In this guide · 8 sections
Short answer
On October 2026 list prices, Gemini 3.8 Flash costs about US$2.25 per 1,000 typical FAQ replies (US$4.50 from January 2027). Claude Haiku 5.5 and GPT-6 Luna cost about US$0.30 for the same work, and Gemini 3.1 Flash-Lite about US$0.80. Test quality on your own questions before choosing.
Gemini Flash releases in 2026: dates that matter
All dates below are from Google's Gemini API changelog, checked on 21 October 2026. "GA" means generally available, the stable version Google supports for production use.
| Date | Change | What it means for a chatbot |
|---|---|---|
| 21 Jul 2026 | Gemini 3.6 Flash and Gemini 3.5 Flash-Lite GA | 3.5 Flash-Lite is Google's low-cost option for high-volume tasks |
| 13 Aug 2026 | Gemini 3.7 Flash GA at an introductory price | Short-lived; deprecated on 8 Oct 2026 |
| 2 Sep 2026 | Gemini 3.8 Flash GA | Google's most capable Flash model |
| 18 Sep 2026 | Gemini 2.5 access limited to past users | Not deprecated, but new projects should use 3.5 Flash-Lite or 3.8 Flash |
| 8 Oct 2026 | 3.7 Flash and 3.5 Flash deprecated | Requests are routed automatically to 3.8 Flash and 3.6 Flash |
The automatic routing matters if a developer built your bot on an older model name: your bot may already be running on a newer model, with different behaviour and price, without anyone changing the code. Check which model string your provider or developer actually calls.
Prices per million tokens, October 2026
Standard paid-tier list prices in US dollars per million tokens, from Google's Gemini API pricing, Anthropic's pricing page and OpenAI's API pricing. Approximate, as of October 2026, before tax.
| Model | Input | Output | Note |
|---|---|---|---|
| Gemini 3.8 Flash | $0.75 (then $1.50) | $3.75 (then $7.50) | Introductory price to 31 Dec 2026 |
| Gemini 3.6 Flash | $0.75 (then $1.50) | $3.75 (then $7.50) | Same schedule as 3.8 Flash |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | Text, image, video, audio input |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | Text, image, video input |
| Claude Haiku 5.5 | $0.10 | $0.50 | Prompts up to 100K tokens |
| GPT-6 Luna | $0.10 | $0.50 | Standard tier |
Google's output prices include thinking tokens. Google, Anthropic and OpenAI all list batch processing at half price, which suits overnight jobs such as tagging products but not a live chat.
Cost per 1,000 FAQ replies, worked through
We use the same assumptions as our guide to what one AI customer reply costs: each reply sends about 2,000 input tokens (instructions, FAQ text and the question) and returns about 200 output tokens. The formula is input tokens times input price plus output tokens times output price.
- Claude Haiku 5.52,000 x $0.10 + 200 x $0.50, per million, x 1,000US$0.30
- GPT-6 LunaUS$0.30
- Gemini 3.1 Flash-LiteUS$0.80
- Gemini 3.5 Flash-LiteUS$1.10
- Gemini 3.8 Flash (to 31 Dec 2026)US$2.25
- Gemini 3.8 Flash (from 1 Jan 2027)US$4.50
Per 1,000 replies, list prices as of October 2026. Excludes caching, which lowers the cost of repeated instructions, and extra reasoning tokens, which raise it.
At a few hundred chats a month, every option costs less than a cup of coffee in model fees. The difference matters at tens of thousands of replies, or when a bot sends long documents with every question.
