One-line guide: the price lists for GPT, Claude, Gemini, DeepSeek, and Kimi are public, but “should someone like me pay a monthly plan or a token invoice?” rarely gets finished on one page. Below, the chat plan and the developer API stay on separate bills, then four kinds of usage put all five on the same month.
Figures were checked against price pages still public on 26 October 2026. This note was written on 27 October. API rates come from OpenAI, Anthropic, the Gemini Developer API, DeepSeek, and Kimi’s USD price note. Plans are cross-checked with the Google AI subscriptions page, Claude pricing, and Kimi’s mainland China price note. Region, tax, and promotions follow the checkout page. For how tokens split into input and output, start with the token price comparison.
Two bills, not one
Each of the five names sits on two rails. A membership in the web app or the coding assistant is a fixed monthly fee with a cap. An API key is priced per million tokens, and none of these vendors sell an unlimited monthly API.
- A plan buys a capped product. ChatGPT Plus, Claude Pro, Google AI Pro, and a Kimi membership cover chat and a coding assistant that can run out.
- An API buys calls from your own software. The model id goes on the request. That invoice is not the membership.
A membership does not zero the API
Google’s pricing page says Gemini 4 Argon is absent from the free tier and from the Google AI Pro API allowance. Kimi says the same in plainer words: a membership covers the product benefits in the plan, not API usage. When Claude Code or Codex usage inside the plan runs out, further API calls are a second invoice.
Which tier is actually being compared
None of these families is a single model. The rows below are the October 2026 tiers a buyer actually chooses between. Older models are still on sale. They are not the default route this month.
| Family | Tiers in this comparison | Where they sit |
|---|---|---|
| GPT | Astra, GPT-6.1 Sol, Luna | Astra is the flagship. Sol is described as near-Astra at about a fifth of the price. Luna is the cheap tier. A prompt over 272K tokens reprices the whole request. |
| Claude | Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5 | Fable matches the flagship price. Opus is the long-running default. Sonnet is the middle of speed and intelligence. Haiku takes classification, extraction, and routing. |
| Gemini | 4 Argon, 3.8 Flash | Argon is the frontier model in the app and on the paid API. 3.8 Flash has a free tier; the paid tier is still introductory. The Argon split is in Argon’s three bills. |
| DeepSeek | V4.1 Flash, V4 Pro | The API name deepseek-flash is served by V4.1-Flash-1017. V4 Pro has no vision. Both have peak and off-peak rates. |
| Kimi | K3.1 Pro, Highspeed | Standard K3.1 Pro matches the Sonnet headline rate. Highspeed is about 2–3× faster and doubles input and output, but not cache hits. Mainland China has a separate yuan list. |
Every worked example compares the same million tokens, not the same paragraph. Claude 4.7 and later tokenize the same text at roughly 30% more tokens than earlier Claude models. Two rows with the same sticker price can still land on different invoices.
Chat and coding: the monthly plan
If the request never leaves the vendor’s own product, read this table before the per-million rates. Kimi’s mainland plans are yuan, billed annually and shown as an effective month. The international page is a different dollar list. Do not convert one into the other and call them the same plan.
| Door | Published monthly price | What it buys | Offsets the API? |
|---|---|---|---|
| ChatGPT Go / Plus / Pro | $8 / $20 / from $100 | Plus fits a few coding sessions a week. Pro also lists $200 and $500. The public note says Pro currently has no five-hour limit. | No |
| Claude Pro / Max | $20 monthly; about $17 on annual billing. Max from $100 | Pro covers chat and capped Claude Code. Max is about 5× or 20× Pro usage. | No |
| Google AI Pro / Ultra | $19.99 / $99.99 or $199.99 | Pro includes Argon in the Gemini app and Search. Ultra is about 5× or 20×, and the Argon trial in Antigravity is written onto Ultra. | No |
| Kimi, mainland China, annual effective month | Go ¥39, Plus ¥79, Pro ¥159, Max ¥559, Ultra ¥1,119 | Kimi Code starts at Plus. Highspeed starts at Pro. Extra-long chat up to 1M tokens starts at Max. | No |
| Kimi, USD, annual effective month | Plus $15, Pro $31, Max $79, Ultra $159 | The international page has no Go tier. The table is an effective month; the charge is the annual total. | No |
| DeepSeek | The public list is the API | A free chat allowance, if the product has one, is not the prepaid API balance. | No |
This table answers “what do I pay every month to stay inside the official app?” It does not answer “how many times will my site call the model?”
