Last updated: November 2026
TL;DR
- The dollar answer: On the API, by our estimates, a 1,500-word blog post costs about 3-7 cents, 1,000 triaged emails cost about $4-$7.50 a month, and a 10-turn customer chat costs about 7-12 cents. These use two example price points ($3 input / $15 output and $5 input / $30 output per million tokens) taken from one vendor-reported price list.
- AI token costs are the per-use charges providers bill for the chunks of text a model reads (input) and writes (output). One token is roughly 0.75 words, so 1,000 tokens is about 750 English words.
- Output tokens cost more than input tokens. One independent source puts the gap at roughly 5-6x, and reasoning models can multiply the bill because their hidden “thinking” is billed like output.
- Surprises come from long context, agent loops and retries, not from the headline price. MindStudio argues that cost per completed task is a better yardstick than price per token.
Token cost: the charge a provider makes each time an AI model reads or generates tokens (small chunks of text). It is usually quoted in dollars per million tokens, with separate rates for input and output.
What is a token, and why do AI tools bill by it?
A token is a chunk of text, either a whole word or part of one. Language models process text as tokens rather than letters, and one independent explainer puts the rule of thumb at 1 token ≈ 0.75 words. The more text a model reads and writes, the more a request costs. According to JumpCloud, a tokenizer converts your text into an array of integers whose length sets the billing footprint, and each output token requires a complete forward pass through the model weights.
Different models split text differently. Tetrate notes the same sentence can take 20 tokens in one model and 25 in another, so identical text can cost different amounts.
Not all text is equally “heavy”:
- Non-English text: Chinese or Arabic can use significantly more tokens than equivalent English, per Tetrate. An independent source says the same for Hindi and Burmese.
- Code and JSON: these often use more tokens than plain prose because of syntax, indentation and special characters.
How does a request turn into a bill?
Every request is billed as input tokens times the input rate, plus output tokens times the output rate, with rates quoted per million tokens (Tetrate). The mechanism is simple, but several things feed into it.
- You send input. This includes your prompt, attached documents, system instructions and, in a chat, the earlier conversation.
- The model generates output. JumpCloud says output tokens are generated sequentially, and each new one requires a complete forward pass through the model weights.
- The provider counts both sides. The totals are multiplied by the per-million rates.
- Extras apply. Cached tokens are reused input, such as a system prompt, that the provider stores so it does not have to reprocess it on every call. How caching shows up in your price depends on the provider, so check its pricing page. Hidden reasoning tokens are billed at the output rate.
A quick conversion: dividing a per-million rate by 1,000 gives the cost of 1,000 tokens. Tetrate cites Claude Sonnet 4.6 at $3 per million input tokens when its page was written, which is $0.003 per 1,000. A 1,300-token request costs just 0.13% of the per-million rate.
Why do input and output tokens cost different amounts?
Output tokens usually cost more because they are generated one after another, each needing a complete forward pass through the model weights (JumpCloud). One independent source describes input tokens as relatively cheap and estimates output at roughly 5-6x the price of input.
The price list we reviewed shows similar ratios:
| Model class (as listed by MindStudio, undated) | Input per 1M tokens | Output per 1M tokens | Output ÷ input |
|---|---|---|---|
| Kimi K2 | about $3 | about $15 | 5x |
| GPT-5.1-class | about $5 | about $30 | 6x |
These prices come from a vendor page and may have changed since, so treat them as illustrations. The practical effect is that tasks producing long answers cost more than tasks that read a lot and answer briefly. According to MindStudio, output-heavy stages such as code writing drive most of the spend.
Reasoning models add a wrinkle. Their hidden chain-of-thought is billed at the output rate, so one source says they can cost several times more for the same question. Turning reasoning off for easy tasks is one of the standard ways to cut usage.
Tip: Before you pick a tool, find out whether it bills per token or per seat. If you can choose a model, test your real task on the cheapest option first and move up only if the quality isn’t there.
How does token pricing map to the plans you see?
A flat subscription charges a fixed monthly fee, while API pay-as-you-go bills the input and output tokens you use at per-million rates. A flat plan is a fixed fee whose included usage is set by the vendor, while the API is metered. The same task is therefore a fixed cost in one case and a variable cost in the other.
