AI API pricing, cut down to one unit
Every model below is priced for the same job — 10,000 input tokens and 2,000 output tokens per request, a thousand times over — so the column on the right is directly comparable. Rates were read from a live price pull on2026-11-18 00:50 UTC; nothing in this table was typed in by hand.
Cheapest in this set: GPT-OSS 20B at $0.360 per 1,000 requests. Most expensive: Claude Fable 5.2 at $160 — a444× spread for the same workload.
Cost per 1,000 requests, cheapest first
| Model | Vendor | $/1M in | $/1M out | $/1M blended (5:1) | $/1k requests | Context |
|---|---|---|---|---|---|---|
| GPT-OSS 20B | OpenAI (open weights) | $0.018 | $0.09 | $0.030 | $0.360 | 131k |
| Nova Micro | Amazon | $0.035 | $0.14 | $0.053 | $0.630 | 128k |
| DeepSeek V4.1 Flash | DeepSeek | $0.051 | $0.203 | $0.076 | $0.916 | 1.05M |
| GPT-5 Nano | OpenAI | $0.05 | $0.4 | $0.108 | $1.30 | 400k |
| Llama 4 Scout | Meta (open weights) | $0.1 | $0.3 | $0.133 | $1.60 | 1.31M |
| Gemini 2.5 Flash Lite | $0.1 | $0.4 | $0.150 | $1.80 | 1.05M | |
| Claude Haiku 5.5 | Anthropic | $0.1 | $0.5 | $0.167 | $2.00 | 1.00M |
| Qwen3.8 Flash | Alibaba Qwen | $0.15 | $0.47 | $0.203 | $2.44 | 1.00M |
| Mistral Small 2603 | Mistral | $0.15 | $0.6 | $0.225 | $2.70 | 262k |
| GLM-5.3 | Z.ai | $0.07 | $1.4 | $0.292 | $3.50 | 1.05M |
| MiniMax M3 | MiniMax | $0.3 | $1.2 | $0.450 | $5.40 | 1.05M |
| Llama 4 Maverick | Meta (open weights) | $0.35 | $1 | $0.458 | $5.50 | 524k |
| Command A+ | Cohere | $0.3 | $1.5 | $0.500 | $6.00 | 192k |
| Mistral Large 4 | Mistral | $0.7 | $2.1 | $0.933 | $11.2 | 1.05M |
| DeepSeek V4 Pro | DeepSeek | $0.9 | $1.5 | $1.00 | $12.0 | 1.05M |
| Grok 4.3 | xAI | $1.25 | $2.5 | $1.46 | $17.5 | 1.00M |
| Kimi K3.1 Pro | Moonshot | $0.24 | $8.402 | $1.60 | $19.2 | 1.05M |
| GPT-6.1 Terra | OpenAI | $1 | $5 | $1.67 | $20.0 | 1.05M |
| Qwen4 Max | Alibaba Qwen | $1.425 | $4.275 | $1.90 | $22.8 | 1.00M |
| Gemini 4 Hydrogen | $1.5 | $7.5 | $2.50 | $30.0 | 1.05M | |
| Grok 4.7 | xAI | $2 | $6 | $2.67 | $32.0 | 500k |
| GPT-6.2 Sol | OpenAI | $2 | $10 | $3.33 | $40.0 | 1.05M |
| Claude Sonnet 5.5 | Anthropic | $2 | $10 | $3.33 | $40.0 | 1.00M |
| Gemini 4 Carbon | $2 | $10 | $3.33 | $40.0 | 1.05M | |
| Nova Premier | Amazon | $2.5 | $12.5 | $4.17 | $50.0 | 1.00M |
| Claude Opus 5.5 | Anthropic | $4 | $20 | $6.67 | $80.0 | 1.00M |
| GPT-6.2 Astra | OpenAI | $8 | $40 | $13.3 | $160 | 1.05M |
| Claude Fable 5.2 | Anthropic | $8 | $40 | $13.3 | $160 | 1.00M |
Sorted by cost per 1,000 requests. Model ids, input and output rates and context windows come straight from the price pull; the blended and per-request columns are computed from those rates.