API list prices per million tokens
Rates below are standard, short context, cache miss. Cache hits are their own column. Output prices include thinking tokens: a short answer on screen can still be billed as a long internal trace.
| Model | Input | Output | Cache hit | What changes it |
|---|---|---|---|---|
| GPT-6 Astra / Claude Fable 5.1 | $10 | $50 | Astra $1.00, Fable $0.25 | GPT prompts over 272K tokens: whole-request input ×2, output ×1.5 |
| Claude Opus 5.5 / Kimi Highspeed | $4 | $20 | Opus $0.20, Highspeed $0.20 | Highspeed does not raise the cache-hit rate |
| GPT-6.1 Sol / Sonnet 5.5 / Kimi K3.1 Pro | $2 | $10 | $0.10 / $0.10 / $0.20 | Same sticker, different tokenizer and cache rules |
| Gemini 4 Argon | $2, then $4 from 1 Dec | $10, then $20 | $0.10, then $0.20 | Intro rate through 30 Nov 2026. Not on the free tier |
| Gemini 3.8 Flash | $0.75, then $1.50 from 1 Jan 2027 | $3.75, then $7.50 | $0.075, then $0.15 | Free tier exists. Cache storage is billed per hour on top |
| DeepSeek V4 Pro | Peak $1.32 / off-peak $0.66 | Peak $3.96 / off-peak $1.98 | Peak $0.044 / off-peak $0.022 | Off-peak is half of peak |
| DeepSeek V4.1 Flash | Peak $0.30 / off-peak $0.15 | Peak $1.20 / off-peak $0.60 | Peak $0.006 / off-peak $0.003 | Same peak window |
| GPT-6 Luna / Haiku 5.5 | $0.10 | $0.50 | $0.01 | Haiku prompts over 100K tokens move to $0.50 / $2.50 |
| Kimi K3.1 Pro, mainland API | ¥13, Highspeed ¥26 | ¥68, Highspeed ¥136 | ¥1.30 | Not the same list as the dollar API |
DeepSeek peak hours are Monday–Friday 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. Every other hour, full weekends, and full holidays are off-peak. Batch on OpenAI, Anthropic, and Gemini is about half the standard rate, with different latency. It does not switch itself on.
Router prices move daily. The OpenRouter weekly usage piece is a routing-layer price, not a substitute for the list above.
Four users, one month
All four rows use a 5:1 input-to-output mix, the standard tier, short context, and no cache hits. The token counts exist so they can sit next to a monthly plan. They are not packages. Mainland Kimi is in yuan. Everything else is in dollars.
| User | Assumption this month | API bill |
|---|---|---|
| Chat only | Stays in the official app | API is $0. Stop at an $8–$20 plan, or Kimi Plus at ¥79 on the mainland list. If the weekly cap holds, do not create a key. |
| Codes a few times a week | 5M input + 1M output | Luna / Haiku $1.00. DeepSeek Flash peak $2.70, off-peak $1.35. Gemini 3.8 Flash $7.50. Sol, Sonnet, Kimi, and intro Argon are all $20. Opus or Highspeed $40. Astra / Fable $100. Mainland Kimi ¥133. |
| An agent every day | 40M input + 8M output | Luna / Haiku $8. DeepSeek Flash peak $21.60, off-peak $10.80. Gemini Flash $60. DeepSeek Pro peak $84.48. The $2/$10 row is $160. The Opus row is $320. Astra / Fable $800. Mainland Kimi ¥1,064. |
| A product already shipped | 200M input + 40M output | Luna / Haiku $40. DeepSeek Flash peak $108, off-peak $54. Gemini Flash $300. DeepSeek Pro peak $422.40. The $2/$10 row is $800. The Opus row is $1,600. Astra / Fable $4,000. Mainland Kimi ¥5,320. |
On the light coding row, a $20 API bill and a $20 Plus or Pro plan look identical. The plan has a weekly cap and a five-hour refresh. The API has neither, and it also does not include the machine that compiles. If the included allowance finishes the work, do not open a second meter.
Two dates rewrite today’s invoice
From 1 December 2026, Argon input and output both double. The Argon share of the $20 / $160 / $800 examples becomes $40 / $320 / $1,600. Gemini 3.8 Flash doubles on 1 January 2027, so $7.50 / $60 / $300 become $15 / $120 / $600. DeepSeek has no calendar like that. It already cuts the rate in half every off-peak hour.
Cache, long context, and peak hours
Cache discounts input only
Take the daily-agent row and hit cache on 70% of input. Sol or Sonnet is 12M × $2 + 28M × $0.10 + 8M × $10 = $106.80, not $160. The expensive line is still the 8M of output. The same shape on DeepSeek Flash at peak is 12M × $0.30 + 28M × $0.006 + 8M × $1.20 = $13.37, not $21.60. Kimi Highspeed cache hits stay at $0.20, so the speed premium lands on fresh input and on output.
