Key facts
- There is no sticker price for one token. Lists are dollars per million tokens. Divide by 1,000,000. On 21 Oct 2026, GPT-6 Sol’s short-context output on OpenAI API pricing is $10 per million, so one output token is $0.00001 and 1,000 of them are $0.01.
- Add three lines: uncached input, cache read, and output. Gemini API pricing says the output price includes thinking tokens. Cache writes and hourly cache storage are extra lines.
- The bill below is arithmetic, not an export from our account. 2,000 uncached input, 6,000 cache hits, 1,500 output. GPT-6 Sol is $0.0202 once, and $8.89 for 440 calls a month (22 days × 20). Output is $0.015 of that call.
- Ten times those counts, still under each short-context cliff, is $88.88 a month on Sol. A chat plan does not bill this formula. Do not add the two charges.
How to turn a per-million price into one token
OpenAI API pricing, Claude API pricing, and Gemini API pricing do not price a single token. They price one million tokens. One token costs that rate divided by 1,000,000. One thousand tokens cost that rate divided by 1,000. GPT-6 Sol and Claude Sonnet 5.5 both list standard output at $10 per million, so 1,000 output tokens are $0.01 and one output token is $0.00001.
Input is usually cheaper than output. Sol input is $2 per million, so 1,000 input tokens are $0.002, one fifth of the output rate on the same model. A single “price per token” that ignores this split understates generation. GPT-6 Luna and Claude Haiku 5.5, for prompts up to 100,000 tokens, list output at $0.50 per million: 1,000 output tokens are $0.0005. GPT-6 Astra output is $50 per million: 1,000 output tokens are $0.05.
- OpenAI’s published rule of thumb is for common English: about 4 characters, or about 0.75 words, per token. It is not a contract price, and it is not a Chinese rate.
- Chinese, code, and JSON have no shared official ratio. The same paragraph can be a different token count at each vendor. Bill the counts the API returns.
- A chat plan’s monthly fee is not a per-token subscription price. Dividing that fee by a guessed token count is not the rate in this article.
Split usage into input, cache, and output
One request is three products added together. Uncached input tokens divided by 1,000,000, times the input rate. Cache-read tokens divided by 1,000,000, times the cache-read rate. Output tokens divided by 1,000,000, times the output rate. The three rates differ. GPT-6 Sol is $2 input, $0.20 cache read, $10 output. Claude Sonnet 5.5 is $2 input, $0.10 cache hit, $10 output. Gemini 3.1 Pro, for prompts up to 200,000 tokens, is $2 input, $0.20 cache read, $12 output.
Output hides tokens you do not see in the chat. Gemini API pricing says the output price includes thinking tokens. OpenAI and Claude do not publish a separate thinking rate; extra generated tokens use the output rate. The field names are whatever that response’s usage object says. The counts below are chosen so the arithmetic can be checked. They are not a log line from a live account.
- 2,000 uncached input tokens: the new question, or the code pasted this turn.
- 6,000 cache-hit tokens: a repeated system prompt or tool list that is already cached. This section prices a hit, not a write.
- 1,500 output tokens. If that response bills thinking as output, those tokens are already inside the 1,500.
- Short context: under OpenAI’s 272,000, Haiku’s 100,000, and Gemini 3.1 Pro’s 200,000. Standard tier, not Batch, Flex, or Fast.
- No search grounding and no regional-processing uplift. Those are different rows on the price cards.
What one token costs at ChatGPT, Claude, and Gemini
The table is the standard rate still on the official pages on 21 Oct 2026, in dollars per million tokens. “Per 1,000 output” is the output rate divided by 1,000. OpenAI rows are the OpenAI API pricing short-context rates (input at or under 272,000). Claude rows are the Claude API pricing global standard rates. Gemini rows are the Gemini API pricing paid standard tier. The Haiku 5.5 row applies only when the prompt is at most 100,000 tokens. Gemini 3.8 Flash’s numbers run through 31 Dec 2026, Gemini 4 Argon’s through 30 Nov 2026, and Gemini 3.1 Pro’s row is the tier at or under 200,000 prompt tokens.
