GPT-6 Luna vs Gemini 3.5 Flash-Lite for Text Extraction

For text-only extraction, GPT-6 Luna starts with lower standard token prices; Gemini 3.5 Flash-Lite explicitly documents audio and video inputs alongside text and images. Choose using correctly extracted records, not just a valid JSON response. Neither price nor a general benchmark establishes which model will read your documents more accurately.

This comparison uses official documentation checked September 25, 2026. We have not measured a head-to-head extraction run. Dollar amounts are direct vendor USD rates and hypothetical usage calculations, not Ofox prices or observed production savings.

Compare the same task first

A useful common task is extracting supplier, invoice number, currency, total and supporting text from an already extracted document string. That separates extraction reasoning from PDF parsing, image preprocessing and speech transcription. If your real workload starts with scans or audio, test that full pipeline separately; a text-only result cannot settle its cost or quality.

Sources: Luna model documentation and Gemini model documentation. The two input lists are not a like-for-like file-format matrix: Luna’s text/image modality list alone does not establish whether a given API accepts PDF files. Verify file-input support and billing at the endpoint you plan to use. Capacity is not a recommendation to fill the window. Retrieve the relevant document region, preserve page or line references and leave room for the answer.

Work out the token bill without inventing a winner

For a short text request under Standard processing, the rates per million tokens are:

A hypothetical set of 1,000 documents using 2,000 ordinary input and 200 billed output tokens each totals 2M input and 0.2M output tokens. Luna costs $0.30; Gemini costs $1.10. This excludes caching, retries, reasoning differences, tools, preprocessing, regional charges and taxes. Equal token volumes are a calculation assumption: the same text can tokenize differently, and actual billed output may include reasoning.

Do not multiply a single long prompt into this short-request example. Luna inputs above 272K tokens trigger double input/cache rates and 1.5 times output rates for the full request. Gemini’s listed caching storage costs $1 per million tokens per hour, in addition to the cache-read rate. Luna lists a separate $0.125 per million cache-write rate at short input. These mechanisms need their own workload model. See OpenAI pricing, Google pricing and the Luna billing guide.

A valid schema can still contain a false total

Define missing data explicitly. If the source does not identify the currency, return a missing-value representation allowed by your schema; do not infer USD because the total contains a dollar sign. Preserve the original evidence span alongside normalized values so a reviewer can trace each field.

A useful acceptance table might look like this:

Google’s structured-output documentation explains the format constraints; the evaluator still needs semantic checks. Use the same field definitions on both sides, while adapting the request schema to each provider’s supported interface. A schema rejection is an integration result, not proof that the model cannot extract the field.

Build a small, revealing evaluation set

Include clean records, missing fields, conflicting totals, locale-specific decimal separators, duplicated pages and source text containing instructions unrelated to extraction. Treat document instructions as data. Keep the expected records outside the model prompt and inspect failures by category rather than averaging away a dangerous error.

Separate structural validity from field accuracy, unsupported-value rate and complete-record acceptance. Count refusal, truncation, timeout and malformed output explicitly. Record all retries and manual review; otherwise a low token bill can hide expensive cleanup. The Luna extraction workflow provides a practical validation starting point.

Choose the pipeline you can actually operate

Luna is a reasonable first candidate when clean text, a compact schema and token cost dominate. Gemini is a reasonable first candidate when its documented audio or video inputs match your ingestion requirements. For PDF workflows, verify each endpoint’s file support and preprocessing before treating one as an operational advantage. These are evaluation priorities, not tested quality rankings.

If easy records work with a small model but exceptions need more work, evaluate an explicit escalation rule. Include the cost of incorrectly accepted records as well as escalation fees; see Luna-to-Sol routing. Publish the eventual decision with the dataset scope and run conditions, so “best for extraction” means a defined job instead of every document anyone might upload.