11 min read
AI pricing has two layers: the monthly subscription a person pays to use a chat app, and the API token rate a company pays to build AI into its own product. What changed this week is the second layer. GPT-6 and Claude Opus 5.5 shipped on September 21–22, token prices fell by as much as half, and not a single subscription price moved.
This piece puts both layers on one page. First a table of ChatGPT, Claude and Gemini subscription tiers; then the API price lists, model by model. Every figure was checked on the companies’ official pages on September 26, 2026. Where a figure could not be verified, the cell is left blank and says so. Won conversions use ₩1,375 to the dollar.[11]
Monthly subscription tiers (individual plans)
| Tier | ChatGPT | Claude | Gemini (Google AI) |
|---|---|---|---|
| Free | Free — $0/month | Free — $0/month | Free — ₩0/month (3.6 Flash, limited 3.1 Pro) |
| Entry | Go — under $5/month in India and some other markets; Korean price unverified | None | AI Plus — ₩7,500/month (≈$5.5), 2× free-tier usage, 400 GB storage |
| Standard | Plus — $20/month (≈₩27,500) | Pro — $20/month (≈₩27,500), or $17/month billed annually (≈₩23,400) | AI Pro — price unverified (4× free-tier usage, 5 TB) |
| Top | Pro — $200/month (≈₩275,000); new signups paused since September 10 | Max — from $100/month (≈₩137,500), 5× or 20× Pro usage | AI Ultra — ₩119,000/month (≈$87) at 5× Pro, ₩300,000/month (≈$218) at 20× Pro, 20 TB |
Dollar plans are billed in dollars; the won figure in parentheses is a conversion. Gemini’s Korean page lists won prices directly. Sources: ChatGPT via TechCrunch reporting, Claude via claude.com/pricing, Gemini via the gemini.google Korea subscriptions page.[7][8][12]

How to read it.
- The standard tier is pinned to the same number. ChatGPT Plus and Claude Pro both cost $20 a month. Google AI Pro is presumably in the same band, but the Korean page renders that one price client-side and the figure could not be confirmed, so it is not in the table.
- The cheapest paid plan is Google AI Plus at ₩7,500. ChatGPT Go sells for under $5 in India and Indonesia; whether it is offered in Korea, and at what price, could not be verified because OpenAI’s pages block automated access.[12]
- Top tiers are sold as multiples, not features. Claude Max is 5× or 20× Pro; Google AI Ultra is 5× or 20× Pro. $100 and ₩119,000 sit in the same slot. ChatGPT Pro is $200, and on September 10 OpenAI paused new signups, citing infrastructure strain from GPT-6 Astra demand. Whether signups have resumed is unconfirmed.[12]
- Only Claude states an annual discount plainly: $17 a month, 15% off. Google is running a 40% annual-plan discount on AI Pro for eligible users through October 31.
- VAT. None of the three Korean pages says whether tax is included. Dollar billing also adds card exchange and foreign-transaction fees.
Nothing in this table changed with this week’s cuts. Anthropic’s one consumer-side change was a limit, not a price: it raised the 5-hour usage caps on Pro, Max and Team.[3] More use for the same money, not less money.
Which plan to pick
Sorting by use case gets to an answer quickly.
- A few questions a day: free is enough. All three companies open a limited slice of their latest generation to free users.
- Daily document and code work: the $20 tier. ChatGPT Plus or Claude Pro is a question of tools, not price. Claude Pro bundles Claude Code (the terminal coding tool) and Chrome and Microsoft 365 integrations; ChatGPT Plus bundles Codex, deep research and image generation.[7]
- Mostly living in Google Workspace (Gmail, Docs, Drive): Google AI Plus at ₩7,500 is the best value. Gemini shows up in Gmail and Vids, and 400 GB of storage comes with it. For summarizing and tidying, the $20 tier is unnecessary.[8]
- Running AI all day (developers with several coding agents open, researchers feeding long documents continuously): only then does the $100-plus band make sense. That band buys usage headroom, not a better model.
- Teams: Claude Team is $20 per seat per month billed annually. ChatGPT Business and Enterprise prices render dynamically and could not be confirmed.
API token prices — the layer that moved
A token is the smallest unit an AI model uses to process text, a little smaller than a word. Companies charge consumers a flat monthly fee, but they charge developers and businesses that build AI into apps by the token, counting input (what you send) and output (what comes back) separately, with output far more expensive. So this price list is effectively the cost sheet for any product with AI inside.
