Every week someone asks a version of the same question: "How much will this actually cost me per month?" And every week the answers are wildly wrong — usually because they're based on last year's prices, or on input tokens only, or on the assumption that output tokens cost the same as input.
Here's a field guide to getting the number right, current as of August 2026.
The one formula that matters
Monthly cost = requests × (input_tokens ÷ 1M × input_price + output_tokens ÷ 1M × output_price)
The trap: output tokens cost 4–8× more than input tokens. GPT-5 is $1.25 per 1M input but $10.00 per 1M output. Claude Sonnet 5 is $2.00 in / $10.00 out. If your workload generates long outputs — agents, code generation, summarization — the output rate dominates your bill, and comparing input prices alone will pick the wrong model.
Quick example: 100K requests/month, 2K tokens in, 500 tokens out.
- GPT-5: 100K × ($0.0025 + $0.005) = $750/mo
- GPT-5 mini: 100K × ($0.0005 + $0.001) = $150/mo
- GPT-5 nano: 100K × ($0.0001 + $0.0002) = $30/mo
Same workload, 25× spread. Model routing — sending easy tasks to nano-tier models — is the single biggest cost lever, bigger than any prompt optimization.
The caching multiplier almost everyone forgets
Both OpenAI and Anthropic offer prompt caching: repeated input tokens (system prompts, documents, conversation history) at roughly 10% of the normal input rate.
If 70% of your input is repeated context — typical for RAG and agents — your effective input price drops from $1.25 to $0.46 per 1M on GPT-5. Across a coding agent running 500K requests/month, that's the difference between $10K and $4K. I built a calculator that models this blend because I kept doing the arithmetic by hand and getting it wrong.
How many tokens is your text, really?
The rough rule: 1,000 tokens ≈ 750 English words (~4 characters per token). But rough rules cost money at scale. For anything serious, count with the actual tokenizer — OpenAI's o200k runs fine in the browser via WASM, which means exact counts with zero data leaving your machine.
Two things the word-ratio hides:
- Code is denser than prose — comments and whitespace eat tokens, so expect 30–50% fewer useful tokens than the word count suggests.
- Language matters enormously — Spanish/French run 1.5–2 tokens per word, and CJK languages can exceed one token per character. The 0.75 words-per-token rule is English-only.
Context windows are free; filling them isn't
The window itself costs nothing — you pay per token actually sent. But a request that fills the 1M-token window on Gemini 3.5 Flash costs $1.50 in input alone, every time, before the model generates a single output token. Long context is a budget decision, not a capability decision. (I keep the full cost-to-fill table here.)
And bigger isn't automatically better: "lost in the middle" degradation means models often miss details buried deep in huge prompts. A well-chunked RAG pipeline over a 200K model frequently beats dumping 1M raw tokens.
Why every price you see should carry a date
LLM pricing changes a few times a year — usually downward when new tiers launch. A pricing table without a verification date is a guess wearing a suit.
When I couldn't find a comparison that showed when each row was last checked against the official provider page, I started maintaining one: a sortable table of 22 models where every row carries its verification date, re-checked monthly against OpenAI, Anthropic, and Google's official pricing pages. Rows awaiting verification are labeled as such — I'd rather show a smaller honest table than a big stale one. The same data feeds head-to-head pages — GPT-5 vs Claude Opus 5, o3 vs GPT-5, budget-tier matchups — each computing monthly cost at three real workloads instead of listing specs side by side.
The decision tree (tl;dr)
- High volume, simple tasks → nano-tier models, route aggressively
- Long outputs (agents, code) → compare output prices first
- Repeated context (RAG, multi-turn) → check cached input rates before base rates
- Huge documents → check what filling the window costs, not just that it fits
- Before committing → run your real token counts through the formula, not a vibe
The tools mentioned above are all free, browser-only, and require no sign-up. But the math works with any calculator — the point is doing the arithmetic with current numbers and the right ratio for your workload.
What's your monthly API bill looking like in 2026, and did anything here surprise you?
Top comments (0)