Short answer: OpenRouter bills a request with the pricing fields attached to the resolved model route: prompt, completion, cache read, cache write, web search, image output, and any prompt-size override. Multiply the published per-token strings by one million to compare against vendor price sheets, pin the route identity in logs, and treat batch or cache variants as separate products rather than discounts you can assume.
| Route | Prompt | Completion | Why it matters |
|---|---|---|---|
| qwen/qwen3.8-max-0902chat, 1M context | $2.00 | $6.00 | 1M-token context with a 1/8 cache-read rate. |
| anthropic/claude-sonnet-5messages, 1M context | $2.00 | $10.00 | Cache reads at $0.20 and cache writes at $2.50. |
| openai/gpt-5.4responses, 1.05M context | $2.50 | $15.00 | Prompts at or above 272,000 tokens move to $5.00/$22.50. |
| moonshotai/kimi-k3chat, 1.05M context | $0.192 | $16.00 | Cheap input, expensive output: an 83x ratio. |
| z-ai/glm-5.2chat, 1.05M context | $0.007 | $6.80 | The largest input/output asymmetry in this set. |
| openai/gpt-oss-120bchat, 131k context | $0.15 | $0.60 | Small context window keeps long-document use off this route. |
What a router actually bills
A router sits between your client and several upstream providers. The model identifier you send selects a catalog entry, and that entry carries the price fields your account is charged against. According to the OpenRouter model API retrieved on November 13, 2026, the fields are prompt, completion, input cache read, input cache write, web search, and image output. Each is expressed as a price per single token, so a string like 0.000002 reads as $2.00 per million tokens.
That detail matters because most provider sheets publish per-million rates while router APIs publish per-token decimals. A decimal mistake of one or two zeroes turns a $2.00 input rate into $20 or $0.20. Convert deliberately: multiply by 1,000,000, write the result down beside the route ID, and use the converted number everywhere in the budget. The catalog also includes variant identifiers such as :batch and family members such as gpt-5.4-mini, so two rows that look similar in a spreadsheet can be different products.
The cache-read fields retrieved for the same six routes were $0.25 (Qwen3.8 Max), $0.20 (Claude Sonnet 5), $0.25 (GPT-5.4), $0.182 (Kimi K3), $0.006 (GLM-5.2), and $0.075 (GPT-OSS 120B) per million tokens. Those are separate meters, not a discount you can apply to the prompt rate at will; the sections below show when they change a budget.
- Prompt and completion are separate meters; never average them into one blended number before you know your token mix.
- Cache read and cache write are separate fields, not a single discount percentage.
- Overrides are attached to the route, so a prompt that grows past a threshold can reprice the whole request.
- Variant suffixes are real routes with their own context limits and price rows.
The route is part of the price
When a model is served by more than one upstream, the router may select among them by availability, latency, or your routing preferences. The visible model name stays stable while the underlying provider, quantization, context ceiling, and price can change. A benchmark that records only the model name therefore cannot explain a cost or quality shift later.
Log the resolved provider and the exact route identifier with every request you sample. If you need reproducibility for a launch, pin the provider or preserve routing metadata; otherwise a cheaper upstream entering the pool can silently change both the invoice and the output distribution. This is also why two teams quoting 'the same model' can report costs that differ by a factor of three without either being wrong.
Overrides reprice a growing prompt
OpenRouter exposes prompt-size overrides as structured pricing. GPT-5.4 is the clear example in this set: below 272,000 prompt tokens it lists $2.50 input and $15.00 output, and at or above that threshold it lists $5.00 input and $22.50 output — double on input and 1.5x on output. Qwen3.8 Max keeps a flat $2.00/$6.00 on the main route but its :batch variant has a 262,144-token context window, so the batch discount is unavailable to the longest documents unless the workload is restructured.
