LLM API Cost Calculator: what your workload costs, and when a flat plan wins
This calculator turns requests, tokens, cache hits and batch share into a monthly bill for named models, then solves for the volume where a flat plan costs less. At 40,000 requests a month, GPT-5.6 costs about $771, Claude Sonnet 4.5 $578, Gemini 3 Flash $109, and DeepSeek V4 Flash $25.
Your workload
Advanced: batch discount and cache writes
Cache-write fees are excluded by default. Set writes to 10 to model one cache write per 100 requests.
Result Builder
List prices in USD per 1M tokens, retrieved 19 Nov 2026 from models.dev. Verify current rates with each provider before committing budget. This is an estimate, not an invoice.
Assumes 3,000 input and 750 output tokens per request, 40% cache hits, 30% batch share at a 50% discount. Cache-write fees excluded.
Monthly cost by model
| Model | Provider | Monthly cost | Per 1K requests | Break-even vs Builder |
|---|---|---|---|---|
| DeepSeek V4 Flash | DeepSeek | $24.60 | $0.62 | 97,551 requests |
| Llama 3.3 70B | Groq | $80.33 | $2.01 | 29,879 requests |
| Mistral Large 4.0.8 | Mistral | $99.25 | $2.48 | 24,182 requests |
| Gemini 3 Flash | $109.14 | $2.73 | 21,990 requests | |
| Claude Sonnet 4.5 | Anthropic | $578.34 | $14.46 | 4,150 requests |
| GPT-5.6 | OpenAI | $771.12 | $19.28 | 3,112 requests |
Rows are sorted by monthly cost. The cheapest row is not automatically the right row: a model that fails your evaluation is not cheap at any price.
What this calculator actually answers
Most LLM cost calculators multiply tokens by a price and stop. That answers “what would this cost per token?” It does not answer the question that decides a budget: “at what volume does a fixed monthly plan become cheaper than metered billing?” This page answers both.
You describe a workload once — requests per month, input and output tokens per request, the share of input that hits a prompt cache, and the share of traffic you can move to a batch queue. The calculator returns a monthly cost for each listed model, a cost per request and per 1,000 requests, and the break-even request count against the selected Plugsky plan. Every result is a URL you can share or paste into a spreadsheet.
The estimate is deliberately conservative. It uses list prices, not negotiated rates. It assumes cache hits and batch discounts apply in the proportions you enter, not the proportions a vendor advertises. It treats the model price as the only variable cost, because that is the part you can control.
The four inputs that change the answer
Requests per month. This is the denominator for cost per request and the unit for the break-even line. If you know daily volume, multiply by 30.4. If you know annual volume, divide by 12.
Input tokens per request. Count the whole prompt: system message, tool definitions, retrieved documents, conversation history and the user turn. A support bot with a 4,000-token RAG context and an 800-token history is a 5,000-token request, not an 800-token request. Under-counting input is the most common source of a 3–5× underestimate.
Output tokens per request. Use the max_tokens you actually set, or the p95 completion length, whichever is larger. Output is often priced 2–5× higher than input, and up to 8× on premium models, so this is the input with the largest leverage per token.
Cache and batch share. These are percentages of your traffic, not discounts. A 40% cache hit rate means 40% of input tokens are billed at the cached-input rate and 60% at the standard input rate. A 30% batch share means 30% of requests move to an asynchronous queue and receive the batch discount — typically 50% — while the other 70% stay synchronous.
How the calculation works
For each model, the calculator blends the input price by your cache hit rate, adds output cost, adds optional cache-write cost, then applies the batch multiplier:
P_blend = (1 - h) × P_in + h × P_cache
C_in = (R × I / 1,000,000) × P_blend
C_out = (R × O / 1,000,000) × P_out
C_write = (R × w / 1,000 × I / 1,000,000) × P_cache_write
C_total = (C_in + C_out + C_write) × (1 - b × d)
R* = plan_price / (C_total / R)
R is requests per month, I and O are input and output tokens per request, h is the cache hit rate, b is the batch share, d is the batch discount, w is cache writes per 1,000 requests, and R* is the break-even request count against the selected plan.
