PII redaction is one of the most economically attractive AI workloads in 2026. The task is narrow, repetitive, and heavily input-driven. You are asking a model to find names, emails, phone numbers, addresses, account IDs, medical identifiers, or other sensitive spans inside text, then return either a masked version or structured spans for downstream redaction. That is not frontier-reasoning wizardry. It is disciplined pattern recognition with enough context to avoid stupid mistakes.
That matters because many teams still buy this workflow like they are commissioning a legal brief. They send every support log, contract, HR packet, and export dump to premium models by default. Then they act shocked when the monthly bill looks like a small payroll expense. The blunt answer is simple: most PII redaction should run on cheap or mid-tier models, with premium review reserved for the messy edge cases.
This guide uses current pricing from AI Cost Check, real model prices from the site catalog, and three practical workload shapes: short support logs, medium contracts or policy files, and large DSAR or HR export packets. If you already know the token math basics, great. If not, read the token guide after this. Either way, the recommendation will be the same: keep first-pass redaction cheap, keep output lean, and pay for premium review only when the risk justifies it.
What actually drives PII redaction cost
PII redaction costs are dominated by input tokens. The model has to read the document. The output should be relatively compact if you design the workflow correctly.
The cheapest production setup does not ask the model to rewrite the full document with every sensitive field replaced. That is the lazy version. The smart version asks for:
- sensitive spans,
- entity types,
- character offsets or quoted text,
- confidence scores,
- optional reasons only for ambiguous cases.
That output format keeps token usage under control and makes the system auditable. Your application can mask the text deterministically after the model returns the spans.
💡 Key Takeaway: PII redaction is cheap when the model returns spans, labels, and confidence. It gets expensive when you make the model rewrite whole documents for no good reason.
There are four cost levers that matter more than everything else:
- Document length. A 2,500-token support log and a 120,000-token export packet do not belong on the same pricing spreadsheet.
- Output style. Span lists are cheap. Full rewritten documents are not.
- Escalation rate. If only 5% of files need premium review, your budget stays sane.
- Context requirements. Long HR files, legal packets, and multi-document exports favor large-context models even when the nominal per-token price is higher than the absolute cheapest option.
For a broader intake-and-extraction view, pair this guide with AI OCR and document processing costs in 2026. OCR and redaction are adjacent problems, but they are not the same budget line.
[stat] $124/month The cost to redact 100,000 contract-sized files on GPT-5 nano using structured spans instead of premium-by-default review.
That number is the real headline. Redaction at scale is affordable. Sloppy architecture is what makes it annoying.
Baseline model pricing for redaction work
These are the most useful candidates from the current pricing catalog for a text-first redaction pipeline:
| Model | Input $/1M | Output $/1M | Context | Redaction fit |
|---|---|---|---|---|
| GPT-5 nano | $0.05 | $0.40 | 128K | Best for short, high-volume files |
| Mistral Small 3.2 | $0.10 | $0.30 | 128K | Cheap structured extraction |
| Llama 4 Scout | $0.08 | $0.30 | 10M | Best cheap option for very long documents |
| GPT-5 mini | $0.25 | $2.00 | 500K | Safe default when quality matters |
| DeepSeek V3.2 | $0.28 | $0.42 | 128K | Strong price-to-quality middle ground |
| Gemini 2.5 Flash | $0.30 | $2.50 | 1M | Strong long-context operational default |
| Claude Sonnet 4.6 | $3.00 | $15.00 | 1M | Premium escalation only |
| Claude Opus 4.7 | $5.00 | $25.00 | 1M | Rare, high-stakes exception lane |
Two opinions worth stating clearly:
- GPT-5 nano is the budget king for short, repetitive redaction jobs.
- Llama 4 Scout is the most interesting cheap option for long files because the 10M-token context removes a lot of chunking nonsense.
Premium Anthropic models are not bad choices. They are just bad defaults. If you send routine redaction work to Sonnet or Opus, you are paying review-lawyer prices for clerical labor.
That is the entire story in one line. Sonnet is useful. Sonnet-by-default is dumb.
Cost per document across three realistic workloads
To keep the math grounded, use these baseline document shapes:
- Short support log: 2,500 input tokens, 250 output tokens
- Contract or policy file: 20,000 input tokens, 600 output tokens
- Large DSAR or HR export: 120,000 input tokens, 2,000 output tokens
Those outputs assume structured spans, entity labels, and confidence metadata. They do not assume a full rewritten redacted document.
