Customer feedback analysis

Gemini 3.6 Flash vs GPT-5.6 Luna
for summarizing customer feedback

This page compares two models on one job: turning open-text survey answers or support tickets into a themed summary. It covers coverage, cost and prompting, and ends with a fair way to test both on your own batches.

Aug 12, 2026 · 10 min read

The bottom line
Gemini covers more and Luna costs less

Gemini 3.6 Flash is the safer single-model choice when one run must cover a large, undivided feedback export without losing rare but important complaints. GPT-5.6 Luna is the better price-first choice once comments are already sharded below its pricing threshold.

That split rests on Google's own long-context retrieval test1, a broader knowledge-work benchmark1 and the published token prices32, not on a dedicated survey-theming benchmark, since none exists publicly for these exact models.

For a staged workflow, use Gemini for initial theme discovery across the full corpus, especially above roughly 272,000 tokens, and use Luna for row-level coding or summarizing pre-clustered shards, where its lower base price applies. If only one model will be deployed, Gemini's stronger published long-context evidence makes it the lower-risk choice for preserving coverage.

Who this is for
Which feedback roles this fits

Start with Gemini01

Insight and research teams

You run one large batch a quarter and need every recurring issue to surface, not just the loudest ones. Gemini's long-context evidence points to fewer missed themes.

Try Luna for shards02

Support ops triaging tickets

You process tickets in smaller, regular batches and care most about cost per run. Luna's lower base price suits frequent, sharded coding work.

Start with Luna03

Teams with small frequent runs

Your exports rarely approach the long-context pricing threshold, so the per-token savings compound across many runs a month.

Use Gemini then audit04

Teams guarding rare complaints

A single missed safety or churn signal is costly. Start with Gemini's one-pass coverage, then have an analyst check that every rare theme still has real evidence behind it.

What we compared
Coverage quality not the app

This page compares the two models through their API in one neutral setup, not one model inside one survey tool against the other inside a different dashboard.

The parts that matter for feedback analysis are covering the whole batch, keeping themes distinct rather than synonyms, tying every example back to real rows, and cost at the batch sizes teams actually run. Official docs come first, then the closest independent long-context and knowledge-work evidence.

We left tools out of the spec table on purpose. A survey platform's dashboard, tagging UI or export pipeline depends on the app around the model, so the same model can behave very differently in a chat product, in the API or inside a workspace. Judging those here would compare software, not theme discovery.

Specs at a glance
The batch-relevant numbers

The model facts that actually affect a feedback run. Tool features are left out, since they change with the app around the model.

Spec
Gemini 3.6 Flash
GPT-5.6 Luna
Why it matters
Context window
1,048,576 input tokens
1,050,000 tokens
Room for a large export of comments in one request12
Max output
65,536 tokens
128,000 tokens
Luna can return a longer structured summary in one pass12
Standard price
$1.50 in / $7.50 out per million
$0.20 in / $1.20 out per million, up to 272,000 input tokens
Luna is cheaper per token at every batch size, most of all under its threshold32
Batch price
$0.75 in / $3.75 out per million
$0.10 in / $0.60 out per million, up to 272,000 input tokens
Both discount roughly in half for asynchronous runs with turnaround up to about a day413
Long-context price
Same rate across the full window
$0.40 in / $1.80 out per million above 272,000 input tokens
Luna's rate rises here but still stays well under Gemini's flat rate32
Reasoning controls
Minimal, low, medium and high thinking levels
None, low, medium, high, xhigh and max reasoning effort
Both let a team trade quality for speed and cost per call52
Structured output
Schema-constrained JSON
Schema-constrained JSON
Both can enforce a fixed theme, row-ID and sentiment schema112

Figures from Google and OpenAI documentation, checked August 12, 2026. The two vendors price and tokenize differently, so treat any cross-model cost comparison as directional, not exact.

Head to head
Where each model wins on feedback

The answer changes by batch size and stage, not by brand. This is the main analysis: which model has the edge on each part of turning feedback into themes, and what backs it up.