| Model | Input | Cache read | Output | Per 1,000 output | What this row covers |
|---|---|---|---|---|---|
| GPT-6 Luna | $0.10 | $0.01 | $0.50 | $0.0005 | Short context, input ≤272K |
| GPT-6 Sol | $2.00 | $0.20 | $10.00 | $0.01 | Short context, input ≤272K |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | $0.01 | Short context; cheaper cache read than Sol |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | $0.05 | Short context, input ≤272K |
| Claude Haiku 5.5 | $0.10 | $0.01 | $0.50 | $0.0005 | Prompt ≤100K; 5× above that |
| Claude Sonnet 5.5 | $2.00 | $0.10 | $10.00 | $0.01 | Global standard |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 | $0.02 | Global standard |
| Gemini 3.8 Flash | $0.75 | $0.075 | $3.75 | $0.00375 | Paid standard through 31 Dec 2026 |
| Gemini 3.1 Pro | $2.00 | $0.20 | $12.00 | $0.012 | Prompt ≤200K |
| Gemini 4 Argon | $2.00 | $0.10 | $10.00 | $0.01 | Through 30 Nov 2026; implicit cache only |
Past the cliff, the rate jumps. GPT-6 Sol above 272,000 input tokens moves from $2 / $10 to $4 / $15. Haiku 5.5 above 100,000 prompt tokens moves from $0.10 / $0.50 to $0.50 / $2.50. Gemini 3.1 Pro above 200,000 prompt tokens is $4 input and $18 output. From 1 Dec 2026, Gemini 4 Argon is $4 input, $20 output, and $0.20 cache read. From 1 Jan 2027, Gemini 3.8 Flash is $1.50 input and $7.50 output. Batch on these tiers is about half the standard rate, and what you trade is latency. OpenAI adds 10% on regional-processing endpoints for models released on or after 5 Mar 2026. Claude’s US-only inference is higher too: Sonnet 5.5 is $2.20 / $11. The table does not include either uplift.
Same usage, what the bill actually is
Put 2,000 / 6,000 / 1,500 into the formula above. The cache line is a hit, not a write. The month is 22 working days times 20 calls, 440 calls, same shape every time. The 10× column multiplies all three counts by 10. The prompt is then 80,000 tokens, still under the 100,000, 200,000, and 272,000 cliffs, so Haiku does not flip to the 5× rate. This is price math, not a download of our account.
| Model | One call | 440 calls / month | 10× usage / month |
|---|---|---|---|
| GPT-6 Luna / Haiku 5.5 (≤100K) | $0.00101 | $0.44 | $4.44 |
| Gemini 3.8 Flash (through 31 Dec 2026) | $0.007575 | $3.33 | $33.33 |
| GPT-6 Sol | $0.0202 | $8.89 | $88.88 |
| GPT-6.1 Sol / Sonnet 5.5 | $0.0196 | $8.62 | $86.24 |
| Gemini 4 Argon (through 30 Nov 2026) | $0.0196 | $8.62 | $86.24 |
| Gemini 3.1 Pro (≤200K) | $0.0232 | $10.21 | $102.08 |
| Claude Opus 5.5 | $0.0392 | $17.25 | $172.48 |
| GPT-6 Astra | $0.101 | $44.44 | $444.40 |
GPT-6 Sol’s $0.0202 splits into $0.004 for 2,000 new input tokens, $0.0012 for 6,000 cache reads, and $0.015 for 1,500 output tokens. Output is about one fifth of the input-side count and still the largest of the three amounts. Cut cost on output first: a smaller model, or less rewriting of whole files, shows up before another round of cache tuning. On this workload, Luna or Haiku is $0.00101 a call and $0.44 a month; 3.8 Flash is $3.33 a month; Sol is $8.89; Sonnet 5.5, GPT-6.1 Sol, and intro-priced Argon are $8.62; 3.1 Pro is $10.21; Opus 5.5 is $17.25; Astra is $44.44. The gap between tiers is an order of magnitude, not a few cents.
Why the first cache write can cost more than no cache
The table above prices a hit. The first time those 6,000 tokens are written, they use the write rate, and they are not also charged as ordinary input. GPT-6 Sol’s cache write is $2.50 per million. Ordinary input is $2 per million. That write is $0.034. Sending all 8,000 tokens as fresh input, with no cache, is $0.031. The first write costs $0.003 more than no cache. Each later hit is $0.0202, which is $0.0108 less than no cache. When the write rate is above the input rate, the cache pays for itself only on later hits.
