On this page · 10 sections
Summary. On 16 November 2026 the lowest-priced current models from Anthropic and OpenAI share one list price: Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, and GPT-6 Luna costs the same $0.10 and $0.50. Gemini 3.5 Flash-Lite, the low-cost model Google recommends for new projects, costs $0.30 and $2.50, three times the input price and five times the output price. Google's older Gemini 3.1 Flash-Lite is cheaper at $0.25 and $1.50, but it has an earliest shutdown date of 7 May 2027 and Google names 3.5 Flash-Lite as its replacement, so teams on 3.1 face a price rise on output. Long prompts change the order: Haiku 5.5 bills the whole request at $0.50 and $2.50 once the prompt passes 100,000 tokens, while Luna doubles input and adds 50% to output only above 272,000 tokens. Our verdict: for short classification and extraction, Haiku 5.5 and Luna tie on price and the choice comes down to quality on your own data; for prompts between 100,000 and 272,000 tokens, Luna is the cheapest; Gemini 3.5 Flash-Lite earns its premium only where you need its audio and video input or Google Search grounding.
About this article. This comparison is for CTOs, platform engineers and product teams in India, the UK and the US who run high-volume model calls: ticket routing, document extraction and retrieval-augmented generation (RAG). We checked every price, limit and date on 16 November 2026 against Google's, OpenAI's and Anthropic's own pricing pages, model pages and deprecation tables. List prices change often, so re-check them before you commit a budget. This is not financial advice and no vendor paid for inclusion. Researched with AI assistance and checked against the primary sources listed in the references.
The price table on 16 November 2026
All prices are US dollars per million tokens on each vendor's standard paid tier, before tax, taken from the three vendors' pricing pages and OpenAI's Luna model page.
| Model | Input | Output | Cache read | Batch input / output |
|---|---|---|---|---|
| Claude Haiku 5.5, prompt up to 100K | 0.10 | 0.50 | 0.01 | 0.05 / 0.25 |
| Claude Haiku 5.5, prompt over 100K | 0.50 | 2.50 | 0.05 | 0.25 / 1.25 |
| GPT-6 Luna, prompt up to 272K | 0.10 | 0.50 | 0.01 | 0.05 / 0.25 |
| GPT-6 Luna, prompt over 272K | 0.20 | 0.75 | 0.02 | 0.10 / 0.375 |
| Gemini 3.5 Flash-Lite | 0.30 | 2.50 | 0.03 | 0.15 / 1.25 |
| Gemini 3.1 Flash-Lite (shutdown from 7 May 2027) | 0.25 | 1.50 | 0.025 | 0.125 / 0.75 |
The Luna rows above 272,000 tokens apply OpenAI's stated multipliers (2x input and cache, 1.5x output for the full request) and its rule that Batch and Flex cost half of Standard. OpenAI lists a 30-minute cache write for Luna at $0.125, and Anthropic lists $0.125 for five-minute cache writes on Haiku 5.5. Google charges $1.00 per million tokens per hour to store an explicit cache, on top of the $0.03 read price.
Two older, cheaper rows are worth knowing about but not building on. OpenAI's gpt-5-nano lists at $0.05 and $0.40 but stops working on 11 December 2026, and OpenAI names gpt-6-luna as its replacement. Google's gemini-2.5-flash-lite lists at $0.10 and $0.40 and is not deprecated, but Google now limits access to the 2.5 models to users who have used them before and tells new projects to start on 3.5 Flash-Lite or 3.8 Flash.
Gemini 3.1 Flash-Lite and the 7 May 2027 date
Google's deprecations table lists gemini-3.1-flash-lite, released on 7 May 2026, with a shutdown date of 7 May 2027, and names gemini-3.5-flash-lite as the recommended replacement. Google says these dates are the earliest a model might be retired and that it will confirm the exact date with advance notice. So treat 7 May 2027 as the first day the endpoint could disappear, not a promise it will last until then.
The move costs money. On the same workload, 3.5 Flash-Lite charges 20% more for input ($0.30 against $0.25) and 67% more for output ($2.50 against $1.50). For an output-heavy pipeline the bill rises by well over half. That is the moment to compare the alternatives rather than accept the default migration.