GPT’s 272K line reprices the whole request
On Astra, 270K input plus 5,000 output is about $2.70 + $0.25 = $2.95. Add input until the prompt is 300K and the entire request moves to long-context rates, about $6.00 + $0.375 = $6.38. The extra cost is not “the 30K that crossed the line.” If an agent keeps appending tool output, compacting before 272K saves more than swapping vendors.
Haiku changes rows after 100K tokens
Haiku 5.5 is two rows: $0.10 / $0.50 up to 100K tokens, and $0.50 / $2.50 once the prompt is over that line. Do not price only the tail at the higher rate. Luna has no 100K step. Its step is 272K, shared with the rest of GPT-6.
DeepSeek’s half depends on the clock
The shipped-product Flash row is $54 if every call is off-peak and $108 if every call is peak. On a Beijing weekday, 01:00–04:00 and 06:00–10:00 UTC cover a morning block and an afternoon block. Weekend jobs sit on off-peak all day. Separately, Search and Fetch calls made through DeepSeek Harness, Claude Code, or Codex start billing at $0.75 per 1,000 successful requests at 00:00 Beijing time on 1 November 2026. Until then they are not charged.
Which bill fits which user
- Chat in the browser. Stay on the cheapest plan that actually includes the model. Mainland Kimi Plus is ¥79 on annual billing. Argon in the Gemini app starts at Google AI Pro, $19.99. Do not prepay an API to chat.
- Personal coding, a few times a week. Finish the ChatGPT Plus, Claude Pro, or Kimi Code allowance first. If that runs out, draft on Luna, Haiku, or DeepSeek Flash, and raise review to Sol or Sonnet before a merge.
- An agent that runs every day. Do not leave the main loop on Astra or Fable. The gap between $800 and $21.60 is the output price, not a longer prompt. Move a single step up to Opus or a flagship only when a failure is expensive.
- A product in production. Default the route to Flash, Luna, or Haiku. Keep the $2/$10 row as the quality backstop. Make Argon’s unit price a config value before December, and do the same for Gemini 3.8 Flash before January.
- Billing from mainland China. Kimi invoices in yuan. DeepSeek’s own API stays in dollars, then splits peak and off-peak. Dividing ¥13 by an exchange rate and comparing it with $2 mixes two tax and currency regimes.
Easy misreads
- “They are all $2/$10, so swap freely.” Sol, Sonnet 5.5, Kimi K3.1 Pro, and Argon through 30 November only share a sticker. Cache hits, long-context surcharges, and tokenizers do not.
- “The answer was short, so it was cheap.” Thinking tokens sit inside the output price.
- “Budget DeepSeek at the peak rate and you are safe.” Weekends and holidays are half price all day. The reverse mistake, using only the off-peak rate for weekday daytime, cuts the estimate in half.
- “Today’s OpenRouter price is the list price.” The router moves daily. The list above is the vendor page.
- “The dollar membership and the yuan membership are one catalog.” Mainland Kimi has Go and different multipliers. The international Plus tier starts at $15. The API is likewise ¥13/¥68 or $2/$10, not both at once.
Questions
Where should an API default point in October 2026?
If the work still fits inside a plan allowance, do not point it at the API. When tokens are the meter, keep the main loop on DeepSeek Flash, GPT-6 Luna, or Haiku 5.5, and keep quality on GPT-6.1 Sol or Sonnet 5.5. Leave Astra and Fable 5.1 for the step whose failure is expensive.
Why do several models cost $2 and $10 per million tokens?
GPT-6.1 Sol, Claude Sonnet 5.5, Kimi K3.1 Pro, and Gemini 4 Argon through 30 November 2026 land on that standard row. Opus 5.5, Kimi Highspeed, and Argon from 1 December land on $4 and $20. Astra and Fable 5.1 are $10 and $50.
For Kimi in China, which currency is the bill?
Whichever currency checkout uses. The mainland membership page is an annual yuan price shown as a month. The standard API is ¥13 input and ¥68 output per million tokens. The international membership and the dollar API are a different column. Neither membership includes API usage.
Why does one DeepSeek model have two prices?
Peak is two UTC blocks on weekdays. Off-peak is half, and it covers the rest of the clock, all weekend, and Chinese public holidays. The model name stays deepseek-flash or V4 Pro. It is not two products.
Written on 27 October 2026. Unit prices follow the vendor pages linked above. Plan tiers follow the membership pages public at the same time. The token counts in the examples are there to sit next to a monthly fee. They are not plans.
ZavCloud Cloud Mac
Keep the model invoice off the Mac that compiles
The monthly fee or the token invoice above pays for the model. Xcode, a long-running agent, and a repository index still need a Mac that stays on.
A dedicated cloud Mac is rented as a machine. It does not share an invoice with OpenAI, Anthropic, Google, DeepSeek, or Kimi. The build and the toolchain can stay in the cloud. The laptop only has to read the result.