Job
Better choice
Why the edge exists
Best evidence
Coverage across one large corpus
Gemini 3.6 Flash
Google's exact-model long-context test shows a wide retrieval gap in Gemini's favor. It is vendor-run rather than a survey benchmark, but it is a like-for-like indication of finding dispersed evidence.
Gemini scores 91.8% against Luna's 74.8% on the GDM-MRCR test at 128,000 tokens1
Retention beyond 272K tokens
Gemini, directional
OpenAI's own results show Luna's retrieval falling away at very long context. No exact Gemini figure appears in that table, so this documents a limit in Luna rather than settling a head-to-head. The matched comparison is the GDM-MRCR row above, which is where this verdict actually comes from.
OpenAI reports 41.3% retrieval for Luna in the 512K to 1M range6, with the like-for-like comparison coming from GDM-MRCR at 128,000 tokens1
Synthesis after the data is already chunked
Luna, slight edge
A broader knowledge-work benchmark gives Luna a higher score than Gemini. The benchmark is wider than feedback analysis, so it supports only a small edge for final synthesis once themes are pre-sorted.
Google's own table reports 1,584 Elo for Luna against 1,421 for Gemini on GDPval-AA v21
Output throughput
Gemini at matched high effort
Independent measurement found Gemini generating output faster at high reasoning effort. Infrastructure changes can move these figures, and short summaries may be dominated by input processing rather than output length.
Artificial Analysis measured about 221 output tokens per second for Gemini against 147 for Luna, both at high effort714
Cost below Luna's pricing threshold
GPT-5.6 Luna
Its standard and Batch input and output prices are lower than Gemini's at every batch size, and the gap is widest whenever the prompt stays under 272,000 input tokens.
Luna lists $0.20 input and $1.20 output against Gemini's $1.50 and $7.50 per million tokens23
Cost for a single very large prompt
GPT-5.6 Luna, narrower margin
Luna's rate rises above 272,000 tokens, but even after that surcharge it stays below Gemini's flat per-token rate. The threshold narrows Luna's cost edge without handing the edge to Gemini.
Above 272,000 input tokens Luna's full request costs $0.40 input and $1.80 output per million, against Gemini's flat $1.50 and $7.5023
Stable machine-readable output
Tie
Both exact models support schema-constrained structured output, so either can hold a fixed theme, row-ID and sentiment schema.
Both vendors document structured output support for these exact models112

Better-choice calls map to dimensions the sources actually evaluated. Where the evidence is a vendor's own long-context or knowledge-work test rather than a survey-theming benchmark, the row says so.

How to test
A fair test on your own batches

A useful test feels boring. Same prompt, same row IDs, same schema, no editing before scoring. Then judge what your team actually pays for: does every important issue show up, are themes distinct, and can the counts be reproduced from the returned row IDs.

Sample01

Pick three to five batches

Cover the range: a normal survey batch, a large support export, a multilingual batch, a batch containing rare but critical complaints, and a batch with overlapping themes.

Prompt02

Give both the same prompt

One prompt, source text, row IDs and output schema for both. Neither model gets a richer version. If you change the prompt mid-test, apply the change to both.

Setup03

Use the same setup

Match the reasoning level and run both through the API configuration the team will deploy. Chat-product behavior can differ from the API, so test where the work will happen.

Scoring04

Score without editing first

Do not edit outputs before scoring. Check whether every important issue is represented, whether examples really support their assigned theme, and how much analyst editing is required. For commercial work, hide the model names and use at least two reviewers.

What the evidence shows
Directional not decisive

The public evidence located does not include a dedicated, independently run benchmark of these exact models on open-text theme discovery. Long-context recall, broad knowledge work and API speed are proxies. Here is what each source helps judge.

Source
What it measures
What it suggests
How to weigh it
Google's GDM-MRCR long-context test
Finding and connecting facts across a large input
Gemini keeps a clear retrieval lead at 128K tokens
Vendor-run and not a survey benchmark, but the clearest exact-model long-context signal1
GDPval-AA v2 knowledge-work Elo
Graded professional deliverables across occupations
Luna scores modestly higher than Gemini
Broader than feedback analysis, so it supports only a small synthesis edge1
Independent topic-modeling research
Whether LLM-based topic methods stay specific on domain-heavy text
Models can produce accessible but overly generic topics without human supervision
A general warning about theme genericity, not a score for either exact model9

The practical conclusion is a two-stage pipeline: clean and code individual comments first, then merge codes into themes. Do not ask either model to produce an unsupported executive narrative directly from a raw export.

How to prompt each one
A different prompt for each stage

The best prompt is not the same for both. Matching the prompt to the model does more for theme quality than the model choice alone.

Gemini 3.6 Flash does best when you give it the full or largest practical corpus, insist on evidence IDs and separate discovery from consolidation. Medium thinking is a balanced default, and it is worth testing low when latency matters5.