Claude Sonnet 5.5’s 5-minute cache write is also $2.50 per million, so that first write is $0.034 as well. The 1-hour write is $4 per million, and that call is $0.043. A hit is $0.10 per million, $0.0196 for this workload. Gemini 3.1 Pro’s cache read is $0.20 per million, $0.0232 on a hit. Explicit cache also has storage: $4.50 per million tokens per hour. Holding 6,000 tokens for one hour is $0.027, more than the request itself. A short call with a cache that stays up for a long time can spend more on storage than on tokens. Through 31 Dec 2026, Gemini 3.8 Flash storage is $0.50 per million tokens per hour, $0.003 for the same 6,000 tokens and one hour. Gemini 4 Argon’s price card lists implicit caching only, not that hourly storage line.
Where a subscription and per-token billing diverge
ChatGPT plans, Claude plans, and Gemini subscriptions sell a usage ceiling inside the chat product. On those pages in October 2026, ChatGPT Plus and Claude Pro are $20 a month (Claude Pro is about $17 a month if you pay annually), and Google AI Pro is $19.99. Higher tiers start near $100. The pages change; use the price you see. Those fees are not a price per token, and $20 divided by a guessed token count is not one either.
Set the monthly arithmetic next to the plan fee. Do not treat them as the same product. Sol at $8.89 is under a personal plan near $20. That does not mean the plan is a bad deal: the plan’s model and these 440 API calls are not the same item. A plan slows you down when the allowance is gone. The API keeps charging the list rate. At 10× usage, Sol is $88.88, and a personal allowance usually runs out first. Only the overflow, if you send it to the API, is paid near that column. Do not add the plan fee and the API bill unless you are paying both.
- Chat in the product, and the allowance is enough: read the plan page. Do not back out a per-token price from the subscription.
- Your own program, an agent loop, or usage you can reconcile: use the API and the formula above. The output rate is the large term.
- An aggregator adds a fee on the credit purchase. That is a second bill; see How Much Does the OpenRouter API Cost? GPT, Claude, and Gemini per Million Tokens and the Monthly Bill. Gemini’s free tier and its billing tracks are in Gemini API Pricing 2026: Token Cost and Free Tier Explained.
FAQ
How much is one token?
There is no single price. On the 21 Oct 2026 standard short-context card, one GPT-6 Sol or Claude Sonnet 5.5 output token is $0.00001 (1,000 of them are $0.01). One GPT-6 Luna or Haiku 5.5 output token, for prompts up to 100,000, is $0.0000005. Input is cheaper, and a cache read is a third rate.
How many tokens is one Chinese character?
There is no official fixed ratio. OpenAI’s rule of thumb covers common English: about 4 characters per token. For Chinese or code, use the token counts the API returns. Do not multiply a character count by a guessed ratio and then by the rate.
Why do the three bills differ for the same paragraph?
The tokenizers differ, so the counts differ, and the rates differ too. This article forces one shared count so the rates can be compared. Your own bill needs each vendor’s usage counts, then that vendor’s rates.
Does caching always cost less?
No. Writing 6,000 tokens once on GPT-6 Sol is $0.034. No cache at all is $0.031. The write rate, $2.50 per million, is above the $2 input rate, so the first write costs more. The saving is on later hits. Storing 6,000 tokens for one hour on Gemini 3.1 Pro also adds $0.027 of storage.
What is one token inside a $20 ChatGPT Plus plan?
You cannot divide it out. Plus, Claude Pro, and Google AI Pro sell an allowance, not the per-million card in this article. $8.89 and $20 can sit side by side. Do not divide them, and do not add them, unless you are paying for both a plan and the API.
The token bill still needs a machine that stays awake
The dollars above land on the API. The process that sends the request still holds the repo, the keys, and a session that should not die when a laptop lid closes. A cloud Mac mini is a macOS machine that stays on, so the Unix toolchain and these API clients can share one Apple Silicon host. A day rental is how you check $8.89 against a real usage log, without buying a workstation first. The machine has to stay up. A closed lid breaks the log you would reconcile.