The token math in this article applies directly to the API. Our sources contain no subscription plan prices or usage caps, so check each vendor’s pricing page for what a flat plan includes.
Tetrate also notes that the same model can carry different per-token prices when called directly versus through Bedrock, Vertex AI or Azure. The access route can therefore change the rate you pay.
What do common freelancer and small-business tasks cost?
For typical writing, triage and chat tasks, the token cost is cents per task, and from about $4 to about $59 a month at the volumes in the table. The figures below are our own arithmetic, using the 0.75 words-per-token rule and the two per-million rate pairs from the table above. The token counts are assumptions, not measurements. Provider caching features could change the totals.
| Task | Assumed tokens | Cost at $3 in / $15 out | Cost at $5 in / $30 out |
|---|---|---|---|
| One 1,500-word blog post (500-word prompt plus notes ≈ 1,000 input; draft ≈ 2,000 output) | 3,000 | $0.033 | $0.065 |
| 1,000 emails triaged a month (≈600 input and ≈150 output each) | 750,000 | $4.05 | $7.50 |
| One 10-turn customer chat (≈14,500 input in total, ≈1,500 output) | 16,000 | $0.066 | $0.1175 |
| 500 such chats a month | 8,000,000 | $33.00 | $58.75 |
The chat row is the one to study. The assumptions are a 500-token system prompt, 50-token customer messages and 150-token replies. Because each turn resends the whole history, input balloons to 14,500 tokens, while the new customer and reply text is only about 2,000 tokens. Tetrate makes the same point: every token in the context window is billed, so long chat histories can cost more than the new responses.
For a scale check, WAV Group estimated 1,000 requests a month at about 20,000 tokens each (roughly 20M tokens, with 30% of input cached). It put GPT-5.2 at about $89.17 a month, GPT-5 Mini at about $12.74 and GPT-5 Nano at about $2.55. WAV Group’s prices were as of December 2025, and it is a vendor-side source. In its scenario, model tier changed the monthly total from about $2.55 to $89.17.
Why do AI bills surprise people?
Bills surprise people because the headline rate hides how many tokens a job really uses. Long context, agents that loop, retries and heavier reasoning all multiply token counts, and the per-million price doesn’t show any of that.
Context length. Every token in the window is billed, including history and system instructions. A preprint on arXiv (submitted 3 Nov 2026) reports that later tokens are costlier to generate than earlier ones, and that the disparity “may rise to as much as 7x” on frontier open-weight models. That is a compute cost for whoever runs the model. The preprint does not show that providers bill later tokens at higher rates or that your invoice changes by that factor. It also reports that frontier models sacrifice little to no accuracy when they keep contexts short through frequent compaction. This is a single preprint, so treat the finding as early evidence that shorter conversations may save money without hurting quality.
Agents and retries. In automated systems, loop traps can turn compute time directly into runaway token cost, according to JumpCloud. Tetrate adds that rate cards omit cache hits, retries and model mix, and recommends measuring blended cost per million tokens across all models and channels.
Scale. WAV Group’s 10-day voice-AI test used over 910,000 input tokens across 58 requests, which would have cost just under 30 cents at paid rates. Extrapolated to 20 sessions of 100 requests a day, it projected nearly 1 billion tokens a month and about $471 in input costs alone, at $0.50 per million input tokens and before output.
Cheaper per token isn’t always cheaper per task. MindStudio reports that Kimi K2 uses about twice as many tokens as GPT-5.1 on comparable tasks, which cancels most of its price advantage. Independent Artificial Analysis data cited on that page put the cost per completed task at about $0.95 for Kimi K2, $1.04 for a GPT-5.1-class model and $2.75 for an Anthropic Opus-class model.
Misconceptions and blurry boundaries
- “Words equal tokens.” They don’t. 0.75 words per token is an English-text average, and code, JSON and non-English text run heavier.
- “The cheapest per-token model is the cheapest overall.” Token usage per task varies by model, as the Kimi K2 example shows.
- “Capping output always saves money.” JumpCloud warns that capping output length too tightly can truncate reasoning and reduce accuracy.
- “Token cost is just the API bill.” The term also covers training cost, which is dataset size times epochs, but that is the provider’s cost, not yours.