On September 21 and 22, frontier models arrived in a cluster: xAI’s Grok 4.7, Xiaomi’s MiMo v2.6, Anthropic’s Claude Opus 5.5, and about an hour later OpenAI’s GPT-6 Sol and GPT-6 Luna.[4]
The figures below are from each company’s official price page, in dollars per million tokens.[1][2][5][6]
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
| GPT-5.6 Luna (previous) | $0.20 | $0.02 | $1.20 |
| GPT-6 Sol | $2 | $0.20 | $10 |
| GPT-5.6 Sol (previous) | $4 | $0.40 | $20 |
| GPT-6 Astra | $10 | $1 | $50 |
| Claude Opus 5.5 | $4 | $0.20 | $20 |
| Claude Opus 5 (previous) | $5 | $0.50 | $25 |
| Claude Fable 5.1 | $10 | $0.25 | $50 |
| Claude Haiku 4.5 | $1 | $0.10 | $5 |
| Gemini 3.1 Pro (preview) | $2 | $0.20 | $12 |
| Gemini 3.8 Flash | $0.75 | $0.075 | $3.75 |
| Grok 4.7 | $2 | $0.50 | $6 |
OpenAI and Google figures are the short-context tier. OpenAI charges double for long prompts; Google charges $4 input and $18 output above 200k tokens.[1][6] Gemini 3.8 Flash is a promotional price through December 31, 2026, and doubles to $1.50 input and $7.50 output on January 1, 2027.[6]

What the table shows.
- Prices have converged within each class. GPT-6 Sol ($2/$10) and Gemini 3.1 Pro ($2/$12) are nearly identical; Claude Opus 5.5 ($4/$20) is double. Anthropic says Opus 5.5 “performs at the level of Claude Fable 5.1 on most work,” and against Fable ($10/$50) it is the cheaper option.[3]
- GPT-6 Luna is the cheapest small model. At $0.10 input and $0.50 output it is one-seventh the price of Gemini 3.8 Flash ($0.75/$3.75).
- “Cached input” needs a word. When the same long instructions or documents precede every request, the second call onward gets that portion at a steep discount. This is where Anthropic cut deepest: from $0.50 to $0.20 per million tokens, a 60% reduction.[2] Developer Simon Willison notes that in long-running automated work, more than 90% of input tokens are billed at the cached rate.[4]
Anthropic’s stated reason: it takes less compute
Anthropic’s announcement says Opus 5.5 “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.” It does not cite competition. OpenAI’s prices appear nowhere in the post. Instead: “Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that.”[3]
Compute is the processing capacity needed to run a model. Less GPU time for the same answer means the company can sell it cheaper.
40% and 20% are different numbers
One filter is needed here. The 40% Anthropic cited and the number on the price list are not the same thing.
The listed token rates fell 20%, input and output alike: $5 to $4, $25 to $20.[2] The 40% adds a second claim on top: that the model uses fewer tokens to do the same job. The first 20% anyone can verify on the public price page; the rest is the company’s own test result.
Some summaries at home and abroad merged the two into “40% off base API rates,” which the price list does not say. TokenPost kept them apart: 20% in the headline, 40% in the body.[11]
Four things a price list cannot measure
The token rate is one variable in the bill. The rest is not on the list.
Different tokenizers count differently. A tokenizer is the rule that splits text into tokens. Anthropic’s documentation states that models from Claude 4.7 onward use a new tokenizer that produces “approximately 30% more tokens for the same text.”[2] Two models with the same listed rate can bill different token counts for the same input.
Thinking time is billed as output. Current models reason at length before answering. That reasoning, visible or not, is charged as output tokens. Google labels the line item “Output price (including thinking tokens).”[6] Turn the reasoning effort up and the per-task cost can multiply while the unit rate stays put.
The old price was already a discount. OpenAI’s price page carries this footnote: “GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026.”[1] The baseline that got halved was never the list price. Willison points out that with a 25% November increase scheduled for GPT-5.6, GPT-6 is effectively half of a promotional rate.[4] Gemini 3.8 Flash’s doubling in 2027 is the same pattern.
Regional processing adds a surcharge. Anthropic charges 1.1× on all tokens when data processing is pinned to the US. OpenAI adds 10% on regional and data-residency endpoints.[1][2] For anyone regulated into fixed-region processing, the list price is not the bill.
Told to think longer, it never answered
Willison tests every new model by asking for an SVG of a pelican riding a bicycle. With Opus 5.5 at maximum reasoning effort, for the first time in the history of the test, nothing came back.
The model settled the pelican’s shin length at “about 95.2,” debated whether the foot should sit “between y=478 and 494” on the pedal, considered a fish in the basket, and exhausted its 128,000-token output limit before drawing a single line. A retry did the same. The two failures cost $2.56 each and nearly 20 minutes. Willison calls max mode “effectively useless,” reasoning that a model that collapses on a trivial drawing cannot be trusted with anything that matters.[4]
Cheaper tokens do not mean cheaper tasks. How you call the model makes a bigger difference than the rate cut.
Where Korean APIs stand
Set next to domestic models: Upstage’s Solar Pro 4 lists $0.30 input and $1.20 output; Solar Mini 4 lists $0.10 and $0.40, and is half price ($0.05/$0.20) through October 22. A Korean company, but priced in dollars, with 10% VAT extra.[9]
On the cheapest row, GPT-6 Luna ($0.10/$0.50) and Solar Mini 4 at list ($0.10/$0.40) are nearly the same. That says nothing about capability; it does establish that at the low end a Korean API is not losing on price. Naver’s HyperCLOVA X price page renders only in JavaScript and could not be read here; no public token rates were found for LG’s EXAONE or SKT’s A.X.
