Comparison · 2026

Multimodal prompt-injection scanner pricing comparison (2026)

If you are shopping for a prompt-injection defender that covers images or audio — not only text — the field is smaller than it looks. Here is what the five credible options cost, what modality each one actually covers, and where the self-serve tier starts.

TL;DR

Four of the five best-known PI defenders are text-only by design or by gating. Only Lakera Guard and Glyphward ship multimodal coverage as a self-serve product in 2026. Lakera sits at $99+/mo and has pushed upmarket since the Check Point acquisition; Glyphward sits at $29/mo with image and audio included. LLM Guard (free, open source) and Promptfoo (free-tier eval) are excellent complements — just not real-time multimodal scanners.

The five options, side by side

Product Image PI Audio PI Self-serve tier Where it sits
Lakera GuardYes (text-in-image; emphasis on text)Limited~$99+/mo (enterprise gated beyond)Text-first, multimodal roadmap; pushed upmarket post Check Point acquisition (per public reporting, Sept–Nov 2025)
LLM Guard (open source)NoNoFree (self-hosted)Text-only by design; excellent complement to a multimodal scanner, not a replacement
Azure Prompt ShieldsText onlyNoPer-call on Azure Foundry (PPU)Azure-gated; cannot be used outside Azure tenant boundaries
PromptfooEval-time, not inference-timeEval-time, not inference-timeFree OSS + Cloud Team $50/moTest harness for red-team suites; not a real-time scanner on live traffic
GlyphwardYes (FigStep, AgentTypo, typographic PI, rendered glyphs)Yes (WhisperInject-class, waveform anomalies, transcript filter)Free 10/day · $29/mo Pro · $99/mo TeamMultimodal-first self-serve; starts below the enterprise tier

Figures are drawn from public vendor websites as of April 2026. Private quotes and enterprise tiers frequently differ — if a specific line has moved since we wrote this, email hello@glyphward.com and we will update it.

How to read the table

“Covers a modality” in this market usually means one of three things. Real-time scanning — your live traffic is sent to the scanner and a verdict comes back in under a second. Eval-time — you test your model against a suite of attacks at CI time, not at request time. Per-call gating — the scanner is bundled inside a larger platform and billed per invocation. These three shapes are not substitutes: Promptfoo is eval-time and extremely good at it, but it does not intercept live user uploads.

The second gotcha is text-in-image vs pixel-level injection. A scanner that runs OCR on images and then passes the OCR output through a text PI pipeline covers a third of the threat surface — it catches unsophisticated FigStep variants but misses anti-OCR rendering, low-res composites, and paraphrased list structures. See FigStep detection for why.

When each option is the right choice

How Glyphward fits

Glyphward is purpose-built for the multimodal gap every text-first defender leaves open: one endpoint that scans images and audio at inference time, free for 10 scans a day, $29 a month for 100,000. No Azure tenancy required, no enterprise sales cycle. Drop-in REST plus Node and Python SDKs.

Get early access · See plan cards

Related questions

Is Glyphward a replacement for Lakera Guard or LLM Guard?

Not a 1:1 replacement — a complement for most teams. If you already run a text PI scanner you are happy with, bolt Glyphward on to cover images and audio and nothing else changes. If you have no scanner and most of your attack surface is image/audio uploads, Glyphward alone is a sensible v1.

Why is Azure Prompt Shields listed as text-only when Microsoft markets content safety broadly?

Content Safety and Prompt Shields are separate products inside Azure AI Foundry. Content Safety covers image moderation (sexual, violent, hateful content) but not prompt injection. Prompt Shields inspects text. Neither scans image-rendered instructions or audio at the waveform.

You listed Promptfoo as "not a real-time scanner" — is that fair?

It is — and it is not a criticism. Promptfoo is excellent at what it does: eval-time red-team runs and CI-gated attack-class reports. Its own docs describe the product as an eval harness. Real-time inference scanning is a different surface, and the team has not marketed Promptfoo as that.

Further reading