Blog · 2026-08-06 · Vynaris Team

What voice-of-customer mining costs per 10,000 support conversations

Mining themes from 10,000 support conversations costs $4.12 to $69.48. Flat extraction is $3.50 on DeepSeek; embedding plus weekly rollup add $0.62. Prices verified 2026-08-06.

Mining themes from 10,000 support conversations costs $4.12 on deepseek-v4-flash to $69.48 on Claude Sonnet 5 introductory pricing on this editable workload. Flat extraction alone is $3.50 on DeepSeek; the embedding-cluster pass and weekly theme rollup add $0.62. Prices verified 2026-08-06.

TL;DR

Each conversation is tagged at 2,000 input and 250 output tokens, then embedded at 2,000 tokens for clustering.
After dedupe, 200 theme clusters each get a 6,000/800 rollup. One weekly digest is 25,000/2,500.
Chat volume is 21,225,000 input tokens and 2,662,500 output tokens, plus 20,000,000 embedding tokens.
Model cost spans $4.1170 to $69.4750, a 16.9x spread. DeepSeek extract plus Sonnet rollup costs $7.9750, 88.5% below all-Sonnet.
Public demand evidence is the Agnost AI HN thread (85 points / 52 comments). Token volumes remain assumptions, not observed traffic.

The workload model

One outcome is 10,000 support conversations turned into theme briefs a product team can read. The pipeline has three stages that flat document extraction does not: per-conversation tagging, an embedding cluster pass, and a cluster-bounded weekly rollup.

Every count below is an assumption. None comes from Vynaris traffic, a customer ticket corpus, or a measured VoC product. Replace them with your tokenizer counts before budgeting.

Assumption       Value                                 Why it exists
---------------  ------------------------------------  -------------------------------------------
Conversations    10,000                                Outcome unit
Extract pass     2,000 in / 250 out each               Tags, severity, feature ask, evidence quote
Embedding        2,000 tokens each                     Conversation body for theme clustering
Theme clusters   200                                   Editable after dedupe (2% of volume)
Cluster rollup   6,000 in / 800 out each               Sample quotes plus theme brief
Weekly digest    25,000 in / 2,500 out                 Cross-theme product memo, once
Embedding price  text-embedding-3-small at $0.02 / 1M  First-party OpenAI rate
Batch discount   Shown separately                      Verdict table uses standard rates

This shape is distinct from invoice extraction. Invoice extraction maps each document to one record and optionally verifies a percentage. VoC mining pays an embedding line for every conversation, then pays rollup on themes, not on conversations. If volume grows 10x while themes stay near 200, extract scales linearly and rollup does not.

It is also distinct from the enterprise RAG assistant playbook. RAG amortizes embeddings for retrieval at query time. This pipeline embeds once to cluster feedback, then spends chat tokens writing theme briefs. The unit is cost per 10,000 conversations processed, not cost per answered query.

Verdict: cost per 10,000 conversations

Model                                                                      Extract 10k  Embed 10k  Rollup 200  Digest   Total / 10k  Per conversation
-------------------------------------------------------------------------  -----------  ---------  ----------  -------  -----------  ----------------
[deepseek-v4-flash](https://vynaris.com/models#deepseek-v4-flash)          $3.5000      $0.4000    $0.2128     $0.0042  $4.1170      $0.0004
[GPT-5.6 Luna](https://vynaris.com/models#gpt-5-6-luna)                    $7.0000      $0.4000    $0.4320     $0.0080  $7.8400      $0.0008
[Gemini 3.5 Flash-Lite](https://vynaris.com/models#gemini-3-5-flash-lite)  $12.2500     $0.4000    $0.7600     $0.0138  $13.4238     $0.0013
[Claude Haiku 4.5](https://vynaris.com/models#claude-haiku-4-5)            $32.5000     $0.4000    $2.0000     $0.0375  $34.9375     $0.0035
[Claude Sonnet 5](https://vynaris.com/models#claude-sonnet-5) intro        $65.0000     $0.4000    $4.0000     $0.0750  $69.4750     $0.0069

Standard prices per 1M input/output tokens are $0.14/$0.28 for DeepSeek, $0.20/$1.20 for Luna, $0.30/$2.50 for Flash-Lite, $1/$5 for Haiku, and $2/$10 for Sonnet through 2026-08-31. Anthropic lists Sonnet at $3/$15 from 2026-09-01. Embedding is $0.02 per 1M tokens on text-embedding-3-small.

