GuidesOneMux

Gemini 4 Neon API Costs for Document Question Answering

Estimate Gemini 4 Neon document question-answering costs using explicit token, cache, and output assumptions, with reproducible Google and OneMux price comparisons.

AI

Summary · TL;DR

A worked 20-question document Q&A estimate shows how cached document tokens and thinking output affect Gemini 4 Neon costs, with Google Standard and OneMux catalogue rates compared on the same token counts.

Key Takeaways

  • In the illustrative 20-question workload, Google Standard token charges total $3.399 with the assumed cache hits, or about $0.170 per question.
  • Applying the published OneMux token rates to the same counts gives $1.281423, or about $0.064 per question, subject to applicable billing and cache conditions.
  • Without cache hits, the same token counts cost $25.059 at Google Standard rates; cache hits should be measured rather than assumed.
  • Output estimates must include both visible answers and thinking tokens, while citation quality needs a separate acceptance check.
On this page (12)

The short answer

For document question answering with Gemini 4 Neon, estimate the cost of an accepted, source-attributed answer, not just the price of a single prompt. The biggest variables are how much document text each question sends, how much of that text actually hits the cache, and how many output tokens—including thinking—the model uses.

In the illustrative workload below, 20 questions over a 400,000-token document bundle cost $3.399 at Google Standard token rates, assuming the document is cached on 19 follow-ups. Applying the published OneMux rates to the same token counts gives $1.281423. Neither figure is a quote or a measured API run.

OneMux publishes this article and provides model access through an OpenAI-compatible API. The Google pricing and cache details discussed here come from Ofox’s Gemini 4 Neon pricing guide; the OneMux rates come from its published catalogue snapshot.

Price snapshot: comparable token units

The OneMux catalogue snapshot was captured on November 21, 2026. The Google Standard prices below are the prices reported in the supplied Ofox guide, which says they were checked on November 19, 2026. Prices can change; check the provider and OneMux model page before budgeting.

Model and price basisInput per 1M tokensOutput per 1M tokensPublished OneMux-to-official ratio
Gemini 4 Neon, Google Standard$3.00$15.00Not applicable
Gemini 4 Neon, OneMux catalogue$1.131$5.6550.377

The ratio is the published price ratio, not a percentage discount. It compares the listed input and output rates; it does not guarantee the same ratio for every invoice or service configuration. Google’s output rate includes thinking tokens, according to Ofox. The worked comparison assumes the same output-token accounting on both price bases; confirm that treatment for your OneMux usage before relying on the estimate.

Cached-input price basisPrice per 1M cached input tokens
Google Standard, as reported by Ofox$0.15
OneMux catalogue cache read$0.05655

The OneMux snapshot also lists a cache-write price of $0.05655 per million tokens. That catalogue label alone does not establish when it applies or whether it is charged alongside another category. The example uses the cache-read price only for assumed hits and does not add a separate write charge. Check current cache and tier conditions before treating it as a forecast. Ofox reports that Google Neon uses implicit caching rather than explicit cache creation; a repeated document is not proof of a cache hit.

Worked estimate: 20 questions over one bundle

This is a synthetic document Q&A workload, not a test of Neon or OneMux. The application sends the document bundle as a shared prefix and asks for answers with citations to document passages.

  • First question: 400,000 total input tokens, no cache hit, and 6,000 output tokens including thinking.
  • Nineteen follow-ups: 402,000 input tokens each: a 400,000-token document prefix assumed to be a cache hit, plus 2,000 other input tokens. Each produces 3,000 output tokens including thinking.
  • Scope: one model request per question, no retries, search, file-search embedding fees, taxes, or other charges. The first question and its instructions are included within its 400,000 input tokens.
  • Attribution: source identifiers and quoted passages are included in the assumed input and output counts; their actual token cost will depend on the format used.

Across the workload, that is 438,000 uncached input tokens, 7,600,000 cached input tokens, and 63,000 output tokens. Cached input is a subset of total input, not an extra copy to charge at both input rates.

ChargeGoogle Standard arithmeticGoogle costOneMux catalogue arithmeticOneMux cost
Uncached input438,000 at $3.00 per 1M$1.314438,000 at $1.131 per 1M$0.495378
Cached input7,600,000 at $0.15 per 1M$1.1407,600,000 at $0.05655 per 1M$0.429780
Output, including assumed thinking63,000 at $15.00 per 1M$0.94563,000 at $5.655 per 1M$0.356265
Total for 20 questionsSum of charges$3.399Sum of charges$1.281423
Average per questionTotal divided by 20$0.16995Total divided by 20$0.06407115

The two columns use identical token counts to isolate the listed rate difference. They do not demonstrate that a particular request will receive the assumed cache treatment through either route.

