2026 AI API Pricing in TWD: Claude, GPT-6, Gemini, DeepSeek

This table lists the multi-provider API unit prices (USD per million tokens) from BazaarLink's public model catalog. Each model's direct official price and its batch, promotion, and peak/off-peak conditions may differ. The TWD figures are budget estimates, using the Bank of Taiwan USD spot selling rate of NT$31.775 as of 2026-09-23. Catalog data and official price verification date: 2026-09-24. You can verify against the public model catalog and each provider's official price list.

Cross-provider price comparison (USD / NT$, per million tokens)

Three things are visible directly from the table: the top tier (GPT-6 Astra at $10 / $50) and the cheapest paid model (DeepSeek V4 Flash at $0.20 / $0.40) differ by 50x and 125x; DeepSeek V4 Pro's $2.40 input is workhorse-level, but its $4.80 output is less than half of Sonnet's, so for generation-heavy uses the results differ greatly; Gemini 3.8 Flash's $0.75 / $3.75 is the price officially marked as valid through 2026-12-31, so keep that in mind when budgeting.

Addendum (2026-10-08): The "cheapest paid model" above is a comparison based on the output unit price of DeepSeek V4 Flash in the table ($0.40). Models with lower input prices are Gemini 2.5 Flash-Lite in the table ($0.10) and Claude Haiku 5.5, listed on 2026-10-07 and not yet in the table (input $0.10 / output $0.50; its output price is still higher than DeepSeek V4 Flash). Neither changes the 50x and 125x figures, which use DeepSeek V4 Flash as the denominator.

OpenAI API Price: GPT-6 pricing, API keys and first call

OpenAI API pricing is calculated by model and token usage. The official price list as of 2026-09-24 shows GPT-6 Sol standard input/output at US$2 / US$10, cache reads at US$0.20, and cache writes at US$2.50 per million tokens; long-context requests have separate price conditions. OpenAI Platform API keys are created from a project and billed separately from ChatGPT subscriptions.

Applying for an OpenAI API key

1. Log in to OpenAI Platform and select your organization and project.
2. Open Settings → Project → API Keys, click Create new secret key, name it, and set permissions according to its purpose.
3. Put it in the OPENAI_API_KEY environment variable. Do not write it into web frontend code or the repository.

Official project key management and API quickstart documentation, verified 2026-09-24: Projects and API keys, API quickstart.

First call to the OpenAI official and compatible APIs

Official Responses API:

export OPENAI_API_KEY="your OpenAI project key"
curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-6-sol","input":"Explain API token pricing in one sentence in Traditional Chinese."}'

Using the OpenAI-compatible Chat Completions format:

export BAZAARLINK_API_KEY="your BazaarLink key"
curl https://api.bazaarlink.ai/v1/chat/completions \
  -H "Authorization: Bearer $BAZAARLINK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-6-sol","messages":[{"role":"user","content":"Explain API token pricing in one sentence in Traditional Chinese."}]}'

OpenAI model price table (2026-09-24)

The following are standard prices per million text tokens. The NT$ columns are converted using the Bank of Taiwan USD spot selling rate of 31.775 as of 2026-09-23, for budget estimates only.

This is the Standard price for short and typical context lengths. Long-context requests for models such as GPT-6 Sol, other service tiers, and tool calls are billed separately under official pricing. For BazaarLink's current prices, you can also check the OpenAI GPT-6 Sol model page, public model IDs and unit prices, and cache and discount fields. Sources: OpenAI official API pricing, GPT-6 Sol detailed pricing, Bank of Taiwan exchange rate, all verified on 2026-09-24.

The cross-provider comparison table that follows retains the original 2026-09-15 snapshot date. The OpenAI section above has been separately updated to 2026-09-24. For other providers' prices, check the date marked on the original table and each provider's official source. Further reading: OpenAI API Taiwan purchasing and billing, Claude API, Gemini API fees and TWD estimate, Gemini API sign-up.

Query date: 2026-09-15. Unit prices in the table are the customer unit prices publicly listed on the BazaarLink Models page, which match each provider's official list prices. Cache prices are taken from the same catalog. The exchange rate is 1 USD = NT$ 31.5, consistent with other price articles on this site. Providers update their official pages often, so check again before acting.