GPT-5.6 Luna does best on inexpensive, parallel row-level coding, with shards kept below the long-context tier where possible and explicit rules for how codes get merged later2.

A Gemini 3.6 Flash prompt: full corpus and evidence IDs

Inductively identify 6-10 themes from these comments.

For each theme return a precise definition, included row
IDs, excluded near-matches, frequency, sentiment mix and
three verbatim evidence excerpts.

Preserve rare themes with operational impact.
Merge themes only after checking every row.

A GPT-5.6 Luna prompt: row-level coding before the final merge

Code each row into one or two provisional themes using
the supplied schema. Do not create the final taxonomy yet.

Return row ID, provisional code, confidence and
supporting phrase.

In the second pass, merge only codes with the same
underlying customer need.

Weak spots
And how to fix them

Neither model is perfect for this job. The useful question is where each one adds cleanup work, and what to change in the prompt.

Model
Weak spot
What it looks like
How to fix it
Gemini 3.6 Flash
Can create polished but broad themes
A tidy-looking taxonomy that hides meaningful sub-issues, a general risk documented for LLM topic methods.
Require inclusion and exclusion rules, row IDs, minimum evidence and a separate 'rare but important' section, then have an analyst review before final counting.
GPT-5.6 Luna
Weaker coverage on one very large prompt
Missed or unevenly weighted evidence above its long-context pricing threshold, where its per-token rate rises too, even though it stays cheaper than Gemini's flat rate.
Shard by time period, product area or deterministic row ranges. Code each shard, deduplicate the provisional codes, then run a smaller consolidation pass.
Both
Can merge neighboring themes or state unsupported prevalence
One comment assigned inconsistently, or a stated frequency that cannot be traced back to specific rows.
Return row-to-theme mappings, calculate counts outside the prose summary, and use blind human review for high-impact conclusions.

Which one to choose
Start from your batch size

One question first. Will the full batch exceed roughly 272,000 input tokens? Then follow the branch that matches most of your feedback pipeline.

Will the batch exceed 272K tokens? Yes, need one-pass coverage Yes, but sharding is acceptable No, cost per run dominates Rare feedback must not disappear Taxonomy is already fixed Gemini 3.6 Flash Test both, lean toward Luna GPT-5.6 Luna Start with Gemini then audit GPT-5.6 Luna

A starting point, not a rule. Test on your own batches before you commit.

Recommendations
Pick by your feedback pipeline

If the batch is large and undivided and rare or safety-critical feedback must not disappear, start with Gemini 3.6 Flash and audit row coverage afterward. Its long-context retrieval evidence is the clearest signal for that risk1.

If the pipeline already shards the corpus, or cost per recurring run is the priority, start with GPT-5.6 Luna at low or medium reasoning. Confirm its lower small-batch cost survives the extra consolidation pass before standardizing23.

For strict JSON output, either model works. Decide based on coverage needs and total pipeline cost rather than schema support alone, since both APIs enforce structure equally well112.

If the goal is wiring either model straight into a survey platform or ticketing system to auto-generate theme reports with no analyst reviewing the output, Playgram is not the right tool. It is a shared chat workspace for people, not a developer API, so that kind of pipeline integration means calling Gemini 3.6 Flash or GPT-5.6 Luna directly instead.

Bottom line
Coverage wins and cost decides later

Gemini 3.6 Flash is the safer single-model choice for turning a genuinely large feedback export into a small, evidence-backed theme set in this exact comparison. GPT-5.6 Luna is the better price-first choice for smaller shards or fixed-taxonomy coding.

The conclusion is not based on a direct survey-theming benchmark. Vendor long-context benchmarks use different ranges and methods, independent performance measurements move with infrastructure, and broad knowledge-work scores do not measure theme coverage directly. Prices and limits can also change quickly.

The safest final step is to test the shape of your own feedback, not a generic prompt from the internet. A fair test needs the same setup for both models: the same export, the same schema and the same place to run them, so the result reflects the models and not the tool around them. In practice that is harder than it sounds, since most teams end up running one model in one dashboard and the other in a different tool, on two separate subscriptions, which tilts the comparison before the first theme comes back. The cleaner the setup, the more the difference you see is really Gemini 3.6 Flash vs GPT-5.6 Luna, and not just which one happened to be easier to reach that day.