What if the cache does not hit?

The same 20 requests contain 8,038,000 total input tokens. If all of them are billed as ordinary input, Google Standard input costs $24.114; adding $0.945 of output gives $25.059, or about $1.253 per question. At the listed OneMux ordinary input and output rates, the equivalent total is $9.447243, or about $0.472 per question.

This comparison holds the workload and output fixed. It illustrates cache sensitivity, not a prediction of cache-hit frequency. Record actual cached and uncached usage before using either scenario in a budget.

Turn the estimate into an attribution-aware budget

  1. Define an accepted answer. Require a document identifier and a page, section, or passage reference that a reader can locate. Decide whether an unsupported answer must be rejected or may state that the evidence is insufficient.
  2. Measure the full request. Include document text, question, instructions, passage labels, and any repeated conversation history. If retrieval selects passages instead of sending the whole bundle, measure the text actually sent; add any separately billed retrieval or embedding charges.
  3. Count all output. Include citations and thinking where the provider bills them as output. Visible answer length alone is not a reliable output-cost estimate.
  4. Check cache usage in records. Separate confirmed cache hits from ordinary input. Start with a no-hit case if you do not have usage evidence, then add a measured-hit scenario.
  5. Count failed attempts. A retry after an incorrect citation is another request. For example, if 20 initial questions cost $3.399 and four retries each cost $0.17, the total rises to $4.079. If all 20 questions are then accepted, that is about $0.204 per accepted answer. The retry cost is an assumption, not a Neon measurement.

A citation can be present but wrong: check that the referenced passage supports the claim and that document identifiers survive retrieval and formatting. A correct refusal when the bundle lacks evidence can also be a valid outcome. Track these outcomes separately from token spend so that a low cost per response is not mistaken for a low cost per useful answer.

Tier and tool charges to check

Ofox reports Google Batch and Flex token rates at half of Standard and Priority at 1.8 times Standard. Those are different service tiers, not interchangeable promises about turnaround or availability. The OneMux snapshot does not establish matching tier terms, so the table above compares only the stated catalogue rates with Google Standard.

The worked example uses no external tools. If the workflow adds Google Search grounding, file search, or another service, price those separately under the applicable terms. Ofox reports that Search can generate multiple billable queries from one request and that file search can add embedding charges. Do not assume an unlisted tool price is zero.

FAQ

Does a larger document always cost more per answer?

Not necessarily. It increases the text potentially sent, but actual cost also depends on retrieval, cache hits, output including thinking, and retries. Compare measured tokens per accepted answer.

Can I assume the whole document will be cached after the first question?

No. The 19 cache hits above are an explicit scenario. Confirm actual cache usage and applicable conditions in your provider’s records.

Conclusion

For a document Q&A budget, calculate both a no-cache baseline and a measured-cache case, then add retries and any separately billed tools. The resulting cost per accepted, attributable answer is more useful than a token price alone.

Sources

Published price snapshot

Captured 2026-11-21T09:00:06.588Z from the OneMux published catalogue. Prices retain website billing units; this is not a live benchmark. Tiered, cached or time-based billing may differ. Unknown values are not zero.

ModelOfficial inputOfficial outputOneMux inputOneMux outputPublished price ratio
Gemini 4 Neon$3$15$1.131$5.6550.377

Current OneMux prices. Compare workload costs only when input/output and billing units match; see each model page for tier conditions.

FAQ

How much does Gemini 4 Neon document question answering cost?

In the article’s illustrative 20-question workload, token charges are $3.399 at Google Standard rates with assumed cache hits, or $1.281423 using the published OneMux rates and the same token counts. Actual costs depend on usage, cache treatment, and additional charges.

Are Gemini 4 Neon thinking tokens included in output costs?

The cited Ofox guide reports that Google bills thinking tokens as output. Confirm the applicable output-token accounting for your OneMux usage before relying on the comparison.

Does repeating a document guarantee a cache hit?

No. The worked example assumes cache hits for illustration. Check actual usage records and the applicable provider conditions.

Related articles