Lower prices also do not translate directly into adoption. A Ministry of Trade survey cited by NewDaily found 37.1% of 685 Korean companies had adopted AI: 65.1% of large firms, 35.6% of small ones, 31.2% of mid-sized ones. The obstacles named were skills shortages and upfront investment, not token rates. Statistics Korea’s 2023 business survey put AI use at 13.5% of large companies and 4.5% of small ones.[10]
What to do with this
- If you pay a subscription, your bill is unchanged. Your usage cap went up. Treat any headline saying “AI subscriptions got cheaper” with suspicion. If you are switching plans, do it for the tools in the table, not the price.
- If you use apps with AI inside, the builder’s cost went down. That does not mean your price will. Whether the saving becomes a lower price or more features is up to each product.
- If you are evaluating for a company, compare cost per task, not the rate card. Tokenizers, reasoning tokens, cache hit rates and residency surcharges are not on the list. Run the same hundred work samples through each candidate and compare the invoices. It is the only reliable method.
- Do not default to maximum reasoning effort. Start at the default and raise it only when the output falls short. Max is not guaranteed to be better, and a failure still costs money.
The tables here will be updated in place whenever a price page changes. Check the date at the top.
Reflection
Competition that was heading toward higher performance at higher prices has shifted one notch toward the same performance at a lower price. Anthropic justifying a cut with “less compute” rather than a benchmark is the tell. In the same post it played down its own benchmark table, writing that “benchmark margins have become a less reliable guide to real-world differences,” and said Sonnet 5.5 and Haiku 5.5 arrive “in the coming weeks.”[3]
How long does a gap last where the cost halved and the consumer price did not move?
Notes
References
- Anthropic: Claude Opus 5.5
- Simon Willison: Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war
- OpenAI API pricing
- Anthropic pricing documentation
- Google Gemini API pricing
- TechCrunch: OpenAI puts Pro subscriptions on hold
Notes
- OpenAI API pricing page (checked 2026-09-26). Footnote: “GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026.” developers.openai.com ↩
- Anthropic, Claude pricing documentation (checked 2026-09-26). Tokenizer note: “approximately 30% more tokens for the same text”. platform.claude.com ↩
- Anthropic, “Claude Opus 5.5” (2026-09-22). Quotes: “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5”, “Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that”, “benchmark margins have become a less reliable guide to real-world differences”. anthropic.com ↩
- Simon Willison, “Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war” (2026-09-22). simonwillison.net ↩
- xAI model documentation, Grok 4.7 pricing (checked 2026-09-26). docs.x.ai ↩
- Google Gemini API pricing (checked 2026-09-26). 3.8 Flash: “$0.75 through December 31, 2026. $1.50 starting January 1, 2027.” 3.1 Pro Preview: “$2.00, prompts <= 200k tokens”, “$12.00, prompts <= 200k tokens”. Output line item: “Output price (including thinking tokens)”. ai.google.dev ↩
- Anthropic, Claude plans page (checked 2026-09-26). Pro $17/month billed annually ($200 upfront) or $20 monthly; Max from $100/month at 5× or 20× Pro usage; Team standard seat $20/month annual or $25 monthly. claude.com/pricing ↩
- Google Gemini subscriptions, Korea page (checked 2026-09-26). “Google AI Plus … 매월 ₩7,500 KRW”, “Google AI Ultra … ₩119,000 KRW/월: AI Pro 대비 5배”, “₩300,000 KRW/월: AI Pro 대비 20배”. The AI Pro card shows no price line. gemini.google ↩
- Upstage Solar API pricing (checked 2026-09-26). upstage.ai/pricing ↩
- NewDaily, “‘이렇게 비용이 많이 들어가?’ … AI는 아직 그림의 떡, 디지털 전환도 버겁다” (2026-08-19), citing the Ministry of Trade, Industry and Energy survey and Statistics Korea’s business activity survey. biz.newdaily.co.kr ↩
- TokenPost, “앤트로픽, 클로드 오푸스 5.5 출시…가격 20% 낮춰” (2026-09-23). Source of the ₩1,375 conversion rate. tokenpost.kr ↩
- TechCrunch, Sarah Perez, “OpenAI puts Pro subscriptions on hold due to Astra demand” (2026-09-10): “temporarily pause subscriptions for its $200-per-month Pro plan, citing strain on its infrastructure”, “the company’s other plans, including the API and lower-cost Go and Plus plans, remain available.” Plus at $20/month per TechCrunch’s pricing explainer (updated 2025-02-25); Go “sub-$5” per TechCrunch 2026-08-27 (India). OpenAI’s own pricing pages block automated access, so won-denominated prices could not be confirmed. techcrunch.com ↩
Source: https://simonwillison.net/2026/Sep/22/opus-and-sol-and-luna/