GPT-5 API Cost in TWD 2026: GPT-5.6 Sol, Terra, Luna
USD/TWD price table for GPT-5 series: GPT-5.6 Sol, Terra, Luna and 5.4-mini per million token prices, plus monthly TWD estimates for three real scenarios.
OpenAI prices in USD, but Taiwan companies planning a budget need TWD figures. This table converts the per-million-token prices of the current main GPT-5 series models (GPT-5.6 Sol / Terra / Luna, plus the lightweight 5.4-mini tier) into NT$, and estimates monthly costs across three real scenarios.
Bottom line up front: GPT-5.6 Terra is the sweet spot for most production tasks (its input is exactly half of the flagship Sol's). For high-frequency simple tasks, use GPT-5.4-mini directly, where the cost is almost negligible. Save Sol for agent workflows that truly need top-tier reasoning.
Want to call the GPT API with a TWD budget view and a company tax ID invoice? BazaarLink provides an OpenAI-compatible endpoint, and the SDK only needs its base URL changed. For the expense-claim process, see the complete guide to claiming AI API expenses for Taiwan companies.
Model cost comparison (USD ↔ NT$)
The exchange rate is 1 USD = NT$ 31.5 (consistent with other pricing articles on the site). MTok means per million tokens. The unit prices below are the customer prices publicly listed on the BazaarLink Models page.
| Model | Input USD/MTok | Output USD/MTok | NT$ input | NT$ output | vs Terra (input / output) | Suitable scenarios |
|---|---|---|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $25.00 | NT$ 158 | NT$ 788 | 2.0× / 1.67× | Most complex agents, deep reasoning |
| GPT-5.6 Terra | $2.50 | $15.00 | NT$ 79 | NT$ 473 | 1× (baseline) | Main production tasks, customer service, coding |
| GPT-5.6 Luna | $1.00 | $6.00 | NT$ 32 | NT$ 189 | 0.4× / 0.4× | High-frequency medium tasks, translation, summarization |
| GPT-5.4-mini | $0.75 | $4.50 | NT$ 24 | NT$ 142 | 0.3× / 0.3× | Lightweight chat, internal tools, batch processing |
The table shows official list prices. Actual checkout unit prices may be lower than this table depending on your plan and current promotions. Always refer to the real-time unit prices shown on the BazaarLink Models page. The scenario estimates below use the official list prices in this table, so they are upper-bound estimates.
There is also a Pro tier (GPT-5.4-pro: $30 / $180 per MTok, about NT$ 945 / NT$ 5,670) for extreme reasoning needs. It is not recommended as a standing choice for general production tasks. For the latest unit prices, see BazaarLink Models.
Three real scenarios (NT$/month)
Scenario 1: Customer service bot (GPT-5.6 Terra, 500 conversations/day)
Assumptions: 500 conversations per day, an average of 3 turns per conversation, and 300 input tokens / 150 output tokens per turn.
- Monthly turns: 500 × 30 × 3 = 45,000 turns
- Monthly input tokens: 45,000 × 300 = 13,500,000 (13.5 MTok)
- Monthly output tokens: 45,000 × 150 = 6,750,000 (6.75 MTok)
Cost:
- Input: 13.5 × NT$ 79 = NT$ 1,067
- Output: 6.75 × NT$ 473 = NT$ 3,193
- Total: NT$ 4,260/month (≈ USD 135)
→ Compared with Claude Sonnet 5 in the same scenario (NT$ 2,977/month), this combination costs about 43% more. If cost is the main consideration, Sonnet 5 is cheaper for this scenario. Response style and latency still differ, so it is worth testing both for a week before deciding.
Scenario 2: Document batch processing (GPT-5.4-mini, 100 documents of 5,000 characters/day)
Assumptions: 100 documents per day, about 6,500 tokens each, and an 800-token output summary.