Quality and speed: published benchmarks
Price per token says nothing about answer quality or how many reasoning tokens a model burns. Artificial Analysis publishes an Intelligence Index and the cost to run its test set. We did not run these tests ourselves; the figures are from its leaderboard on 21 October 2026.
| Model (setting) | Intelligence Index | Cost per task | First chunk (s) |
|---|---|---|---|
| Gemini 3.8 Flash (high) | 41 | $1.24 | 24.9 |
| Claude Haiku 5.5 (high) | 38 | $0.08 | 20.9 |
| Claude Haiku 5.5 (medium) | 34 | $0.05 | 9.4 |
| GPT-6 Luna (high) | 33 | $0.03 | 11.2 |
| Gemini 3.5 Flash-Lite | 22 | $0.19 | 7.3 |
| GPT-6 Luna (non-reasoning) | 18 | $0.01 | 0.9 |
Two points stand out. Gemini 3.8 Flash scores well but uses more tokens per task, so its cost per task is far higher than its token price suggests. And high reasoning settings make replies slow: a customer waiting 20 seconds for an answer about opening hours will leave. For FAQ bots, a low or no-reasoning setting is usually the right trade.
When Gemini Flash still makes sense
Pick Gemini Flash or Flash-Lite if
- You already build on Google Cloud or Vertex AI
- The bot needs Google Search or Maps grounding
- You process images, video or audio in the same flow
- Your own test shows clearly better answers on your questions
Pick Haiku 5.5 or GPT-6 Luna if
- The bot answers short text questions from a fixed FAQ
- Volume is high and cost per reply matters
- You want the lowest list price per token
- You need fast replies with little or no reasoning
Our guide to the best model for a customer service chatbot covers non-price factors such as refusal behaviour and language support, and the full AI API pricing comparison includes the larger models.
Run your own 30-minute test
Collect 30 real customer questions
Take them from your inbox, WhatsApp or chat logs, with personal details removed. Include a few awkward ones: complaints, questions you cannot answer, mixed English and Chinese.
Write one prompt and FAQ block
Use the same instructions and FAQ text for every model so the comparison is fair.
Run each candidate model on the paid tier
Use two or three models at a low reasoning setting. Record each answer, the token counts the API reports and the time to reply.
Score answers yourself
Mark each answer correct, partly correct or wrong, and note any invented facts. One invented refund policy matters more than a cent of savings.
Pick the cheapest model that passes
Then re-run the test whenever the vendor changes the model behind your model name.
Do it yourself or get help?
If you have a developer, switching models is usually a one-line change; the work is in the test and in keeping the FAQ content accurate. If you have no developer, a ready-made assistant is quicker. Our AI Website Assistant (S$790) is trained on your website, FAQ and documents, captures leads, hands over to WhatsApp or email and reports usage after 30 days. IT Sorted is a Singapore-based team of developers with over 10 years of experience; you pay upfront, send the brief in writing and follow progress on your order page.
Frequently asked questions
Google's Gemini API changelog lists Gemini 3.8 Flash as generally available on 2 September 2026. It is described as Google's most capable Flash model. Gemini 3.7 Flash, released in August, was deprecated on 8 October 2026 and its requests are routed automatically to 3.8 Flash.
Yes. Google's pricing page lists an introductory price of US$0.75 per million input tokens and US$3.75 per million output tokens through 31 December 2026, rising to US$1.50 and US$7.50 from 1 January 2027. Budget for the higher price if your bot will run into next year.
Not on October 2026 list prices. Claude Haiku 5.5 and GPT-6 Luna are both US$0.10 per million input tokens and US$0.50 per million output tokens, lower than any current Gemini Flash or Flash-Lite model. Gemini 2.5 Flash-Lite is similar in price, but Google now limits 2.5 access to existing users.
It is not a good idea. Google's pricing page marks content sent on the free tier as used to improve its products, and rate limits are lower. For a bot that handles customer names, phone numbers or orders, use the paid tier, where Google states content is not used to improve its products.
Models that reason before answering produce hidden thinking tokens, which are billed as output. A model with a moderate token price can use many more tokens per task. Artificial Analysis measures the total cost of running its test set, which captures this. For simple FAQ replies, a low reasoning setting keeps costs and delays down.