A cost-estimate checklist you can reuse
- Write down the task and how often it runs per month.
- Estimate input tokens: words ÷ 0.75, plus system prompt and any history that gets resent.
- Estimate output tokens the same way. Add extra if you use a reasoning model.
- Look up the current input and output rates per million tokens for your chosen model.
- Calculate: (input tokens × input rate + output tokens × output rate) ÷ 1,000,000.
- Multiply by monthly volume, then add a safety margin for retries and longer-than-expected context. Size the margin from the 10-run test in the tip below.
- Set a per-task token budget. JumpCloud notes that hard budgets bound worst-case spend even when other safeguards fail to trigger in time.
Tip: Run your real task 10 times, check the token counts in your provider’s usage dashboard, and compare them with your estimate before you scale up. Use the gap between the two to size your safety margin.
Where can you check current token prices?
Check the model provider’s own pricing page first, since rates change. The prices in this article come from the cited sources at their own dates: WAV Group’s are from December 2025, and the Tetrate and MindStudio pages are undated in our inputs. Verify any figure before you budget with it. If you route through a cloud platform, check that platform’s price too, because Tetrate notes the same model can be priced differently on Bedrock, Vertex AI or Azure.
To compare across models and providers, one independent source points to OpenRouter. WAV Group recommends logging token usage by user and task type in dashboards so you can see what’s driving spend.
Limits and risks
- Security and data access: according to JumpCloud, tracking token consumption helps security teams prevent resource-exhaustion attacks and budget overruns from infinite generation loops. Review your provider’s usage dashboard regularly.
- Failure modes: loops, retries and ever-growing chat history inflate usage. Rate limiters in your own application logic can count tokens and halt execution once a budget threshold is passed.
- Costs: commitment-based pricing can lower rates, but you pay for unused capacity if usage falls short, per Tetrate. WAV Group suggests that self-hosting small or large language models, using on-demand GPU capacity, or fine-tuning can improve cost efficiency compared with plug-and-play APIs.
- Quality trade-offs: WAV Group argues that cheap Nano or Mini models suit routine work, and that premium models should be saved for high-stakes tasks. That is a vendor-side recommendation, so test it on your own work.
- When not to use metered API billing: in our view, if you can’t monitor usage or set a budget, metered billing is the riskier choice. Check whether your vendor offers a fixed-price plan and what its limits are.
FAQ
Is AI becoming too expensive to use?
For typical freelancer tasks, our estimates suggest no: a blog post costs a few cents and 1,000 triaged emails cost roughly $4-$8 at the example rates above. Costs get painful at scale or with agents and long contexts: WAV Group projected about $471 a month in input costs alone for a heavy voice-AI workload.
What are tokens in AI?
Tokens are the chunks of text a language model reads and writes, either whole words or word fragments. In English, 1 token is roughly 0.75 words, so 1,000 tokens is about 750 words. Providers bill separately for input tokens and output tokens.
What is the current price of AI?
It depends on the model and provider, and prices change. As of the sources we reviewed, one vendor page lists Kimi K2 at about $3 input and $15 output per million tokens, and a GPT-5.1-class model at about $5 and $30. Claude Sonnet 4.6 was listed at $3 per million input tokens. Check the provider’s pricing page for today’s rates.
What is the 30% rule in AI?
None of our sources define a standard “30% rule” for AI costs. The only 30% figure we found is WAV Group’s assumption that 30% of input tokens were cached in its monthly cost scenario. If you saw the phrase elsewhere, check its source for the intended meaning.
Why does the same text cost different amounts on different models?
Each model uses its own tokenizer, so the same sentence can take 20 tokens in one model and 25 in another (Tetrate). Per-token prices also differ between models and between access routes such as Bedrock, Vertex AI or Azure.
How can I reduce my token costs?
Independent sources suggest four steps: turn off reasoning for easy tasks, use the smallest model that does the job, ask for shorter answers, and cache reused content like system prompts and documents. Keep chat histories short, and set a hard token budget per task.
Sources: arXiv preprint 2611.05306, Tetrate: cost per token, WAV Group: AI token costs, MindStudio: token cost optimization, JumpCloud: what is token cost in AI.