Reference
Every major provider's published list price, input and output, side by side — Frontier, Balanced, and Fast tiers, in one table.
At a glance
Grouped by tier, cheapest input first. The lowest input and output price in each tier is marked.
| Provider | Model | Tier | Input $/1M | Output $/1M | Context | Notes |
|---|---|---|---|---|---|---|
| Gemini 3.1 Pro Preview | Frontier | $2.00lowest | $12.00lowest | 1M | preview pricing, ≤200K token prompts | |
| Anthropic | Claude Opus 4.8 | Frontier | $5.00 | $25.00 | 1M | — |
| OpenAI | GPT-5.5 | Frontier | $5.00 | $30.00 | 1.1M | — |
| Gemini 3.5 Flash | Balanced | $1.50lowest | $9.00lowest | 1M | — | |
| OpenAI | GPT-5.4 | Balanced | $2.50 | $15.00 | 1.1M | — |
| Anthropic | Claude Sonnet 5 | Balanced | $3.00 | $15.00 | 1M | intro pricing $2 / $10 thru 2026-08-31 |
| Gemini 3.1 Flash-Lite | Fast | $0.25lowest | $1.50lowest | 1M | — | |
| OpenAI | GPT-5.4 mini | Fast | $0.75 | $4.50 | 400K | — |
| Anthropic | Claude Haiku 4.5 | Fast | $1.00 | $5.00 | 200K | — |
How it works
Every provider bills input and output separately, and at different rates. Input is everything you send — the prompt, system instructions, retrieved documents, tool results, prior turns in a conversation. Output is only what the model generates back. Providers price output several times higher than input because generation is the more expensive half of the computation, so a chatty, verbose model reply costs more than a terse one even at identical input size.
The context window is the maximum combined input + output size a model can handle in one request, measured in tokens (roughly ¾ of a word each). A bigger window lets you stuff in more history, more retrieved documents, or a larger codebase before the model loses track of earlier content — but it doesn't change the per-token price, so a huge context window only costs more if you actually fill it.
Every major provider ships roughly three tiers. Frontier models are the most capable and most expensive — reasoning-heavy or high-stakes tasks. Balanced models trade a little quality for a meaningfully lower price and are the default choice for most production traffic. Fast (sometimes called "mini" or "lite") models are cheapest and quickest, built for high-volume, low-complexity work like classification, extraction, or simple chat turns. Routing easy requests to a Fast model and reserving Frontier for the requests that need it is one of the highest-leverage AI cost controls available.
If your requests repeat the same large system prompt, tool definitions, or reference document, prompt caching lets the provider skip re-processing that prefix and charge a small fraction of a normal input-token rate for the cache read instead. It only helps the input side, and only on the portion of the prompt that stays identical between calls — but on a workload with a stable prefix and a high cache-hit rate, it can meaningfully cut the input line of your bill. Exact caching mechanics and discount rates vary by provider and change often, so this page keeps the discussion qualitative — check the provider's own pricing page for current cache rates.
Where a row below has a note, it's usually flagging non-standard pricing: an intro or preview rate that reverts to a higher published price after a stated date, or a context-length cap that only applies above a certain prompt size. Always check the note before budgeting off a listed price — "intro pricing" in particular is a temporary number, not the steady-state rate.
Worked example
A representative chat turn — 3,000 input tokens, 500 output tokens — costed against every model, per 1,000 requests.
| Model | Tier | Cost per 1,000 requests |
|---|---|---|
| Gemini 3.1 Pro Preview | Frontier | $12.00 |
| Claude Opus 4.8 | Frontier | $27.50 |
| GPT-5.5 | Frontier | $30.00 |
| Gemini 3.5 Flash | Balanced | $9.00 |
| GPT-5.4 | Balanced | $15.00 |
| Claude Sonnet 5 | Balanced | $16.50 |
| Gemini 3.1 Flash-Lite | Fast | $1.50 |
| GPT-5.4 mini | Fast | $4.50 |
| Claude Haiku 4.5 | Fast | $5.50 |
Running your own volume, mixed models, or prompt caching? The LLM cost calculator computes this live against your actual monthly tokens. Running a multi-step agent instead of single calls? See the AI agent cost guide for why cost compounds across steps.
FAQ
Pricing is per-token and varies by tier. As of 2026-07-08, Frontier models run roughly $2–$5 per million input tokens and $12–$30 per million output tokens; Balanced models are noticeably cheaper; Fast/lite models are cheapest of all, often under $1 per million input tokens. See the full table above for exact, current figures per model.
Generating tokens is more computationally expensive than reading them — the model has to run a forward pass for every output token it produces, one at a time, versus processing the entire input in a single pass. Every major provider reflects that asymmetry in price, typically charging 4–6x more per output token than per input token.
It depends on the tier you're comparing and your input/output mix, since the cheapest model changes tier by tier and the table above marks the lowest input and output price in each. For an exact number against your own monthly token volume — including the effect of prompt caching — use the LLM cost calculator, which recomputes every model's cost live against your usage.
List prices move more often than most infrastructure pricing — new model releases, intro/preview discounts expiring, and periodic rate changes are all common. This page's figures were verified as of 2026-07-08; always cross-check against each provider's official pricing page (linked above) before budgeting off a number here.
This page and the calculator both work from published list prices — useful for planning, but not your real bill, which depends on your actual usage, any negotiated rates, and volume discounts. CostMon connects directly to your provider's cost API (starting with the Anthropic Admin cost report) and normalizes real AI spend into the same dashboard as your cloud and SaaS costs, so you see what you're actually being charged, daily, not a monthly total that arrives too late to act on.
Connect your first providers in minutes and give finance and engineering one number they both trust. Flat pricing — never a percentage of your bill or your savings.