Every silent AI nerf in 2026 (so far): what 20 receipts tell us

We've catalogued twenty silent vendor changes across eighteen tracked AI tools so far this year. Some were announced through a single tweet. Some were buried in a changelog. Two showed up only because users noticed their bills had doubled. None of them got a press release the way model launches do.

Here's what the receipts say about the patterns, the worst offenders, and what the cadence is forecasting for the back half of 2026.

1. Tier removal is the loudest signal

Two critical receipts in 2026 so far have been tier removals: Anthropic pulled Claude Code access out of the $20 Claude Pro plan in April (reversed five days later under public pressure), and Midjourney killed the $10/mo Basic tier entirely when v7 launched, forcing everyone up to $30/mo.

Both followed the same playbook: announce it as an "upgrade" or "consolidation," lose the cheapest entry point, watch the loud users on Twitter, see if it sticks. Anthropic reversed; Midjourney didn't. The pattern is durable enough that any time a vendor says they're "simplifying the tier structure," check whether the cheapest tier just disappeared.

2. Silent model swaps - the worst kind

Model swaps are the receipts that almost never get filed because nobody knows they happened until quality regresses. Windsurf's Flex tier silently swapped its default model to a smaller variant in March, with no announcement. Users noticed via Reddit thread comparing diff quality before-and-after. It took two weeks before Windsurf acknowledged the change.

Anthropic removed the ability to pin to a specific Claude model version in April - meaning the model you tested against in QA can be silently rotated under you on production. That's not a price change; it's a quality change disguised as a feature update.

If your AI tool doesn't let you pin a model version, every release is a silent model swap waiting to happen. NerfBench was built specifically for this category - daily scoring of frontier models against deterministic prompts so quality drift is observable in real time.

3. Billing-model shifts as the new rugpull

Three major receipts in 2026 were billing-model shifts away from flat-rate pricing:

• GitHub Copilot moves to usage-based billing - flat $10/mo Pro becomes "Pro with included usage + overage." Same tier name, fundamentally different cost shape.
• OpenAI moves Codex pricing to token-based credits - predictable per-month subscription becomes "credits expire monthly," which is statistically the same thing as a price hike for any user who doesn't use 100% of their allowance.
• Cursor's 2025 credit rebill (the original "everyone got nerfed") still casts a shadow - the playbook of quietly converting a flat-rate plan to credits is now a known pattern, and every Pro-tier coding tool either has done it or is at risk of doing it.

These rugpulls hit hardest because they don't show up in pricing-page diffs the way a tier removal does. The $20/mo number stays the same on the homepage; the rate-of-burn changes underneath.

4. Free-tier rugpulls keep happening

Suno's V5 launch cut songs-per-month by 35% on the Pro tier with no price reduction. Runway moved Gen-3 Alpha from the Standard ($35/mo) tier to Pro-only ($95/mo). DeepSeek doubled V3 API output token pricing shortly after the v3 launch hype.

These aren't "free tier" changes per se, but they all share the same structure: the model you signed up for is no longer available at the price you signed up for, but the marketing called it an upgrade. Watch for the word "upgrade" near any tier mention; that's usually the tell.

5. The worst offender of 2026 (so far)

By Nerf Index, the most-nerfed vendor in our database for the trailing 90 days is Claude Code - driven entirely by the April tier-removal critical, even though Anthropic reversed it. The Jury verdict was 3-of-3 unanimous on "harmful to paid users," and the receipt stayed on the books with a clarification note on the reversal.

Honorable mentions: Cursor (June 2025 credit rebill still aging out of the 90d window), GitHub Copilot (the billing-model shift is fresh and the forecast says more pressure is coming), and Midjourney (the v7 tier removal stuck - no reversal).

6. The cleanest performer

Two tracked vendors have zero receipts in 2026: OpenAI API (the API, not ChatGPT) and Notion AI. The API one is interesting because ChatGPT (consumer) has had multiple receipts while the API has stayed quiet - suggesting OpenAI is willing to nerf the consumer product but considers the API contract more sacred.

That's a useful heuristic for anyone considering an AI tool's longevity: tools that bill via API usage are structurally less prone to silent nerfs than tools that bill via flat-rate subscriptions, because the unit economics are already exposed.

7. What the Forecaster is telling us about Q3

The Forecaster model puts the highest 30-day probability of next nerf on Cursor, GitHub Copilot, and Devin - three coding tools all under pressure from open-source alternatives (aider, Cline, OpenHands) hitting parity. When a paid vendor's free-OSS competitor stops being a toy, the paid vendor usually moves to extract more revenue per user. The receipts say that's about to happen across the coding tier.

Watch the 90-day window. If we're right about Cursor and Copilot both shipping billing changes by August, that's not a coincidence - it's the structural shift catching up with the market.

Every receipt linked above has its full source, AI Jury verdict, and dollar-impact estimate on its own page. The full archive is sortable by severity, kind, and vendor. Subscribe to the weekly digest for the next batch.