Cache and batch discounts are not additive. A 50% cache discount and a 50% batch discount do not produce a 100% discount; they apply to different slices of the same request. With a 40% cache rate and a 30% batch share at 50%, the batch multiplier is 0.85, not 0.70.
The flat-plan break-even
Once you have a cost per request, the break-even request count against a plan is plan price ÷ cost per request. At 40,000 requests a month, GPT-5.6 costs $771.12, or $0.019278 per request. Against Builder at $60, the break-even is 3,112 requests a month. Above that volume, Builder is cheaper than per-token billing on list prices alone — before counting seats, the absence of a per-token invoice and the removal of month-to-month variance.
| Plan | Price/mo | Requests/day | Tokens/day | RPM | Seats |
|---|---|---|---|---|---|
| Free | $0 | 500 | 250K | 30 | 0 |
| Hobby | $8 | 1,000 | 2M | 120 | 0 |
| Starter | $20 | 2,000 | 5M | 240 | 1 |
| Builder | $60 | Unlimited fair use | Unlimited fair use | 600 | 5 |
| Scale | $120 | Unlimited | Unlimited | 1,200 | 20 |
Annual billing is 20% off. “Unlimited fair use” is subject to fair-use RPM, concurrency and max input/output limits; see pricing.
Break-even against Builder is not the whole decision, because Starter and Hobby have published daily caps. A workload can be below Builder’s break-even and still fit Starter for a third of the price. Check the caps before recommending an upgrade.
What the estimate excludes
- Cache writes. Some providers charge a premium to write a cache entry. Use the advanced input to add them.
- Long-context tiers. Several models raise the per-token price above a context threshold. If your p95 prompt crosses a tier boundary, model the tier, not the headline rate.
- Retries and failed requests. Providers generally bill tokens processed, and retries process tokens again.
- Reasoning tokens. Reasoning models bill internal tokens as output. A 500-token visible answer can be a 3,000-token bill.
- Storage, egress, vector databases, observability and gateway fees. These are real costs and belong in a total-cost model, but they are not model spend.
- Negotiated and committed-use discounts. Enterprise contracts are not list prices.
Methodology and data notes
Prices are list prices in USD per 1 million tokens, retrieved from the models.dev catalogue on 19 November 2026. They are cross-checked against the provider names shown in the table. Model prices, context windows and max output change; verify current rates with each provider before committing budget.
The calculator applies cache hits to input tokens and batch discounts to the whole request. It does not add the two discounts together. The default table excludes cache-write fees; set “cache writes per 1,000 requests” to include them. The estimate is for model spend only, not an invoice.
Related: context window calculator · rate-limit calculator · token calculator · AI API pricing index · all models · plans.
FAQ
Does this calculator include cached and batch discounts?
Yes. It applies the cache hit rate to input tokens and the batch share to the whole request, then multiplies by the batch discount. The two are applied in sequence, not added together, because they cover different parts of the same request.
Why is my invoice higher than the estimate?
The usual causes are retries, reasoning tokens billed as output, cache writes, long-context price tiers, and prompts larger than the sample you measured. The estimate covers list-price model spend only and excludes storage, egress, gateway fees and negotiated rates.
At what volume is a flat plan cheaper?
Divide the plan price by your cost per request. At 3,000 input and 750 output tokens per request, 40% cache hits and 30% batch share, Builder at $60 breaks even at about 3,112 requests a month on GPT-5.6 and about 21,990 on Gemini 3 Flash. Use the calculator for your own numbers.
Can I use this for a model that is not listed?
The six listed models cover the common price bands. For an unlisted model, use the same formula with the provider’s input, output and cached-input prices; the calculation is identical.
Want a fixed bill instead of a per-token invoice?
Plugsky plans start at $8/month. Builder is $60/month with unlimited fair use and a 7-day money-back guarantee. Compare the break-even above, then check the caps.
See plans and caps → Start free