Log-scale horizontal chart showing VoC mining model cost per 10k conversations from $4.12 on deepseek-v4-flash to $69.48 on Claude Sonnet 5, plus a routed DeepSeek-extract Sonnet-rollup bar at $7.98. — Voice-of-customer mining cost per 10k conversations. Source: first-party provider pricing verified 2026-08-06. Token counts and the 200-cluster count are editable assumptions.

The arithmetic behind $4.1170

DeepSeek on the full pipeline:

Extract: 10,000 x (2,000 x $0.14 + 250 x $0.28) / 1M = $3.5000
Embed:   10,000 x 2,000 x $0.02 / 1M                 = $0.4000
Rollup:     200 x (6,000 x $0.14 + 800 x $0.28) / 1M = $0.2128
Digest:           (25,000 x $0.14 + 2,500 x $0.28) / 1M = $0.0042
Total                                                   = $4.1170

Sonnet uses the same chat and embedding volumes:

Extract $65.0000 + embed $0.4000 + rollup $4.0000 + digest $0.0750 = $69.4750

Use the cost calculator for the chat half with your measured totals. Add the embedding line separately: tokens embedded times $0.02 per million.

Why clustering changes the unit economics

Flat extraction of the same 10,000 conversations on DeepSeek is $3.5000. The VoC-specific overhead is $0.6170: $0.4000 embedding plus $0.2128 rollup plus $0.0042 digest. That is 15.0% of the DeepSeek total and only 6.4% of the Sonnet total, because extract dominates expensive models.

Cluster count is the controllable VoC lever. Extract and embed scale with conversations. Rollup scales with themes.

Theme clusters  DeepSeek total  Sonnet total  DeepSeek rollup
--------------  --------------  ------------  ---------------
50              $3.9574         $66.4750      $0.0532
100             $4.0106         $67.4750      $0.1064
200             $4.1170         $69.4750      $0.2128
400             $4.3298         $73.4750      $0.4256
800             $4.7554         $81.4750      $0.8512

Going from 200 to 800 themes adds $0.6384 on DeepSeek or $12.0000 on Sonnet. Set the cluster threshold from how many briefs a product team will actually read, not from embedding cost. A vector database and semantic search over conversation embeddings are infrastructure lines outside this model bill.

At 100,000 conversations with the same 200 themes, DeepSeek rises to $39.2170 and extract is 89.2% of that bill. Rollup stays $0.2128. That is the merge-check against flat extraction: the unique cost surface is theme-bounded synthesis on top of linear tagging.

Where routing changes the unit economics

Run every extract on DeepSeek. Keep text-embedding-3-small for clustering. Spend Sonnet only on the 200 theme rollups and the weekly digest.

Stage         Model                   Cost / 10k
------------  ----------------------  ----------
Extract 10k   deepseek-v4-flash       $3.5000
Embed 10k     text-embedding-3-small  $0.4000
Rollup 200    Claude Sonnet 5 intro   $4.0000
Digest 1      Claude Sonnet 5 intro   $0.0750
Routed total                          $7.9750

The routed total is 88.5% below all-Sonnet and 1.94x all-DeepSeek. The extra $3.8580 over all-DeepSeek buys a stronger writer only on theme briefs a human will read.

Model routing earns that premium only if Sonnet rollups reduce PM edit time or improve theme precision on a labeled sample. If DeepSeek briefs need the same rewrite minutes, all-DeepSeek wins on cost.