- Monthly input tokens: 3,000 × 6,500 = 19,500,000 (19.5 MTok)
- Monthly output tokens: 3,000 × 800 = 2,400,000 (2.4 MTok)
Cost:
- Input: 19.5 × NT$ 24 = NT$ 461
- Output: 2.4 × NT$ 142 = NT$ 340
- Total: NT$ 801/month (≈ USD 25)
→ At the same volume, Luna would be NT$ 1,068/month (about 1.3×), and Terra would be NT$ 2,670/month (about 3.3×). For low-complexity batch tasks, mini is the obvious choice.
Scenario 3: Agent workflow (Sol for planning + Terra for execution, 50 runs/day)
Assumptions: 50 complete agent tasks per day. Each task has 1 Sol planning call (2,000 input + 800 output tokens) and 4 Terra execution calls (1,500 input + 600 output tokens per call).
- Monthly tasks: 1,500
- Sol: 3 MTok input, 1.2 MTok output
- Terra: 9 MTok input, 3.6 MTok output
Cost:
- Sol input: 3 × NT$ 158 = NT$ 474
- Sol output: 1.2 × NT$ 788 = NT$ 946
- Terra input: 9 × NT$ 79 = NT$ 711
- Terra output: 3.6 × NT$ 473 = NT$ 1,703
- Total: NT$ 3,834/month (≈ USD 122)
→ Running the same tasks entirely on Sol would cost about NT$ 5,670/month (1.5 times as expensive). The mixed strategy of flagship for planning and main model for execution saves about NT$ 1,836 per month.
Usage rebate: how much of this money comes back
BazaarLink's usage rebate is not a percentage discount. It is a fixed milestone rebate. Each time eligible spending in the month crosses a threshold, a fixed amount is issued. Thresholds reached are cumulative (not only the highest tier is paid), and the portion above the last threshold earns a proportional rebate once more. Rebates are credited directly to your account balance.
For the GPT series, the applicable program is the OpenAI GPT rebate program:
| Monthly spending crosses | Rebate |
|---|---|
| USD 20 | + $2 |
| USD 50 | + $3 |
| USD 100 | + $5 |
| USD 200 | + $10 |
| USD 500 | + $30 |
| Portion above USD 500 | Another 10% |
The effective rebate percentage is not fixed; it fluctuates with the month's spending. It is best when you land exactly on a threshold, and it gets diluted between thresholds:
| Eligible monthly spending | Rebate | Effective percentage |
|---|---|---|
| USD 19 | $0 | No rebate |
| USD 49 | $2 | About 4.1% |
| USD 100 | $10 | 10% |
| USD 199 | $10 | About 5.0% |
| USD 500 | $50 | 10% |
| USD 1,000 | $100 | 10% |
Applying the three scenarios above:
| Scenario | Monthly cost | Rebate | Effective equivalent |
|---|---|---|---|
| Scenario 1 customer service bot (Terra) | About USD 135 | $10 | About 7.4% |
| Scenario 2 document batch (5.4-mini) | About USD 25 | $2 | About 8.0% |
| Scenario 3 agent workflow (Sol + Terra) | About USD 122 | $10 | About 8.2% |
In Scenario 3, Sol and Terra are both in the same program, so their spending counts toward the thresholds together. This is also why concentrating traffic from the same model family is more cost-effective.
⚠️ Not every OpenAI model is included in the program. The Pro variants gpt-5.6-luna-pro, gpt-5.6-terra-pro and gpt-5.6-sol-pro, as well as embedding, image and video model types, are currently not included. Matching uses the exact model ID, not prefix matching, so please refer to the list on the explainer page.
A few things to know first:
- Settlement is based on the UTC calendar month. Rebates for the previous month are credited on the 1st, and each month is calculated anew; nothing carries over to the next month.
- The rebate is a balance credit, not a payment item, so it is not refundable, not separately invoiced, and does not expire.
- Free plan usage, bring-your-own-key (BYOK) usage, and usage offset by subscription points do not count.
- Rebates go to the wallet that generated the usage: personal usage goes to the personal wallet, and organization usage goes to the organization wallet.
