Same support bot: $10 or $965 a month.
$10 a month on GPT-6 Luna, $965 on GPT-6 Astra: a 100x spread. Same workload, 100,000 support chats a month (650 input and 63 output tokens each), at each provider's official list price checked 2026-11-12. Quality is not equal: the Artificial Analysis Intelligence Index scores these two at 21 and 55, so the cheapest model is not a like-for-like swap.
Prices checked 2026-11-12 · 5 providers · 16 models · token counts from 1,400 real calls · method · prices.json · raw CSV
What the same work costs across providers
Pick a workload and volume. Monthly cost is calls times tokens times each provider's official list price, the way providers bill. Up and to the left is better.
Assumed per call: Token counts per workload come from our own run of 1,400 real calls; other vendors' tokenizers count differently, and reasoning modes add hidden output tokens (in that run, reasoning variants used 1.3 to 8.8 times the output tokens), which the multiplier approximates.
Prices: each provider's official pricing page (OpenAI, Anthropic, Google, DeepSeek, Qwen), checked 2026-11-12. Quality: Artificial Analysis Intelligence Index (general, not task-specific). Prices change: Gemini 4 Carbon and Gemini 4 Argon double on 2026-12-01, Gemini 3.8 Flash on 2027-01-01; DeepSeek is shown at peak-hour prices (off-peak is half).
How much would you save?
Pick the model you use now, describe your workload mix and volume, and set how many Artificial Analysis index points you can give up. It finds the cheapest model from any provider above that stays inside that band.
Batch mode leaves out models whose provider prints no batch discount. Savings assume your calls look like the four workloads above; your own traffic can differ, which is exactly what a bill audit tests.
Method and raw data
Monthly cost = calls per month x (input tokens x input price + output tokens x output price), at each provider's official list price, which is how providers bill. Token counts per workload (support bot, RAG answers, classification, summarisation) come from our own run of 1,400 real calls on fictional business data; the raw results are in raw_results.csv. Tokenizers differ by provider, so another vendor's count for the same prompt will differ a little.
Prompt caching assumes the fixed system prompt is a cache hit on every call, and batch applies the provider's printed 50% discount; both are best-case figures. Quality is the Artificial Analysis Intelligence Index, a general reasoning, knowledge and coding index that is not specific to these workloads; the reasoning-effort variant is shown per row. Prices were checked 2026-11-12 and change: Gemini 4 Carbon and Gemini 4 Argon double on 2026-12-01, Gemini 3.8 Flash on 2027-01-01.
Want this on your traffic, not ours?
On a free 15-minute call, send us last month's LLM invoice split by model and we will tell you where it leaks and what a quality-checked cut looks like. If it is worth doing properly, the full audit runs on an eval built from your own requests.
Book a 15-min call