Scenario 1: short support logs and ticket transcripts
This is the common “remove customer identifiers before downstream analysis” case. Think support tickets, abuse reports, complaint summaries, CRM notes, or chat transcripts.
| Model | Cost per file |
|---|---|
| GPT-5 nano | $0.0002 |
| Mistral Small 3.2 | $0.0003 |
| Llama 4 Scout | $0.0003 |
| DeepSeek V3.2 | $0.0008 |
| GPT-5 mini | $0.0011 |
| Gemini 2.5 Flash | $0.0014 |
| Claude Sonnet 4.6 | $0.0113 |
| Claude Opus 4.7 | $0.0188 |
At this workload size, the premium tiers are economically absurd unless the redaction also includes delicate legal interpretation. For ordinary support and ops data, use GPT-5 nano or Mistral Small 3.2 and move on with your life.
Scenario 2: contracts, policies, and compliance documents
This is where people start overbuying. A contract is longer, yes. That does not mean it suddenly deserves Opus-level spending.
| Model | Cost per file |
|---|---|
| GPT-5 nano | $0.0012 |
| Llama 4 Scout | $0.0018 |
| Mistral Small 3.2 | $0.0022 |
| DeepSeek V3.2 | $0.0059 |
| GPT-5 mini | $0.0062 |
| Gemini 2.5 Flash | $0.0075 |
| Claude Sonnet 4.6 | $0.0690 |
| Claude Opus 4.7 | $0.1150 |
The cheap models are still cheap. The premium models are still expensive. The main difference is that long-context flexibility starts to matter more, which makes Scout and Gemini Flash more attractive than their raw per-token ranking might suggest.
📊 Quick Math: Redacting 1,000 contract-sized files costs about $1.80 on Llama 4 Scout, $6.20 on GPT-5 mini, and $69.00 on Claude Sonnet 4.6.
Scenario 3: DSAR packets, HR exports, and large mixed files
This is the ugly stuff: long employee records, data exports, combined case files, or giant internal packets that may contain identifiers across many sections.
| Model | Cost per file |
|---|---|
| GPT-5 nano | $0.0068 |
| Llama 4 Scout | $0.0102 |
| Mistral Small 3.2 | $0.0126 |
| GPT-5 mini | $0.0340 |
| DeepSeek V3.2 | $0.0344 |
| Gemini 2.5 Flash | $0.0410 |
| Claude Sonnet 4.6 | $0.3900 |
| Claude Opus 4.7 | $0.6500 |
The cheap lane is still viable, but this is where context-window convenience matters. If you can fit the whole file cleanly into Scout or Gemini Flash without chunking and reconciliation overhead, that is usually the correct engineering decision.
Monthly cost at scale, where the bad habits get expensive
Per-document math is useful for demos. Monthly math is what forces discipline.
100,000 short support logs per month
| Model | Monthly cost |
|---|---|
| GPT-5 nano | $22.50 |
| Mistral Small 3.2 | $32.50 |
| Llama 4 Scout | $27.50 |
| DeepSeek V3.2 | $80.50 |
| GPT-5 mini | $112.50 |
| Gemini 2.5 Flash | $137.50 |
| Claude Sonnet 4.6 | $1,125.00 |
| Claude Opus 4.7 | $1,875.00 |
100,000 contract-sized files per month
| Model | Monthly cost |
|---|---|
| GPT-5 nano | $124.00 |
| Llama 4 Scout | $178.00 |
| Mistral Small 3.2 | $218.00 |
| DeepSeek V3.2 | $585.20 |
| GPT-5 mini | $620.00 |
| Gemini 2.5 Flash | $750.00 |
| Claude Sonnet 4.6 | $6,900.00 |
| Claude Opus 4.7 | $11,500.00 |
This is the moment where “quality-first” hand-waving usually falls apart. If your first-pass redaction workflow really needs Sonnet on all 100,000 contracts, your prompts, chunking strategy, or validation design probably suck.
⚠️ Warning: The fastest way to blow your redaction budget is to ask the model to rewrite the full document. A 20K-token contract that returns a 20K-token redacted rewrite costs about $0.0076 on Llama 4 Scout instead of $0.0018, and $0.3600 on Claude Sonnet 4.6 instead of $0.0690.
That is not a subtle difference. It is a tax on laziness.
The routing policy I would actually ship
A real PII-redaction system should have three lanes.
Lane 1: deterministic pre-filter
Use regexes, validators, and known schema rules first. Email addresses, obvious phone numbers, SSNs, passport patterns, and database field labels should not require a model call every time. Save the model for contextual cases, not the low-hanging fruit.