- Each plan is calculated separately, so spending on different plans does not count toward the thresholds together.
For the complete rules and the currently effective model list, see the usage rebate explainer.
BazaarLink vs buying directly from OpenAI
The API is fully OpenAI-compatible, so existing code only needs its base URL changed. The differences are in billing currency and invoicing process:
| Direct purchase from OpenAI | BazaarLink | |
|---|---|---|
| Billing currency | USD (foreign currency on a credit card) | Settled internally in USD; site prices are shown in USD |
| Invoice | US company receipt (PDF) | Unified invoice (personal or company tax ID) |
| Tax ID | Not supported | Supported |
| Prepaid balance | Some plans | Supported (balance does not expire) |
| Top-up service fee | None | Charged at the rate shown on the top-up page |
| Best for | Individual developers who do not need expense claims | Taiwan companies that need expense claims and TWD budget visibility |
→ For the full expense-claim process, see How do Taiwan companies claim AI API expenses? 2026 complete guide.
Closing
Want to test latency and response quality before deciding? The BazaarLink free trial (no credit card required) lets you call GPT-5 series models directly. Related pricing articles: Claude TWD cost estimate, Gemini TWD cost estimate and DeepSeek TWD cost estimate. For the latest unit prices, see BazaarLink Models.
Explore current model discounts and learn how usage rebates add credit to your balance.
View discounts and usage rebatesFAQ
How do GPT-5.6 Sol, Terra and Luna differ, and which one should I choose?
They are from the same generation but serve different roles. Sol ($5 / $25 per MTok) is the flagship for the most complex reasoning and agent tasks. Terra ($2.5 / $15) is the main production model, with input exactly half of Sol's, and is the first choice for most customer service, writing and coding tasks. Luna ($1 / $6) is the lightest and suits high-volume simple tasks. A common strategy is to start with Terra, move up to Sol when you hit a bottleneck, and move down to Luna at high volume. (The table shows official list prices; the actual checkout unit price follows the model catalog shown on the site.)
Roughly how much does GPT-5.6 Terra cost per month?
It depends on usage. Estimating 500 customer service conversations per day (3 turns each, with 300 input and 150 output tokens per turn), the monthly cost is about NT$ 4,260 (approximately USD 135). Light development and testing usually falls between a few hundred and one or two thousand NT dollars per month.
Which GPT model is cheapest for high-frequency batch tasks?
GPT-5.4-mini ($0.75 / $4.50 per MTok) is the cheapest option in the table. For the same document processing volume (100 documents per day, about 6,500 tokens each), the monthly cost is about NT$ 801, roughly 3/4 of Luna's cost and 3/10 of Terra's. For low-complexity tasks such as classification, extraction and summarization, the spending is almost negligible.
How does buying GPT API on BazaarLink differ from buying directly from OpenAI?
The API is fully OpenAI-compatible, so your code does not need to change. The differences are in billing and invoicing. BazaarLink supports a TWD budget view, unified invoices (personal or company tax ID), and prepaid balances that do not expire. A service fee is charged on top-ups, and the rate is shown in real time on the top-up page. For Taiwan companies that need expense claims, the invoicing process is usually worth more than the service fee saved.
Which exchange rate is used to convert USD to NT$?
This table uses an estimate of 1 USD = NT$ 31.5, consistent with other pricing articles on the site. BazaarLink settles internally in USD. The TWD figures are for budget estimates only, and actual charges are calculated in USD.
Is there a rebate for the GPT series, and roughly how much can I get back?
Yes. GPT models are covered by the OpenAI GPT rebate program. Each time eligible spending in the month crosses $20, $50, $100, $200 or $500, a fixed rebate of $2, $3, $5, $10 or $30 is issued respectively. Thresholds reached are cumulative, and spending above $500 earns an additional 10%. The effective percentage varies: landing exactly on a threshold gives 10%, while amounts between thresholds are diluted, and nothing is given below $20. Note that Pro variants, embedding, image and video model types are not included. The active list is determined by the explainer page.
TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API