In short
For 10 million input and 2 million output tokens a month, current list prices run from about $2 on budget models to $200 on top-tier ones, a 100-fold spread. Flagship everyday models cluster around $40. Prompt caching and batch processing can reduce costs substantially, and some prices are about to move: Gemini 4 Argon is scheduled to double on 1 December 2026.
Current list prices
All prices are in US dollars per million tokens, standard tier, from vendor pages checked on 4 November 2026. This is a selection of commonly used models, not a complete list. New models appear every few weeks, so treat the vendor pricing pages as the source of truth.
The right-hand column shows the monthly cost of a reference workload: 10 million input tokens and 2 million output tokens, with no discounts.
| Model | Input | Output | Reference workload |
|---|---|---|---|
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | $1.80 |
| GPT-6 Luna | $0.10 | $0.50 | $2.00 |
| Claude Haiku 5.5 (prompts up to 100K tokens) | $0.10 | $0.50 | $2.00 |
| DeepSeek Flash (off-peak) | $0.15 | $0.60 | $2.70 |
| DeepSeek Flash (peak) | $0.30 | $1.20 | $5.40 |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | $5.50 |
| Mistral Large | $0.50 | $1.50 | $8.00 |
| DeepSeek V4 Pro (off-peak) | $0.66 | $1.98 | $10.56 |
| Gemini 3.8 Flash | $0.75 | $3.75 | $15.00 |
| GPT-6.1 Terra | $1.00 | $5.00 | $20.00 |
| GPT-6.1 Sol | $2.00 | $10.00 | $40.00 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $40.00 |
| Gemini 4 Argon | $2.00 | $10.00 | $40.00 |
| Claude Opus 5.5 | $4.00 | $20.00 | $80.00 |
| GPT-6 Astra | $10.00 | $50.00 | $200.00 |
| Claude Fable 5.1 | $10.00 | $50.00 | $200.00 |
Other sources: DeepSeek and Mistral.
Price changes already scheduled
Google's page lists dated increases, which means a price you test today may not be the one you pay at launch:
- Gemini 4 Argon: $2/$10 through 30 November 2026, then $4/$20 from 1 December. The reference workload goes from $40 to $80.
- Gemini 3.8 Flash and 3.6 Flash: $0.75/$3.75 through 31 December, then $1.50/$7.50 from 1 January 2027. The reference workload goes from $15 to $30.
Google also says Gemini 2.5 Pro and 2.5 Flash are deprecated, with requests routed to newer models from 1 December 2026, so check any code that still names them. DeepSeek's rates vary by time of day: its peak period covers 01:00–04:00 and 06:00–10:00 UTC on weekdays, and off-peak is half the peak rate.
Why the sticker price misleads
Output costs more than input. Generally 3–5 times more, as the table shows, so summarising long documents and generating long answers are very differently priced. Work out your own input-to-output ratio.
Caching can change everything. If each request repeats a long system prompt or document, cached input is far cheaper. Listed cached rates include $0.10 per million on GPT-6.1 Sol and on Claude Sonnet 5.5, compared with $2 standard; DeepSeek's Flash goes as low as $0.003 off-peak. Write fees and expiry rules differ by vendor, so check how yours works. For repetitive workloads, this may matter more than which model you choose.
Batching halves many prices. Google, Mistral and Anthropic list half-price batch rates, and OpenAI offers Flex and Batch pricing. If a job does not have to finish immediately, such as nightly classification, always use it.
Tools cost extra. Search and other tools can carry their own charges. Google lists 5,000 free grounded search requests a month, then $14 per 1,000. DeepSeek charges separately for search and fetch.
Reasoning counts. On Gemini, Google's listed output price includes thinking tokens. Models that reason longer may produce more billable output than the visible answer suggests.
Cheap per token is not cheap per result. If a budget model needs a retry or human correction, the saving may vanish. Measure cost per successful task, not per token. How to evaluate an LLM feature explains how.
How to estimate your own bill
- Measure a real request. Log actual input and output tokens across a few typical examples. See what a token is if you are starting out.
- Multiply by volume, using requests per user and the number of users you expect.
- Split out what's cacheable, such as system prompts and shared documents.
- Add 20–30% as a working margin for retries and growth. That margin is our rule of thumb, not an industry standard.
- Set a hard monthly limit with your provider before going live.
How to choose
- High-volume, simple tasks: classification, extraction and routing usually run well on the budget tier. Test Luna, Haiku, Flash-Lite or DeepSeek Flash first.
- General product features: start in the $20–$40 band, and move down if quality holds.
- Hard reasoning or coding: pay for the top tier only on the requests that need it.
- Most real products: use several models. Route easy requests to a cheap model and escalate difficult ones. This often saves more than any negotiation.
Price is only one factor. Data handling, rate limits, latency and reliability also matter, particularly with customer data. Read how to choose an LLM for your product and how to control AI API costs. If you are thinking of running models yourself, compare with the cheapest ways to run an LLM.
Want help designing an AI feature that won't surprise you on cost? Book a call.
Common questions
Which LLM API is the cheapest?
On current list prices for this workload, Gemini 2.5 Flash-Lite, GPT-6 Luna and Claude Haiku 5.5 are among the lowest, around $2 for 10 million input and 2 million output tokens. The cheapest useful model depends on your task, so test quality as well as price.
Why is output priced higher than input?
Generating tokens uses more compute than reading them, and vendors price accordingly. In the models listed here, output typically costs three to five times as much as input. That makes long responses a major driver of cost, so limit output length where you can.
How much can prompt caching save?
It can save a lot on repetitive workloads. Listed cached input rates can be a small fraction of standard rates, such as $0.10 against $2 per million on some models. Actual savings depend on how much of each prompt repeats and on each vendor's rules.
How do I keep LLM costs under control?
Measure tokens per request, route simple work to cheaper models, cache repeated context, use batch pricing for non-urgent jobs, and limit output length. Set a hard spending limit with your provider and review cost per successful task regularly.