A second Google change lands sooner. From 1 December 2026, API calls to the deprecated gemini-2.5-flash route to 3.5 Flash-Lite and are billed at 3.5 Flash-Lite rates. The text price does not change, because 2.5 Flash also lists at $0.30 and $2.50, but the model behind your calls does, so re-run your evaluation set before that date. We track the OpenAI side of the same cycle in our GPT-5 and o3 shutdown plan.
Specifications that affect cost
| Item | Claude Haiku 5.5 | GPT-6 Luna | Gemini 3.5 Flash-Lite |
|---|---|---|---|
| Model ID | claude-haiku-5-5 | gpt-6-luna | gemini-3.5-flash-lite |
| Context window | 1M tokens | 1,050,000 tokens | 1,048,576 tokens |
| Max output | 128K tokens | 128,000 tokens | 65,536 tokens |
| Input types | Text and images | Text and images | Text, image, video, audio, PDF |
| Default thinking | Adaptive, medium effort | Reasoning, medium effort | Minimal |
| Long-prompt surcharge | Over 100K tokens | Over 272K tokens | None listed |
| Earliest retirement | Not before 7 October 2027 | None announced | None announced |
Sources for the table: Anthropic's models overview, OpenAI's Luna model page and Google's 3.5 Flash-Lite model page and migration guide. Anthropic's overview gives Haiku 5.5 a retirement date of not sooner than 7 October 2027. Google lists no shutdown date for 3.5 Flash-Lite, and OpenAI's deprecations page lists none for Luna.
The default thinking row matters more than it looks. OpenAI bills reasoning tokens as output tokens, and Anthropic reports thinking as part of billed output. Luna's reasoning effort defaults to medium, and so does Haiku 5.5's. Google says 3.5 Flash-Lite defaults to the minimal thinking level and recommends leaving it there for high-volume extraction, routing and classification. Left on defaults, a one-word classification label from Luna or Haiku can carry hundreds of hidden output tokens. Set Luna's effort to none or low and lower Haiku's effort setting for simple labels, then measure with our LLM token counter and your own logs.
Worked costs for three high-volume workloads
The numbers below use list prices only and assume thinking is turned down so that billed output is close to the visible answer. Rupee figures use the ECB euro reference rates of 13 November 2026, crossed to ₹94.24 per US dollar and ₹124.65 per pound sterling, rounded to two decimals.
Workload 1: ticket classification
One million requests a month, each with a 500-token prompt and a 10-token label. That is 500 million input tokens and 10 million output tokens.
- Haiku 5.5 and Luna: 500 x $0.10 = $50, plus 10 x $0.50 = $5, so $55 a month (about ₹5,183).
- Gemini 3.5 Flash-Lite: 500 x $0.30 = $150, plus 10 x $2.50 = $25, so $175 (about ₹16,492).
- Gemini 3.1 Flash-Lite, until it is retired: $125 plus $15, so $140.
Workload 2: document extraction
200,000 documents a month, 4,000 input tokens each and 600 tokens of JSON out. That is 800 million input and 120 million output tokens.
- Haiku 5.5 and Luna: $80 plus $60, so $140 (about ₹13,194). Through each vendor's batch API it halves to $70 (about ₹6,597).
- Gemini 3.5 Flash-Lite: $240 plus $300, so $540 (about ₹50,890). Through batch, $120 plus $150, so $270.
Extraction rarely needs an answer within seconds, so batch is the single biggest saving here for all three vendors.
Workload 3: RAG question answering
300,000 questions a month, each sending 8,000 tokens (instructions plus retrieved chunks) and getting a 400-token answer. That is 2,400 million input and 120 million output tokens.
- Haiku 5.5 and Luna: $240 plus $60, so $300 (about ₹28,272).
- Gemini 3.5 Flash-Lite: $720 plus $300, so $1,020 (about ₹96,125).
- Gemini 3.1 Flash-Lite: $600 plus $180, so $780.