Lane 2: cheap first-pass model
For most text, use GPT-5 nano, Mistral Small 3.2, or Llama 4 Scout depending on document length. Ask for spans, entity types, and confidence.
Lane 3: premium escalation
Escalate only when confidence is low, the document is regulated, the text is ambiguous, or the file mixes many identities and contexts. This is where GPT-5 mini or Claude Sonnet 4.6 earns its keep.
For 100,000 contract-sized files per month, a smart routed stack looks like this:
| Stack | Routing policy | Monthly cost | When to use |
|---|---|---|---|
| Cheapest high-volume stack | 95% GPT-5 nano, 5% GPT-5 mini | $148.80 | Support ops, standard privacy cleanup, moderate risk |
| Long-document stack | 95% Llama 4 Scout, 5% Claude Sonnet 4.6 | $514.10 | HR archives, policy packets, long mixed files |
| Premium-by-default mistake | 100% Claude Sonnet 4.6 | $6,900.00 | Use only if you enjoy lighting money on fire |
The second line is the one I would ship for serious long-form redaction. It buys context headroom and a premium escape hatch without letting the entire pipeline inherit premium pricing.
✅ TL;DR: Use deterministic rules first, cheap models for the bulk pass, and premium review only for ambiguous or regulated documents. That is the whole playbook.
If you also need labeling, downstream QA, or human-review economics, the companion read is AI data labeling costs in 2026. Redaction quality is not just model quality. It is also queue design.
Where teams waste money on redaction
1. They treat redaction like generation
Redaction is an extraction-and-masking problem. If your prompt reads like you are commissioning a polished rewrite, you are paying output-token rates for zero extra business value.
2. They use premium models to compensate for missing policy logic
If your rules for “what counts as sensitive” are not clearly defined, teams reach for a bigger model. That is backwards. Fix your taxonomy first. A sharp prompt and a validation layer beat brute-force premium spend.
3. They ignore chunking until long files explode
If the model cannot fit the file cleanly, the system becomes fragile. This is why AI contract review costs in 2026 and redaction economics overlap: long documents punish sloppy context management.
4. They never measure false positives separately from false negatives
Over-redaction harms usability. Under-redaction harms compliance. These are different failures. The right system tracks both and routes the uncertain cases. Throwing everything at Opus is not a measurement strategy.
5. They forget that auditability has token consequences
You do want confidence scores, labels, and reasons for borderline cases. You do not want verbose essay explanations for every field. Keep the default response terse and trigger expanded rationales only when the file is going to human review.
The right mental model is simple: redaction is a cheap bulk workflow with a small expensive exception queue. Build it that way.
Frequently asked questions
What is the cheapest model for AI PII redaction in 2026?
For short, high-volume text files, GPT-5 nano is the cheapest practical option in the current catalog at $0.05 input and $0.40 output per million tokens. For very long files, Llama 4 Scout is often the smarter cheap choice because its 10M-token context reduces chunking overhead.
How much does AI PII redaction cost per contract?
For a contract-sized file using 20,000 input tokens and 600 output tokens, the cost is roughly $0.0012 on GPT-5 nano, $0.0018 on Llama 4 Scout, $0.0062 on GPT-5 mini, and $0.0690 on Claude Sonnet 4.6. Use the calculator if your files are longer or your output format is heavier.
Should I use a premium model for privacy redaction?
No, not by default. Premium models belong in the escalation lane for ambiguous legal language, heavily mixed identity contexts, or regulated edge cases. For the bulk pass, cheap and mid-tier models are the correct economic choice.
Is it better to return spans or a fully redacted document?
Return spans unless you have a very specific reason not to. Span-based output is cheaper, easier to audit, and easier to combine with deterministic masking. Full rewritten documents increase output cost and make validation harder.
What is the best way to cut redaction costs without increasing risk?
Combine deterministic pattern rules, a cheap first-pass model, confidence thresholds, and a premium escalation queue. That architecture cuts cost harder than prompt tweaking and usually improves review quality at the same time.
Calculate your own redaction stack before shipping it
If you are budgeting a privacy workflow, do not guess. Run the numbers in AI Cost Check, then compare the surrounding workflows that usually travel with redaction: AI OCR and document processing costs, AI contract review costs, and what AI tokens are.
The correct strategy is boring and profitable: keep bulk redaction on cheap models, reserve premium reasoning for the ugly exceptions, and make the model emit the smallest useful output. That is how you protect privacy without turning compliance into a ridiculous API bill.