Q3 spending in won and euro
The figures come from KED Global's card-spending report on domestic card payments for AI services, covering the third quarter of 2026 and published in October 2026, after the quarter closed. It counts money actually charged to cards, not survey answers.
Cumulative generative AI spending in South Korea reached 1.3258 trillion won across the first three quarters of 2026, about 1.8 times the 740.8 billion won recorded for the whole of 2025. Q3 alone was 579.6 billion won, 26.5% more than Q2.
The report's own conversion puts 1.3258 trillion won at "over $900 million", which implies roughly 1,470 won to the dollar. At that rate Q3 works out to about $394 million, or near €336 million if EUR/USD sits around 1.17.
| Service | Q3 2026 share | Q3 spend (bn KRW) | ≈ USD | ≈ EUR |
|---|---|---|---|---|
| Claude | 51.2% | ~296.8 | ~$202M | ~€172M |
| ChatGPT | 36.1% | 209.4 | ~$142M | ~€121M |
| Google Gemini | 9.7% | ~56.2 | ~$38M | ~€33M |
| Wrtn | 2.2% | ~12.8 | ~$9M | ~€7M |
Only ChatGPT's quarterly total of 209.4 billion won is stated directly in the source. The Claude, Gemini and Wrtn figures are derived from the published revenue shares, so the last digits are approximate.
ChatGPT's September jump
Quarterly shares hide the movement. Reconstructing monthly card spending from the Q3 total and the reported August and September figures gives three different curves. ChatGPT: about 60.4 billion won in July, 64.7 in August, 84.3 in September. Claude: about 100.6, then 98.2, then 98.0. Gemini: about 19.5, 18.5, 18.2.
ChatGPT added 19.6 billion won of monthly spending in September alone. Claude was flat to slightly down for three months running. Gemini lost a little each month. The Korean market grew in Q3 because one participant grew.
KED Global ties the September surge to GPT-6 Astra, released on September 3 with claimed improvements in computer use, software engineering and agent workflows. Corporate card spending on ChatGPT rose 43.1% over the quarter, individual card spending 44.7%, so the rebound was not confined to one buyer type.
The ranking did not flip. Claude's 98.0 billion won in September is still 16.3% more than ChatGPT's 84.3 billion won. On the revenue measure, Anthropic's lead in Korea survived the launch.
Count, size and repeat purchase
Revenue is one axis. Payment count is another, and there ChatGPT leads. It processed 3.56 million card payments in Q3, up 33.4% on Q2 and 43.2% of every transaction in the category. The average payment rose from about 54,500 won to 58,800 won, close to $40 and roughly double a single $20 monthly subscription.
The retention number is the one worth sitting with. Of everyone who paid for ChatGPT in August, 90.0% paid again in September. For Claude the same measure was 81.2%, for Gemini 75.8%. A single month's cohort is not annual churn, and a launch month flatters anybody's figures. Even so, a nine-point gap in repeat purchase does more damage to a subscription business than one quarter of revenue share suggests.
What the models cost per million tokens
Now the other half of the story, the part a European team budgets. List prices per million tokens, converted at the same rough EUR/USD rate used above:
| Model | Input (USD) | Output (USD) | Input (≈EUR) | Output (≈EUR) |
|---|---|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 | €8.55 | €42.70 |
| Claude Fable 5.1 | $10.00 | $50.00 | €8.55 | €42.70 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | €1.71 | €8.55 |
| Gemini 4 Argon | $2.00 | $10.00 | €1.71 | €8.55 |
| Grok 4.7 | $2.00 | $6.00 | €1.71 | €5.13 |
| Mistral Large 4 | $0.50 | $1.50 | €0.43 | €1.28 |
| DeepSeek-V4.1-Flash (off-peak) | $0.15 | $0.60 | €0.13 | €0.51 |
On sticker price the two flagships are twins. GPT-6 Astra and Claude Fable 5.1 both list at $10 input and $50 output per million tokens — five times Claude Sonnet 5.5 on output and 33 times Mistral Large 4. Astra's cached input drops to $1 per million tokens, which softens the bill for repeated prompts against the same document. Anthropic's cheap tier goes further: Haiku 5.5 sits at $0.10 input and $0.50 output.
Token price is not the invoice, though. Artificial Analysis puts GPT-6 Astra at $3.26 per completed task and Claude Fable 5.1 at $7.63. Identical list prices, and the cheaper model per million tokens does not exist here — the gap of more than 2x runs entirely the other way, driven by how many tokens each task consumes and how often the agent has to restart.
That is the number I would put in front of a procurement team: same sticker, $3.26 against $7.63 for the same finished job. Our AI Arena benchmark rig tracks tokens per second, time to first token and VRAM on a single RTX 5060 Ti, which is useful for local models and irrelevant for a hosted flagship. Cost per finished task is the metric that survives the comparison.
European prices and EU rules
All four services in the Korean data are sold in the EU. API pricing is published in dollars and converted at billing. Mistral Large 4 comes from a French provider, which matters to teams that need a processor inside the EU for GDPR reasons. That is a contractual and architectural argument rather than a performance one, and on raw price Mistral Large 4 undercuts every flagship in the table.
Regulation adds a line to the budget. Article 50 of the AI Act (Regulation (EU) 2024/1689) became legally binding across the EU on 2 August 2026: synthetic audio, video, image and text have to carry machine-readable marking, and users must be told when they are dealing with an AI system. Teams that ship a support bot or a content pipeline on top of these APIs own the disclosure at the point where a person meets the output. Annex III high-risk obligations were pushed to 2 December 2027 in the Digital Omnibus package the European Commission proposed in November 2025, so conformity work for those systems is not due this year.
Limits of the comparison
Card spending is narrower than total AI revenue. It captures subscriptions and card-settled corporate seats, and says nothing about contracts billed by invoice or usage folded into larger cloud commitments. South Korea also has unusually heavy corporate card use, which flatters the corporate line in this dataset. A European market measured the same way would probably look different.
The Q3 ranking itself is unambiguous: Claude 51.2%, ChatGPT 36.1%, Gemini 9.7%, Wrtn 2.2%, with the remainder split between smaller services. September's monthly totals were 98.0 billion won for Claude and 84.3 billion won for ChatGPT.
Does ChatGPT now lead South Korea's paid AI market?
No, not on money. Claude holds 51.2% of Q3 revenue. ChatGPT leads on transaction count with 3.56 million payments and a 43.2% share, and on repeat purchase at 90.0% against 81.2% for Claude. The two measures disagree, and which one matters depends on whether you sell seats or usage.
Why is the average ChatGPT payment close to $40?
The report gives the average payment and its rise from about 54,500 to 58,800 won, but does not break the figure down by plan or billing term. Annual prepayment and higher subscription tiers would both raise an average above a single monthly fee, and the data does not say which contributes more.
Is cost per completed task a fair way to compare providers?
It depends on the task mix in the benchmark. Agentic tasks with tool calls and file handling consume far more tokens than a chat reply, so a suite weighted toward one or the other will shift the ranking. The per-task figure is still the closest thing to a real budget number, because it folds in retries and wasted output that a per-token price hides.