Searches for "AI API price", "cost", "discounts" and "promotions" turn up the same set of articles, because in Taiwan these four things are one question: how much each million tokens actually costs, how much my use case will cost per month, and what methods can make it cheaper. This article puts four mainstream models in one table, runs estimates on the same set of scenarios, and then splits "discounts" into four mechanisms that really exist, including the two that BazaarLink does not offer.

Summary first:

• Flagship models are priced the same: GPT-5.6 Sol and Claude Sonnet 5 are both $2 / $10. Price alone cannot decide; choose by quality and latency.
• The output unit price drives the bill: most models' output costs 5x their input, so for generation-heavy workloads look at the output columns on the right first.
• There are four kinds of "discount", and two are only available when buying direct: cache reads (BazaarLink has them; for most models they are 10% of the input price), promotional prices (yes, mirroring official), Batch API 50% off (we do not have it), and DeepSeek off-peak half price (only available when buying direct). One more exists only with us: usage rebates.
• The cheapest and the most expensive differ by 50–125x: model choice affects cost more than any discount.

Why the bill is higher than you calculated: look at output first, then caching

Long document, short answer (summarization, classification, extraction) is input-dominated. Short question, long answer (article generation, code generation, reasoning models that output their thinking process) is output-dominated. First determine which type your workload is, then check the table. Otherwise you may choose a model that is "cheap on input but expensive on output".

Four discount mechanisms: which are only available when buying direct, which we have, and which only we have

If your workload is "not urgent, high volume, and can wait 24 hours", the Batch API is the largest discount, and it requires buying direct. We are not skipping past this.

Our cache prices (as a fraction of the input price)

For the GPT and Claude series, cache writes cost 1.25x the input price. Caching only pays off when the prefix repeats more than once.

Three scenarios, same assumptions, four providers side by side (monthly)

Assumes all runs are 30 days and caching is not counted. The caching effect is demonstrated separately in the next section.

Customer service bot: 500 conversations/day, 3 turns per conversation, 300 in / 150 out per turn

Monthly tokens: input 13.5 M, output 6.75 M.

For the same task, the cheapest and the most expensive differ by 20x.

Document batch: 100 documents/day, each 6,500 tokens read and 800 tokens summarized

Monthly tokens: input 19.5 M, output 2.4 M.

For the same task, the cheapest and the most expensive differ by 11x.

Long document analysis: 50 documents/day, 50,000 tokens each, 2,000 output

Monthly tokens: input 75 M, output 3 M.

For the same task, the cheapest and the most expensive differ by 28x.

How much caching saves: using the customer service bot as an example

In each turn of the customer service bot's input, the system prompt and the first few turns of history repeat. Assume 60% of input tokens hit the cache and use Claude Sonnet 5:

• No caching: 13.5 × $2.00 + 6.75 × $10.00 = $94.5/month
• With caching: 13.5 × (40% × $2.00 + 60% × $0.20) + 6.75 × $10.00 = $79.9/month
• Savings of 15%. The output half is unchanged, so the overall percentage saved is always less than the caching discount itself.

For caching details (which models support it, how prefixes are counted, minimum lengths), see a separate article: What is prompt caching and how much it saves.

How to choose: by use case, not by brand

Taiwan dollars, invoices, rebates: three things beyond the bill

• TWD pricing and unified invoices: Buying direct gives you a USD overseas receipt. BazaarLink charges in TWD top-ups, and eligible payment channels can issue unified invoices (personal or business tax ID). For the expense reporting process, see this article.
• Usage rebate: when monthly spending on the same provider (OpenAI GPT / Gemini / DeepSeek) crosses $20 / $50 / $100 / $200 / $500, you receive $2 / $3 / $5 / $10 / $30 respectively (DeepSeek: $4 / $6 / $10 / $20 / $60). Above $500, 10% is returned again (DeepSeek 20%). Nothing below $20; it becomes stable only from $500 upward, full rules.
• Key spending cap: each API key can have a spending limit. Once reached, requests are blocked; it is not just an alert.

Standalone TWD pricing articles by provider

GPT-5 series, Claude, Gemini (including free credits), DeepSeek. For a summary of free credits, see this article.

TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API