Short answer: your bill is (tokens × rate) times a multiplier stack most calculators ignore. For a realistic 3,000-message month, estimates run from $2 on GPT-6 Luna to $78 on Claude Opus 5.5, and the naive estimate misses by 30–100% once reasoning tokens, cache writes, retries, and long-context tiers get in. Five steps below get you within a usable band before you sign up for anything.

The pricing page shows $2 per million input tokens. Your invoice shows $47. Both numbers are real, and the gap is what this post is about. You can close it before spending a cent, and it takes about twenty minutes.

What an LLM API cost calculator misses

Search "llm api cost calculator" and you get a full page of tools: llm-prices, Helicone, benchlm, gptforwork, openreplay, Artificial Analysis, and several more. They're fine tools. All of them compute the same thing: your token counts multiplied by a rate card.

The problem is upstream of the multiplication. A calculator can't know that your agent retries failed tool calls 30% of the time, that your reasoning model bills thinking tokens at output rates, or that Google's promo price doubles on January 1, 2027 (it does). Those are assumptions, and assumptions are where estimates die.

So the method here is: build the assumptions first, then multiply. If you run a single model at modest volume and don't want to think about any of this, your provider's own dashboard is genuinely enough. This is for people choosing a model or committing to a monthly budget.

Step 1: Build your token profile

Rate cards are per 1M tokens. Tokens are roughly 4 characters of English text, so a 2,000-word document is about 3,300 tokens. For chat work, measure where you can: most providers count tokens in the response usage object, and Anthropic publishes a token counting endpoint you can call before you send.

Three profiles cover most cases:

Profile	Input / message	Output / message	Examples
Light chat	1,500	300	Quick questions, translation
Working session	4,000	800	Multi-turn drafts, code review
Agent run	20,000+	3,000+	Coding agents, research loops

Use the working-session profile for the rest of this post: 4,000 input and 800 output tokens per message.

Step 2: Split input into uncached, cache reads, and cache writes

Not all input tokens cost the same. Providers give steep discounts for re-reading a cached prefix, and some charge extra to write one. Anthropic's docs lay out the multipliers explicitly: a 5-minute cache write costs 1.25x base input, an hourly write 2x, and a read costs 0.1x (0.05x on Opus 5.5, 0.025x on Fable 5.1). OpenAI discounts GPT-6 cached reads 90%.

A realistic working session repeats its system prompt and history, so assume 40% of input arrives as cache reads, 5% as cache writes, and 55% as fresh input. For our 4,000-token message that's 2,200 uncached, 1,600 cached reads, 200 cache writes. Whether caching pays for itself depends on how often a prefix gets reused; the write premium is the part most estimates forget.

Step 3: Audit the multiplier stack

This is the step every quick calculation skips. Line items that silently resize a bill:

Reasoning tokens. Billed at output rates, on top of the visible answer. A "short" answer from a thinking model can carry thousands of invisible output tokens. Details in our reasoning tokens cost guide.
Long-context tiers. GPT-6 Sol runs $2/$10 until you cross into long-context territory, where it's $4/$15 (double input, half again on output). Anthropic's Claude 4.6-and-later models bill the full 1M window at standard rates with no surcharge. Same job, different shape of bill.
Clock-based pricing. DeepSeek's deepseek-flash is literally twice the price in peak hours (01:00–04:00 and 06:00–10:00 UTC weekdays). Our example profile runs $2.53 a month off-peak and $5.07 at peak.
Data residency. Anthropic's inference_geo: "us" adds 1.1x to every token category. OpenAI's regional endpoints carry a 10% uplift.
Router fees. OpenRouter took 5.5% on credit purchases (minimum $0.80) and 5% on BYOK usage above the $25K monthly allowance, as verified in early September. Check the current schedule before you model it.
Per-call tools. Web search is $10 per 1,000 calls at OpenAI and Anthropic; Gemini's grounding is 5,000 free a month then $14 per 1,000.

Step 4: Scale to volume and add an agent multiplier

Messages per day × 30 gives base volume. Working session, 100 messages a day: 3,000 a month, 12M input tokens and 2.4M output.

