Stealth Pricing: When the List Price Lies
The list price is increasingly disconnected from the actual cost. Six mechanisms documented in the dataset — none of them show up on a pricing page.
FairMind published "List Price Is Lying" on May 5, 2026, documenting how three AI vendors changed effective costs through hidden mechanisms while headline prices stayed flat. Our dataset confirms the pattern extends well beyond AI infra.
The Mechanisms
| Mechanism | Company | What Happened | Effective Cost Impact | Source |
|---|---|---|---|---|
| Tokenizer inflation | anthropic | Opus 4.7 tokenizer produces 32-45% more tokens for same text | +12-27% for prompts >2K tokens | FairMind |
| Model-generational pricing | openai | GPT-5.5 at $5/$30 per 1M tokens vs GPT-5.4 at $2.50/$15 | +100% list price, +49-92% real cost | FairMind |
| Bundle-justified hikes | microsoft | M365 price increases (+5-33%) bundled with new security features (Defender, Intune). Frontline SKUs hit hardest: F1 +33%, F3 +25%. Enterprise E3 +8%, E5 +5%. | +5-33% depending on SKU | Microsoft |
| Feature unbundling | gong | Forecast, Engage, Enable, Data Cloud split into paid add-ons | +25-56% effective cost | Sybill |
| Silent policy change | zendesk | Auto-bills overage resolutions without warning or discounted rate (since Jan 2026) | Unpredictable overage exposure | CorePiper |
| Access restriction | notion | AI add-on ($8/member/mo) discontinued for new Free/Plus subscribers (May 2025). Existing add-on users grandfathered. Basic AI writing features remain on Plus; full AI (Agents, Ask Notion) requires Business+. | Forced tier upgrade for full AI access | AI Productivity |
Why This Is Different from the Fee Retreat
The fee-retreat documents explicit capacity cuts — Relay slashing credits 60%, Replit trimming allocations 20%. Those are visible on the pricing page. The customer knows they're getting less.
Stealth pricing is the opposite: the pricing page looks the same (or better), but the effective cost goes up. Three distinct patterns:
| Pattern | How It Works | Example |
|---|---|---|
| Same price, more units consumed | The unit of measurement changes so more units are needed for the same work | Anthropic tokenizer: same prompt, 32-45% more tokens billed |
| Same price, less included | Features or capacity removed from existing tiers without price adjustment | Notion discontinuing AI add-on for new Free/Plus subscribers; Gong unbundling modules |
| Higher price, bundled justification | Price increases packaged with "new features" that most users didn't ask for | Microsoft M365 +5-33% hikes bundled with Defender and Intune additions |
The Scale
Vertice SaaS Inflation Index (2026):
| Metric | Value |
|---|---|
| SaaS inflation rate | 12.2% overall; 13.2% in March 2026 |
| SaaS costs per employee | $9,100 (end 2025), up from $7,900 (2023) |
| Renewals with reduced value, no price decrease ("shrinkflation") | 28% of Q4 2025 renewals |
| Vendors deliberately masking rising prices | 60% |
28% of renewals delivering less value at the same price. 60% of vendors deliberately masking increases. This isn't anecdotal — it's the majority behavior.
AI Inference: The New Hidden Variable
AI pricing introduces a new stealth vector: the cost of the underlying model changes independently of the product price.
| Layer | Who Controls It | Who Pays | Visibility |
|---|---|---|---|
| Product price | Vendor | Customer | High — on pricing page |
| Model cost (tokens) | Model provider | Vendor (passed through or absorbed) | Low — buried in API docs |
| Tokenizer efficiency | Model provider | Everyone downstream | None — not disclosed on any pricing page |
| Prompt caching | Vendor implementation | Shared savings | Varies — some vendors pass through, some absorb |
When Anthropic changes its tokenizer, every company building on Claude sees cost changes that their own customers never see on a pricing page. The cost stack has more layers than the pricing page reveals.
What to Watch
Tokenizer drift will become a recurring pattern. Every major model update could change token counts for identical prompts. Companies building on LLM APIs need to monitor tokens-per-prompt as a cost metric, not just price-per-token.
Bundle-justified hikes will accelerate. Microsoft proved the playbook: add security features nobody requested, raise prices 5-33%, frame it as "enhanced value." Expect Salesforce, ServiceNow, and Adobe to follow.
Feature unbundling is the inverse of AI bundling. While most companies are bundling AI into existing plans (credit-convergence), some are simultaneously unbundling non-AI features into paid add-ons. Gong is the clearest example — the net effect is higher TCO even as "AI is included."
See also: fee-retreat, credit-convergence, incumbents-leading-ai-pricing