Test both in one workspace
Right here inside Playgram

That's the practical case for the steady setup just described, and it also makes recurring review cycles easier. When both models sit in one workspace, an insights team can send the same export to each, compare the theme sets side by side, and hand a shard from one model to the other without setting it up again.

Take one real support export, the kind with a long tail of rare complaints buried in it, and run that exact comparison in Playgram: paste the comments once, put the theming prompt in front of the latest Gemini and GPT models, and keep refining with whichever one holds coverage better, without re-uploading the export or starting a new session for the second opinion.

The same memory carries across the team too, not just this one comparison, over the latest GPT, Claude, Gemini and Grok models and many more, all in one place12. The line-up is curated, so retired models are turned off and new ones are added as they ship.

Team memory

Shared across everyone and every model.

Project memory

Scoped to a campaign or document set.

Personal memory

Your own working style, kept private.

Fair pricing
Pay per usage, not per seat

Upgrade as needed, and only pay for what you actually use

Save ~17% with the annual plan

Pro

$50/ month

Perfect for small and medium teams

Unlimited users & infinite memory

Multi-LLM chats

Granular access control to models

EU data residency

Get started

Ultra

$200/ month

Best for large, growing teams

Unlimited users & infinite memory

Multi-LLM chats

Unlimited use of DeepSeek V4 Flash

Granular access control to models

Choose US or EU data residency

Get started

Enterprise

Get in touch

Unlimited Credits

For organizations with advanced needs

Unlimited users & SSO

Priority Support

Unlimited use of DeepSeek V4 Flash

Granular access control to models

Choose US or EU data residency

Book a call

30-days money back guarantee

Frequently asked
questions

Gemini 3.6 Flash is the safer choice when a single run must cover a large, undivided export without losing rare but important complaints. On Google's own long-context test, Gemini scores 91.8% against Luna's 74.8% at 128,000 tokens. GPT-5.6 Luna scores slightly better on a broader knowledge-work benchmark and costs less per token at every batch size we priced, an edge that is largest below its 272,000-token pricing threshold, which makes it attractive once comments are already sharded into smaller batches.

GPT-5.6 Luna is cheaper at every batch size we priced. Below its 272,000-token threshold, its standard rate of $0.20 per million input tokens and $1.20 per million output is well under Gemini's $1.50 and $7.50. On a 250,000-input, 5,000-output run, that works out to roughly $0.06 for Luna against $0.41 for Gemini at standard rates. Above 272,000 input tokens, Luna's full request is billed at $0.40 input and $1.80 output, which narrows the gap but does not close it: a 500,000-token run costs about $0.21 for Luna against Gemini's flat-rate $0.79. Both offer a Batch endpoint at roughly half the standard price for turnaround up to about 24 hours.

It is a real risk for both, and more so once a prompt exceeds either model's strongest tested range. Independent research on LLM-based topic methods found that models can produce polished but overly generic themes that hide meaningful sub-issues. Require row-level IDs, a minimum evidence count and a dedicated 'rare but important' section in the output, and have an analyst review the taxonomy before final counting.

Shard once the export is likely to exceed roughly 272,000 tokens: that is also where GPT-5.6 Luna's per-token price rises, though even after that rise it stays cheaper than Gemini's flat rate. The stronger reason to shard at that point is coverage, not cost: Gemini's published long-context retrieval score is the better-tested case for handling one very large prompt without losing rare themes. Either way, the safer pattern is a two-stage pipeline: code individual comments first, then merge the codes into themes, rather than asking either model for an unsupported executive summary straight from a raw export.

Yes. Both APIs support schema-constrained structured output, so a team can require fields such as theme name, row IDs, sentiment and frequency on both sides. Schema support does not by itself guarantee accurate counts or well-separated themes, so calculate counts outside the prose summary and check that every listed example actually supports its assigned theme.

No. Run the same prompt on both and compare the answers, or switch between them mid-conversation. You choose after reading both answers instead of guessing up front.

Related comparisons

Claude Sonnet 5 vs Gemini 3.6 Flash for marketing copyClaude Opus 5 vs Gemini 3.6 Flash for finance memosGrok 4.5 vs Gemini 3.6 Flash for competitor battlecardsGPT-5.6 Terra vs Gemini 3.6 Flash for presentation outlines

One export for
both models

Send the same survey or ticket export to the latest Gemini and GPT models, keep the schema in one place, and see which one holds coverage with less shard-and-merge work. Set it up in a minute.

Get startedSee the pricing