Agents multiply that. A coding agent that makes 12 model calls per task costs like 12 messages, plus retries. OpenAI's own AutomationBench number for GPT-6 Sol at xhigh effort is $0.27 per completed task, and task-level dollars are the number to watch as agents grow. If independent tests of a model show it burning more tokens per task (as several did for Grok 4.7 this week), the per-token rate is the wrong lever.

One lever cuts the other way: batch. Async work through any batch API is about half price. Our 100K-row classification job at 1,500 input and 100 output tokens per row costs $400 at standard GPT-6 Sol rates and $200 batched.

Step 5: Put a band on it, then check

Don't produce one number. Produce a band: expected usage at P50, and a P90 built on 1.5x volume with a worse cache hit rate. Then spend one week at low volume, read the actual usage numbers and invoice line items, and re-fit. The first estimate is a hypothesis. The guide to monitoring and capping API spending covers keeping the P90 from becoming a surprise.

Worked example: 3,000 messages a month, seven models

Profile: working session (4,000 input, 800 output per message), 40% cache reads, 5% cache writes, 55% uncached. All rates read from provider pricing pages on September 22, 2026.

Model	Input	Output	Cache read	Est. monthly
GPT-6 Luna	$0.10	$0.50	$0.01	$1.98
deepseek-flash (off-peak)	$0.15	$0.60	$0.003	$2.53
Gemini 3.8 Flash*	$0.75	$3.75	$0.075	$14.76
Grok 4.7†	$2.00	$6.00	n/a	$38.40
GPT-6 Sol	$2.00	$10.00	$0.20	$39.66
Claude Sonnet 5	$2.00	$10.00	$0.20	$39.66
Claude Opus 5.5	$4.00	$20.00	$0.20	$78.36

Prices per 1M tokens. * Google's rate through December 31, 2026; it doubles January 1, 2027. Excludes $0.50 per million tokens per hour of cache storage. † xAI publishes no cache discount in its model table, so every input token is priced at base rate. Conservative by up to 20%.

Per message that's $0.0007 on Luna and $0.0261 on Opus 5.5. Per thousand messages, $0.66 to $26.12. If your numbers come out dramatically different from this shape, check your cache split first; it moves the answer more than any rate card does.

Which estimate should you trust?

If you…	Do this
Run one model, low volume	Use your provider's dashboard and skip this
Compare models for a switch	Run these five steps on your real profile
Build agents	Add a 2–3x task multiplier and measure task dollars
Have a backlog that can wait	Model it at batch rates, roughly half
Set a budget	Commit to the P90 number and cap it

Where this method doesn't help

It estimates, and estimation has a floor. Early product usage wobbles more than any spreadsheet predicts, which is why the band matters. It also says nothing about output quality, and quality differences between a $2 model and a $4 model are not visible in rate cards at all. If a cheap model writes copy you rewrite from scratch, its effective cost is whatever the rewrite cost you.

And there's a real case against managing any of this: if computing these tables once a quarter sounds miserable, a flat $20 chat subscription that bundles usage is a rational buy. The math in our BYOK versus subscriptions comparison shows where that trade flips back.

FAQ

Can't I just plug numbers into an LLM API cost calculator?

You can, and the number won't be wrong so much as incomplete. The calculator multiplies the assumptions you feed it, so the five steps above are really about making those assumptions match your workload first.

What do 100 messages a month actually cost?

At the working-session profile in this post, $1.32 on GPT-6 Sol and $2.61 on Claude Opus 5.5. Volume that low rarely justifies optimization work.

Why is my first month higher than my estimate?

Development usage wobbles: failed runs, retries, larger prompts while you tune. Budget the P90 band for month one and re-fit once the usage numbers settle.

Does a workspace subscription change the provider math?

No. With BYOK, provider billing stays separate from the app fee, so the monthly total is the flat workspace price plus the token math from this post.

Multimodel Chat keeps the two sides of the bill separate: Pro is $2/month, and provider usage stays on your own key at the rates above, with no token markup. See what the workspace fee covers →

Start your free trial → — 7 days, all providers, no credit card required.

How to Estimate LLM API Costs in 2026 (Before You Commit)