Estimate per-request, monthly, and annual API costs from official prices, and check that the request fits the context window.
No API key · Latest fact review 2026-10-25; dates vary by field
Token prices for 38 models come from AI Compute Tracker's reviewed table (opens in a new tab), reviewed 2026-10-24, a companion site owned and operated by the same team as AI Wiki; the other models and fields come from this site's registry.
A planning estimate, not a quote or invoice. Check the selected model's effective dates and caveats, and confirm the current provider price before committing a budget.
This tool uses manually reviewed pricing assumptions with the dates shown below. For the separately refreshed price snapshot and named intelligence benchmarks, see API prices on the front page. Check the source and rate configuration before using an estimate.
Select models to compare, estimate costs, and check deployment options.
Prices checked Oct 24, 2026
Calculated in your browser. Settings are included in the page URL and shared between tools: anyone with that URL can see them, and opening it sends the URL to AI Wiki. Do not enter confidential workload details.
Presets (illustrative defaults, not provider guidance)
Per request
$0.0505
22K total tokens
Per month
$1,111.00
22K requests (1,000 x 22 working days)
Per year
$13,332.00
12 equal months
Cache writes bill at $2.50 per 1M (30-minute TTL) / $5.00 (12-hour) and are not modeled here. OpenAI bills prompt-cache writes at 1.25x the uncached input rate.
22K combined input + requested output
Context: 1.05M
Max output: 128K
1.02M context tokens remain before the published window limit.
Cached input is billed at 5 percent of the input rate, half of GPT-6 Sol's cached price. Flex is priced like Batch, Fast mode costs 2x Standard, and regional processing adds 10 percent where available.
Does not accept the none or minimal reasoning efforts, and tool calling requires the Responses API (Chat Completions works without tools).
Ranked comparison
Sorted by estimated monthly cost. Select a row to inspect its rates and constraints above. Deprecated and retired models are hidden until you widen the lifecycle filter.
Transparent methodology
Cached tokens are carved out of total input, and only that subset receives a recorded cache rate. Batch combines with cached-input rates where the provider publishes a batch cached rate; otherwise the cached share bills at the batch input rate. Published cache-write fees are disclosed, not silently added.
Input cost and output cost form a per-request estimate. Requests per day multiplied by working days produces the monthly volume, and announced price changes blend into the monthly and yearly projections.
Input plus requested output is compared with the context window, and crossing a documented long-context threshold moves the whole request to the provider's tier rates. Output is also checked independently when the provider publishes a separate maximum.
Common questions
The calculator multiplies the selected model's published per-million-token input and output rates by your tokens per request, then multiplies the request estimate by requests per working day and working days per month. When a provider has announced a dated price change, such as introductory pricing with a published end date, the monthly and yearly projections blend the current rates with the announced successor rates across that horizon; the per-request figure always uses today's rates. For providers that publish separate peak and off-peak prices, the rates blend by the share of traffic you send in peak hours.
No. Cached input is treated as a subset of the total input. The cached subset receives a published cached-input rate when one is recorded; otherwise it receives the ordinary input rate. Cache reads are not the whole story: several providers also bill to write cache entries, so the calculator shows the model's published cache-write prices beside the savings row without adding them to the estimate. DeepSeek publishes cache-hit and cache-miss input prices rather than a write fee, and the tool labels its savings row accordingly.
No. Batch mode appears only when the model registry contains official Batch input and output rates. The calculator does not assume a percentage discount or infer missing prices. Batch combines with cached-input rates where the provider publishes a batch cached-input rate, as Anthropic, OpenAI and Google do; for other models the cached share bills at the batch input rate. Documented long-context tiers still apply in Batch mode, scaling the published batch rates by the provider's tier multipliers.
No. It covers the recorded text-token prices, documented long-context tiers, and announced successor rates. Taxes, regional pricing, minimum commitments, cache-write and cache-storage fees, searches, grounding, tools, fine-tuning, media, partner platforms, and other feature charges may change the actual bill.