Overview
Compare what two language models would cost for your workload, side by side. Set tokens per call and calls per day, pick two models from a price board of major providers, add a prompt-caching assumption, and the page shows cost per call, per day, per month and per year for both, plus the annual saving from switching. A copyable markdown report summarises the comparison. All arithmetic runs locally.
How It Works
Pick Model A (the baseline) and Model B (the alternative) from the two dropdowns; each option shows its input and output price per million tokens and its context window. Enter input tokens per call, output tokens per call and API calls per day, or use a volume preset such as Hobby prototype, Growing SaaS or Enterprise RAG. Slide the prompt-cache hit ratio if you use caching. Three cards compare monthly cost, annual run-rate and net annual savings, and Copy Report Markdown puts a formatted comparison on your clipboard.
Step-by-Step Usage Guide
- Choose Model A and Model B from the dropdowns.
- Enter input tokens, output tokens and calls per day, or press a volume preset.
- Set the prompt-cache hit ratio if your workload uses caching.
- Read the monthly, annual and savings cards, then copy the markdown report.
Technical Specifications & Standards
Per-call cost is (input tokens / 1M) x input price + (output tokens / 1M) x output price, with the input rate blended when caching is enabled: the share of tokens you model as cache hits is charged at the board's cached-input price instead of the full input price. Day cost multiplies by calls per day, the month by 30 and the year by 365 - no weekday or seasonality shaping. The price board is a static catalogue bundled with the page and dated in its notes, listing frontier, reasoning, efficient, specialised and open-weight models across OpenAI, Anthropic, DeepSeek, Google, xAI and Mistral. That dating matters: LLM API pricing changes frequently, sometimes quarterly, so treat every figure as a snapshot and verify against the provider's current price page before budgeting. You can add your own models with custom prices via the Add Custom Model dialog on the token estimator tool, persisted in localStorage. Reasoning tokens billed as output, batch discounts and tiered pricing are not modelled.
Targeted Use Cases
- Deciding whether a cheaper alternative model is worth the quality trade-off for an agent fleet.
- Budgeting an LLM feature's annual run-rate before writing the proposal.
- Comparing a frontier baseline against a high-efficiency model at 5,000 calls a day.
- Showing finance a documented cost model with its caching assumptions stated.
Notes & Gotchas
- Measure real tokens per call from logs rather than guessing - prompts with retrieval context inflate input counts.
- Set the cache ratio to what you actually observe; 50 percent is a guess, not data.
- Re-check the price board against provider pages quarterly; pricing drifts.
- Use annual, not monthly, numbers when comparing - 30-day months hide seasonal volume.
Frequently Asked Questions
Where do the prices come from?
From a static price board bundled with the page, labelled with its catalogue date. LLM pricing changes often, so verify current rates with the provider before signing off a budget.
How does prompt caching change the cost?
The calculator charges the fraction of input tokens you set as cache hits at the model's cached-input rate and the rest at full price, then blends them into one effective input rate.
Are reasoning tokens or batch discounts included?
No. The model is simple input and output token pricing. Reasoning tokens billed as output, batch discounts and enterprise tiers are not modelled.
Can I add a model that is not listed?
Yes, via the custom-model dialog on the token and cost estimator tool. Custom prices are stored in your browser's localStorage and appear alongside the built-in board.