Batch when the digest can wait a day

VoC mining is usually asynchronous. OpenAI, Anthropic, and Gemini publish Batch rates at half or near-half of standard for these chat models. Embedding stays at $0.02 / 1M in this sheet.

Batch path                   Total / 10k
---------------------------  -----------
GPT-5.6 Luna Batch           $4.1200
Gemini 3.5 Flash-Lite Batch  $6.9119
Claude Haiku 4.5 Batch       $17.6688
Claude Sonnet 5 intro Batch  $34.9375

Luna Batch at $4.1200 lands within $0.0030 of DeepSeek standard. If your stack is already on OpenAI and a one-day lag is acceptable, batch processing closes most of the provider gap without a second vendor. DeepSeek has no Batch row here because we did not find a first-party Batch price on its pricing page for this run.

Build notes for VoC mining

Emit structured outputs: theme tags, severity, feature request, quoted span, and an abstain flag. Free-form summaries are harder to cluster and audit.
Store conversation IDs with every tag and every cluster membership. Theme briefs without lineage are not actionable.
Cap cluster count. A 2,000-theme digest is not a product memo; it is a second ticket queue.
Prefer chunking long threads before embedding if your tokenizer blows past the assumed 2,000 tokens. Re-run the sheet with measured lengths.
Keep PII redaction outside the model line. Redaction failures are a privacy incident, not a cost-per-token win.
Measure theme precision and recall on a labeled week before trusting automated roadmaps. Cheap tags still invent feature requests.

When this workload does not need a router

Skip routing when DeepSeek rollups clear your edit bar. At $4.1170 per 10k conversations, a routing service built only to shave model spend is ceremony.

Skip the weekly LLM digest when a BI rollup of structured tags already answers the product question. If severity and feature-request fields are reliable, a SQL group-by may beat another generation pass.

Skip embeddings when you only need per-ticket classification into a fixed taxonomy. Fixed labels are flat extraction. Do not pay the cluster tax for a closed enum.

The honest tradeoff: if product leaders will not read theme briefs or change a roadmap from them, do not build the pipeline. The $0.62 VoC overhead is small next to Sonnet extract, but it is wasted when the output has no owner.

FAQ

What does voice-of-customer mining cost per 10,000 support conversations? On this workload, $4.1170 on deepseek-v4-flash to $69.4750 on Claude Sonnet 5 introductory pricing, including embeddings.

How is this different from document extraction? Extraction ends at structured fields per item. VoC mining adds an embedding cluster pass and theme rollups whose cost tracks cluster count, not conversation count.

How much do clustering and rollup add? On DeepSeek, $0.6170 above a flat $3.5000 extract. That is 15.0% of the DeepSeek pipeline total.

Should every theme brief use the strongest model? Not by default. Route Sonnet onto rollups only if labeled samples show better theme quality or less PM edit time.

Is Batch worth it? Luna Batch totals $4.1200, essentially matching DeepSeek standard, when a day of latency is acceptable.

Sources

OpenAI API pricing, captured 2026-08-06: GPT-5.6 Luna standard $0.20/$1.20 and Batch $0.10/$0.60; text-embedding-3-small $0.02 per 1M input tokens.
Anthropic pricing, captured 2026-08-06: Claude Haiku 4.5 $1/$5; Claude Sonnet 5 $2/$10 through 2026-08-31 then $3/$15; "Save 50% with batch".
Gemini API pricing, captured 2026-08-06: Gemini 3.5 Flash-Lite standard $0.30/$2.50 and Batch $0.15/$1.25.
DeepSeek API pricing, captured 2026-08-06: deepseek-v4-flash $0.14 cache miss and $0.28 output.
Launch HN: Agnost AI, page-verified 2026-08-06: 85 points / 52 comments. Demand signal for extracting feedback from conversations, not a measured cost receipt.

All conversation counts, token volumes, cluster counts, and rollup sizes are shown assumptions. Replace them before using these totals in a product budget.