📊 LLM Pricing Comparison
Every major model side by side, priced per million tokens, taken from each provider's own pricing page. Enter the volume you actually expect and the table sorts by what it would cost you rather than by headline rate.
Reading an AI price list without being misled by it
Every provider quotes a price per million tokens, which makes the models look comparable. They are not quite. The same headline rate can mean different things depending on how output is priced, whether caching is available, whether the figure is promotional, and whether a tier switch is waiting above a certain prompt length.
The table above sorts by what a workload would actually cost you rather than by headline input rate, because the input price alone is close to meaningless when output costs three to eight times as much.
The four things a headline price hides
- The input-to-output ratio. Grok charges 3x input for output; some OpenAI tiers charge 8x. Two models with the same input price can differ by a third on a generation-heavy workload.
- Tiers. OpenAI prices the same model four ways, from Flex to Ultrafast, a six-fold range. The model name alone does not tell you what you will pay.
- Thresholds. Claude Haiku 5.5 costs five times more above 100,000 prompt tokens, and the threshold counts cache reads too, so a cached long prompt can quietly cross it.
- Expiry. Several Gemini prices are promotional with a published end date. Budgeting against them without noting the date is how a forecast goes wrong by 100%.
Prices here were checked on 2026-10-26 against each provider's own pricing page. Models marked with an arrow have an increase already announced. This is a list that goes stale: treat any AI price older than a quarter as a rough guide.
Cheap models are not a compromise any more
The spread on this table is extreme — roughly a hundredfold between the cheapest and dearest input rate. The gap in capability has not widened at anything like the same pace, which makes model selection by far the biggest cost lever available. Classification, extraction, routing and summarising rarely need a flagship; drafting, reasoning and code review often do.
The practical approach is to route by task rather than pick one model for everything. Run the cheap model first, check the result against a simple rule, and escalate only what fails. A system that sends 90% of traffic to a model costing a twentieth as much pays for the routing logic in its first week.
Discounts worth knowing
| Mechanism | Typical saving | Catch |
|---|---|---|
| Prompt caching | 90–97.5% on cached input | Write premium of 1.25x on first use |
| Batch API | 50% on input and output | Results within hours, not seconds |
| Off-peak (DeepSeek) | 50% | Tied to specific UTC hours |
| Flex tier (OpenAI) | 50% | Lower priority, slower |
None of these require changing model or prompt. Anything that is not latency-sensitive — overnight processing, bulk classification, content generation for later review — should be on a batch or flex tier, and that is usually a one-line change.
A worked example
A pipeline processes 10 million input and 2 million output tokens a month. On Claude Fable 5.1 that is $100 plus $100, so $200. On Sonnet 5.5 it is $20 plus $20, so $40. On GPT-6 Luna it is $1 plus $1, so $2. Put the bulk classification on Luna and the final review on Sonnet, and a bill that was $200 becomes closer to $10 — without touching the prompts.
Worth pairing with the ai token calculator, which uses the same inputs for a different question, and with the ai conversation cost for the step after.
Frequently Asked Questions (FAQs)
Which LLM is cheapest?
On input rate, the lightest tiers from OpenAI and Anthropic are the cheapest here at $0.10 per million. But total cost depends on your output ratio, so sort by your own volume rather than by headline price.
Why is output priced higher than input?
Input is processed in parallel in a single pass, while output must be generated one token at a time. The ratio ranges from 3x on Grok to 8x on some OpenAI tiers.
What is prompt caching worth?
Between 90% and 97.5% off the input rate for the cached portion, after a write premium of about 1.25x on the first call. It pays back within two or three reuses and is the largest routine discount available.
Are batch prices always 50% off?
On the providers listed here, yes, for both input and output. The trade is latency: results arrive within hours rather than seconds, which suits anything processed overnight or in bulk.
How often do these prices change?
Often enough that a quarterly recheck is the minimum. Models get repriced, launched and retired within months, and some prices on this page are promotional with a published expiry date.
Should I use one model for everything?
Rarely. The hundredfold spread in price is much wider than the spread in capability for simple tasks, so routing cheap work to a cheap model and escalating only what fails is usually the single biggest saving available.
Do these prices include tax?
No. They are list prices in US dollars before any local sales tax, VAT or withholding, and before any committed-use or enterprise discount you may have negotiated.