Run your own workload
The table assumes a fixed request shape. Change it to your own traffic — everything is computed in the browser, nothing is sent anywhere.
| Model | Monthly cost | Per 1,000 requests |
|---|
What actually moves the bill
- Output tokens cost more than input tokens — often 3–6× on the same model. A table that only advertises the input rate hides most of the invoice.
- Reasoning models bill invisible tokens. A "thinking" model can emit thousands of output tokens you never display; budget for it at the output rate.
- Batch and cache tiers are separate prices. Batch endpoints commonly halve the rate, and prompt caching discounts the repeated prefix — neither is in the list price above.
- Context length is a price risk, not a feature. Long-context surcharges are billed per token beyond a threshold on several providers.
Method and limits
Source: OpenRouter's public model API, pulled2026-11-18 00:50 UTC. Rates are USD per million tokens for the model as served, before any platform fee on credits. The comparison assumes no caching, no batch discount, no reasoning overhead and no retries. Where a vendor publishes several variants of a family (pro / max / mini), only the variant shown is priced — the family name alone is not a price.
Questions
- What does "cost per 1,000 requests" mean on this table?
- It is the same request priced across every model: 10,000 input tokens plus 2,000 output tokens, repeated 1,000 times. That is 10M input and 2M output tokens. It replaces the guesswork of comparing an input rate against an output rate: one column, one currency, one workload.
- Are these the prices the providers charge directly?
- They are per-model list rates read from OpenRouter's public model API on 2026-11-18 00:50 UTC, which mirrors each provider's published list price for that model. Direct provider contracts, committed-spend discounts, provisioned throughput and batch tiers can differ, and OpenRouter adds a fee on credit purchases. Treat the table as a comparison of list prices, not as a quote.
- Why can the same model differ in price between vendors?
- Open-weight models such as Llama 4 or GPT-OSS are served by several hosts, and each host sets its own rate, quantisation and context limit. A single "cheapest Llama" figure is meaningless without naming the host, which is why the table lists the model id as served rather than the family name alone.
- How often is this rebuilt?
- The table is regenerated from a fresh price pull rather than edited by hand; the pull behind this version is dated 2026-11-18 00:50 UTC. Prices on frontier models have moved several times within a single quarter, so a table without a check date should be treated as decoration rather than data.
Found a rate that has moved since 2026-11-18 00:50 UTC? Send the provider's own pricing URL through thecontact page and the row gets re-checked. Figures on this page are list prices, not a quote, and AIListScout has no affiliate relationship with any model provider listed.
Frequently asked questions
What does "cost per 1,000 requests" mean on this table?
It is the same request priced across every model: 10,000 input tokens plus 2,000 output tokens, repeated 1,000 times. That is 10M input and 2M output tokens. It replaces the guesswork of comparing an input rate against an output rate: one column, one currency, one workload.
Are these the prices the providers charge directly?
They are per-model list rates read from OpenRouter's public model API on 2026-11-18 00:50 UTC, which mirrors each provider's published list price for that model. Direct provider contracts, committed-spend discounts, provisioned throughput and batch tiers can differ, and OpenRouter adds a fee on credit purchases. Treat the table as a comparison of list prices, not as a quote.
Why can the same model differ in price between vendors?
Open-weight models such as Llama 4 or GPT-OSS are served by several hosts, and each host sets its own rate, quantisation and context limit. A single "cheapest Llama" figure is meaningless without naming the host, which is why the table lists the model id as served rather than the family name alone.
How often is this rebuilt?
The table is regenerated from a fresh price pull rather than edited by hand; the pull behind this version is dated 2026-11-18 00:50 UTC. Prices on frontier models have moved several times within a single quarter, so a table without a check date should be treated as decoration rather than data.