If the first 2,000 tokens of every prompt are the same instructions and are served from cache, 600 million tokens move to the cache-read price. Haiku 5.5 and Luna then cost $180 plus $6 plus $60, so $246 (about ₹23,183), and 3.5 Flash-Lite costs $540 plus $18 plus $300, so $858 (about ₹80,858). These totals leave out cache-write charges and Google's hourly storage fee, which depend on traffic pattern.
| Workload per month | Haiku 5.5 | GPT-6 Luna | Gemini 3.5 Flash-Lite |
|---|---|---|---|
| Classification, 1M short calls | $55 | $55 | $175 |
| Extraction, 200K documents | $140 | $140 | $540 |
| Extraction via batch | $70 | $70 | $270 |
| RAG, 300K questions | $300 | $300 | $1,020 |
| RAG with 2K cached prefix | $246 | $246 | $858 |
Where long prompts flip the ranking
The tie between Haiku 5.5 and Luna breaks once prompts grow. Anthropic prices each Haiku 5.5 request by its total prompt length, cache reads and writes included, and a request over 100,000 tokens pays the higher rates for the whole request. OpenAI applies its long-prompt multipliers only above 272,000 input tokens. Google lists one flat price for 3.5 Flash-Lite.
Per 1,000 requests with 1,000 output tokens each:
| Prompt size | Haiku 5.5 | GPT-6 Luna | Gemini 3.5 Flash-Lite |
|---|---|---|---|
| 150,000 tokens | $77.50 | $15.50 | $47.50 |
| 300,000 tokens | $152.50 | $60.75 | $92.50 |
The arithmetic for the 150,000-token row: Haiku 5.5 is 0.15 x $0.50 plus 0.001 x $2.50, so $0.0775 a request; Luna is 0.15 x $0.10 plus 0.001 x $0.50, so $0.0155; Flash-Lite is 0.15 x $0.30 plus 0.001 x $2.50, so $0.0475. At 300,000 tokens, Luna's rates become $0.20 and $0.75, giving $0.06075 a request. If your RAG design stuffs whole contracts or case files into the prompt, Luna is the cheapest of the three by a wide margin, and Haiku 5.5 is the most expensive. Keeping prompts under 100,000 tokens with tighter retrieval is usually the better fix; we compare that trade-off in long context against RAG.
Which one is capable enough
Price per token is only half of cost per task. A cheaper model that needs a retry, a longer prompt or more output tokens can cost more per correct answer. All three vendors aim these models at the same jobs: Anthropic describes Haiku 5.5 as built for classification, extraction and routing, OpenAI calls Luna its most efficient model for focused, high-volume tasks, and Google pitches 3.5 Flash-Lite at high-volume agentic work, simple data extraction and document parsing.
We did not run our own benchmark for this piece, so we make no quality ranking. Our Gemini 3.8 Flash, Haiku 5.5 and Luna comparison covers published third-party scores. For a decision, build 200 to 500 labelled examples from your own traffic, run all three with thinking turned down, and compare accuracy, output length and cost per correct answer.
Gemini 3.5 Flash-Lite has three real advantages that can justify paying more:
- Input types. It accepts audio, video and PDF directly; Haiku 5.5 and Luna take text and images.
- Grounding. Google Search grounding is built in, with 5,000 free search requests a month shared across Gemini 3 and newer models, then $14 per 1,000, subject to a daily cap of 1,500 requests.
- Free tier. 3.5 Flash-Lite has a free tier for prototyping, though Google states free-tier data may be used to improve its products and paid-tier data is not.
Region and residency charges
Residency settings add a premium on two of the three. Anthropic applies a 1.1x multiplier to all token prices when you request US-only inference, including Haiku 5.5's long-prompt rates. OpenAI charges a 10% uplift on regional processing endpoints for models released on or after 5 March 2026, and lists EU data residency for Luna on Standard, Fast, Flex and Batch. Indian teams handling personal data under the DPDP Act should decide where inference runs before comparing prices, because the premium applies to every token.
How to pick, in order
- Short calls under 100,000 tokens: shortlist Haiku 5.5 and Luna, which tie on list price, and choose on your evaluation results.
- Prompts between 100,000 and 272,000 tokens: Luna, which charges a fifth of Haiku 5.5's input rate in that band.
- Anything that can wait: use batch, which halves the price at all three vendors.
- Audio, video or Search grounding: Gemini 3.5 Flash-Lite, accepting a 3x to 5x higher token price.
- On Gemini 3.1 Flash-Lite or 2.5 Flash today: re-test before 1 December 2026 and before May 2027 rather than taking the default replacement.
A routing layer that sends each call to the cheapest model that passes its test set usually beats committing to one vendor.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our RAG knowledge assistant service builds retrieval pipelines and runs model evaluations on your own documents so you can pick the cheapest model that meets your accuracy bar. Talk to us at /contact-us/.
Last updated: 16 November 2026.