The operational consequence is a bimodal budget. A retrieval pipeline that normally sends 180,000 tokens can cross the threshold on a handful of long documents and reprice those requests entirely. Compute the cost of the tail, not just the median: sort real request sizes, price the bucket below the threshold at one rate and the bucket above it at the other, and check whether truncation, summarization, or a different route is cheaper than the override.
Cache and batch are separate products
Cache reads are typically a fraction of the prompt rate: $0.25 against $2.00 on Qwen3.8 Max, $0.20 against $2.00 on Claude Sonnet 5, and $0.182 against $0.192 on Kimi K3. Those ratios look like an easy 87–90% discount, but they apply only to the tokens that hit a cache the provider recognizes. Cache writes, prefix stability, and eviction rules decide whether the discount ever materializes. A cache read rate on a prefix that changes every request is a number you will never see on an invoice.
Batch variants are similarly conditional. Claude Sonnet 5 at $1.00/$5.00 is half the interactive rate, and GPT-5.4 batch at $1.25/$7.50 follows the same logic. Batch requires asynchronous completion and a workload that can wait. Using a batch rate in an interactive chat estimate does not save money; it produces a forecast that the first month of production usage will contradict.
A worked monthly estimate
For example, take 120,000 requests per month with 1,800 input tokens and 400 output tokens, of which 35% of input tokens are served from cache. In APIDir's analysis, on Qwen3.8 Max the uncached input is 140.4M tokens at $2.00 ($280.80), cached input is 75.6M at $0.25 ($18.90), and output is 48M at $6.00 ($288.00), for $587.70 per month and $4.90 per thousand requests. Without the cache the same workload costs $720.00, so the computed cache saving is $132.30.
Now price the identical workload on Kimi K3: uncached input 140.4M at $0.192 ($26.96), cached 75.6M at $0.182 ($13.76), output 48M at $16.00 ($768.00), total $808.72. The input rate is about 10x cheaper than Qwen's, yet the bill is 38% higher because output dominates. The calculator below lets you replace these assumptions with your own token mix; the answer changes fastest when you change output tokens, so measure those first.
What to record before you commit
Record the exact route ID, the resolved upstream provider, prompt and completion usage per request, cache read and cache write fields, the model's context ceiling, and the date you retrieved the price. Keep a provider invoice sample beside the estimate so a pricing-field change is visible rather than inferred. Recheck the model API before any budget review, and treat a rate older than its review window as unverified even when the page still loads.
If your product cannot tolerate a mid-month repricing, put the budget guard at the request level: cap output tokens, cap agent steps, and log requests that cross a context threshold. Those three controls bound the two meters that move the bill, and they are cheaper to implement than a reconciliation process after the invoice arrives.
Model this route on your token mix
Enter the rates you just retrieved for one exact route. The estimate applies uncached input, cache-hit input, output, and batch share separately; it is a planning calculation, not a provider quote.
Planning result
Review the example inputs, replace each rate with the exact current value from the linked source, then calculate. APIDir does not send these values to a provider.
Questions developers ask
Is OpenRouter cheaper than calling a provider directly?
Not as a rule. The router adds access and routing, but its price for a given route may be lower, equal, or higher than a direct account depending on the upstream and the deal. Compare the exact route identity, context ceiling, cache terms, and rate limits, not just the headline number.
Why does the same model show different prices?
Model families contain multiple versions, providers may serve the same name at different rates, and variants such as :batch carry different meters and context limits. Copy the full route identifier rather than the display name.
Do cache discounts apply automatically?
Only when the provider recognizes a cacheable prefix and your request matches its cache rules. Measure the cache-read fields on real responses before budgeting a discount.
What happens above the context override threshold?
The route's override pricing applies to the request. GPT-5.4, for example, doubles input and raises output to $22.50 per million once the prompt reaches 272,000 tokens.
Sources and check dates
Every rate in this guide was retrieved from the linked page on Nov 13, 2026. Prices can change; recheck the source before purchasing.