Check your real LLM API cost from the usage response: catches the 272K-token price cliff, near-zero cache hit rate, and batch/flex/fast mode multipliers. Astra, DeepSeek and GLM presets, or your own rate card. Private, in-browser.
Updates automatically as you type.
Calculated in your browser; your inputs are not sent to us. Privacy
Results will appear here.
✓Basic automated checks passed. Formula accuracy is not established by these checks. See check scope and results
Calculation details
Every figure above is calculated locally in your browser from the assumptions shown. No inputs are sent anywhere. See the methodology section below for how this tool works.
Use this resultShare the current inputs, or ask ChatGPT to explain the calculation in context.
Ask an AI to explain this result
Share this result
Planning estimate only, not financial, tax, or legal advice. Verify assumptions and current rules before making decisions.
This tool runs in your browser. Your calculator inputs and results are never transmitted to us or to ad/affiliate partners. Ads and sponsored links may set third-party cookies to serve and measure them, but they never receive your calculation values. If you explicitly save a scenario, its permitted fields stay in local browser storage until you clear them. See our Privacy Policy.
Use this result
Share the current inputs or ask ChatGPT to explain the calculation in context.
No unrelated authority link is added merely to make this page look sourced. Tool-specific external sources are shown only when they directly support the method, rate, or rule.
Assumptions to Verify
Inputs are treated exactly as entered.
Verify important decisions with current authoritative information.
Privacy: The calculation runs in this browser. Shared links contain the page address only, not your inputs or results.
A worked example: crossing the 272K cliff on GPT-6 Astra
Take one request with 100,000 fresh input tokens, 200,000 cached input tokens, and 2,000 output tokens on the Astra preset ($10/M fresh, $1/M cached, $50/M out). Total input is 300,000 tokens, which is 28,000 past the 272,000 cliff, so the whole request reprices: input x2, output x1.5.
Fresh input: 100,000 / 1,000,000 x $10 = $1.00, doubled to $2.00. Cached input: 200,000 / 1,000,000 x $1 = $0.20, doubled to $0.40. Output: 2,000 / 1,000,000 x $50 = $0.10, times 1.5 = $0.15. The bill is $2.55. Without the cliff the same request costs $1.00 + $0.20 + $0.10 = $1.30, so the penalty is $1.25, nearly doubling a single call.
The per-token cliff rate is brutal at the margin: those 28,000 excess tokens did $1.25 of damage, about $44.6 per million tokens of overflow versus the $10/M sticker rate. Retrieving or compacting to pull input back under 272,000 saves more than doubling the model's output budget would.
Where estimates drift from the invoice
Cached-token accounting is the biggest source of error. Providers bill cache reads only on exact prefix matches, and interactive apps usually show cache hit rates of 60 to 90 percent, while a naive loop that regenerates a timestamp or session id inside the system prompt can sit at 0 percent. On DeepSeek V4 Flash ($0.28/M fresh vs $0.028/M cached), a 90 percent hit rate on 200,000 input tokens costs $0.0112; a 0 percent hit rate costs $0.056, a 5x difference on every call.
Rate cards move. The Astra figures here were verified in September 2026; the DeepSeek and GLM presets are marked indicative in the tool itself and should be confirmed against the pricing page before any monthly budget is committed. Providers have cut cached rates by 50 percent or more within a single quarter.
The calculator prices one request, not a conversation. Multi-turn sessions resend the full history, so call ten on a 40,000-token context bills 40,000 input tokens again; multiply the per-request result by expected turns, subtract nothing, and treat the sum as a ceiling. Finally, real invoices include taxes, minor rounding differences, and occasionally failed-but-billed requests, so expect a few percent of slack above any estimate.
