updated nightly

• Human-readable, per-task pages: https://ainetcafe.com/costs
• JSON API, no key, CORS open: https://ainetcafe.com/api/model-costs
• Raw runs as CSV: https://ainetcafe.com/api/model-costs/raw.csv
• Callable from a bare URL by any agent: https://ainetcafe.com/t/model_costs?days=30

Citation

AI NetCafe real model cost dataset. https://ainetcafe.com/costs (CC BY 4.0)