On April 7, Anthropic’s Claude Code changelog quietly recorded a single line: “Changed default effort level from medium to high for API-key, Bedrock/Vertex/Foundry, Team, and Enterprise users.” No blog post. No announcement. No apology. Just a version bump (v2.1.94) reverting what had been done five weeks earlier.

What happened between March 3 and April 7 is a case study in how AI companies behave when compute economics collide with user trust. And given that Anthropic’s ARR just surpassed $30 billion (up from $9 billion at year-end 2025). At this scale, every compute decision is an economics decision.

Quick Summary: What Happened

Detail	Info
March 3, 2026	Anthropic silently lowered Claude’s default effort from “high” to “medium”
Early April	GitHub quality complaints spiked 3.5x vs. baseline; major coverage followed
April 2	AMD senior director files forensic GitHub issue; engineers publicly call out the decline
April 7	Anthropic reverses the change in v2.1.94 — no public explanation
Who Was Affected	Enterprise and API users, agent workflows, complex engineering tasks
Official Source	Claude Code CHANGELOG v2.1.94

Bottom line: Anthropic throttled Claude’s performance for five weeks without disclosure, got caught by power users, and reversed course quietly. The incident exposed real tensions between its scaling economics and the pricing model that enterprise customers built their workflows on.

What Actually Happened

On March 3, Anthropic changed Claude’s default “effort” level (a parameter controlling how many tokens the model spends on reasoning before responding) from “high” to “medium.” Claude Code head Boris Cherny described it at the time as “the best balance across intelligence, latency and cost for most users,” according to VentureBeat.

Users weren’t told. No changelog entry at the time, no email, no release note flagged as customer-impacting. For individual users prompting casually, the change may have been invisible. For professionals running automated pipelines, agent workflows, or complex engineering tasks — it wasn’t.

The Numbers Behind the Complaints

The Register asked Claude itself to analyze its GitHub issue tracker filtered for quality complaints since January 2026. The results were stark: April had already accumulated 20+ quality issues in just 13 days, on pace to exceed all of March’s 18. March was itself already 3.5x the January–February baseline.

Something changed in March. The timing aligned exactly with the effort reduction.

Stella Laurenzo’s Forensic Audit

On April 2, Stella Laurenzo — senior director of AI at AMD — filed a GitHub issue titled “[MODEL] Claude Code is unusable for complex engineering tasks.” It wasn’t a frustrated complaint. It was a forensic analysis.

Laurenzo had analyzed 6,852 Claude Code sessions, 17,871 thinking blocks, and 234,760 tool calls. The findings:

• Read:Edit ratio collapsed from 6.6 to 2.0: a 70% drop in research before making code edits
• User interrupts per 1,000 calls jumped 6.5x: the model was going off-rails far more often
• Estimated costs increased 122x month-over-month — as users fought degraded output by retrying

She then went public. In a statement covered by PC Gamer and The Register, Laurenzo declared: “Claude has regressed to the point it cannot be trusted to perform complex engineering.”

That’s not a power-user tweet. That’s a technical director at a Tier 1 AI hardware company saying the product broke her team’s workflows. Senior engineers at AMD and Microsoft made similar observations publicly in the same window. Claude had gone from their preferred coding assistant to something that required constant supervision.

Why This Matters

The Real Pressure: What $30B ARR Actually Costs

Anthropic’s ARR jump from $9 billion to $30 billion in roughly four months is extraordinary. Bloomberg reported it alongside news of a new Broadcom chip deal. But that revenue growth comes with a brutal compute bill.

Here’s the math that explains the throttle decision: Anthropic’s Max tier at $200/month had become structurally indefensible. Heavy users — particularly enterprises running agent loops — were consuming compute worth $1,000–$5,000 per month at standard API rates. At a $200 flat fee, every power user was subsidized by lighter accounts.

The effort reduction was, in effect, a band-aid on that pricing model. Less effort per request = fewer output tokens = lower compute cost per seat. The problem: the users who noticed the degradation fastest were exactly the power users generating the most compute load. The fix hurt the people it was designed to protect the business from.

Pricing Is Being Rebuilt From Scratch

Gizmodo reported on April 15 that Anthropic is shifting enterprise pricing to a usage-based model: $20/seat/month base fee, with all compute billed at standard API rates. The old flat-fee structure — up to $200/user/month — is being retired.

This is the right structural move. The flat subscription model that made AI mainstream was always a temporary state; it works when AI is a novelty for casual users and breaks when enterprises are running a quarter-million tool calls per month. But the timing couldn’t be worse. Announcing a pricing overhaul days after a hidden performance cut — even one that was reversed — creates a narrative problem that one clean rollout would have avoided.

For a broader look at how AI vendor pricing is shifting, see our AI pricing comparison for 2026.

The Transparency Problem

Anthropic’s brand rests heavily on being the “safe and honest” AI company — the principled alternative to OpenAI’s move-fast culture. Our Anthropic vs OpenAI analysis covers how that positioning has defined the company’s enterprise pitch.

This incident challenged that brand directly. The degradation wasn’t announced. The reversal wasn’t announced. Both were discovered by users doing forensic analysis of their own session data and filing GitHub issues with quantitative evidence.

Fortune’s coverage made the stakes explicit: the situation threatens Anthropic’s IPO trajectory by undermining its transparency positioning. When your go-to-market story is “we’re the honest AI company,” silently shipping a performance reduction and reverting it after public backlash is a particularly damaging sequence of events.

The contrast with a different approach is worth sitting with. “We’ve updated Claude’s effort defaults to balance speed and cost — override anytime with /effort high” would have been a three-line release note. Most users would have accepted it. A few enterprise customers might have pushed back. Instead, the silence turned a reasonable product decision into an integrity story.

What You Should Do Now

If you’re running Claude-dependent workflows, three things are now relevant:

1. Override the default explicitly. The /effort flag is real and documented. For complex engineering tasks, agent workflows, or anything where thoroughness matters, specify /effort high rather than trusting the default to stay where it is. Defaults can change again.

2. Benchmark your own workflows. Laurenzo’s approach — tracking Read:Edit ratios, user interrupts, and retry rates — is the right framework for any agentic Claude usage. If you don’t have baseline metrics, you have no early warning system for future changes.

3. Model the new pricing before it hits. The shift to $20/seat + API usage rates will significantly change costs for heavy users. Run the math before your next renewal. Our AI cost optimization guide has a framework for evaluating AI spend.

For teams evaluating alternatives, our best AI coding assistants guide covers how Claude compares to Cursor, Copilot, and others specifically on the complex engineering tasks at issue here.

The Bigger Picture

The flat subscription model that got AI into every enterprise — $20/month, use it as much as you want — was always running on borrowed time. It worked when usage was conversational. It breaks when you’re running 234,000 tool calls a month.

What Anthropic is navigating now is the compressed version of what every AI provider will face: the power users who drive your strongest retention are the ones making your unit economics impossible. The Max-tier users running agent loops at $200/month were simultaneously Anthropic’s most sophisticated customers and its most costly accounts. The effort reduction was an attempt to fix that without a pricing change. The backlash showed it couldn’t.

The usage-based shift is the correct destination. But execution matters enormously. Moving to usage-based billing with advance notice and a transition period is manageable. Quietly throttling, getting caught, reversing course, and then announcing a price restructure days later is a different experience for customers.

GitHub Copilot moved to usage-based billing last year with months of notice and clear migration documentation. How Anthropic handles the next few months of communications will determine whether this incident is remembered as a blip or a turning point.

What It Means for Claude’s Enterprise Position

Claude’s capabilities on complex tasks remain genuinely strong — the underlying model wasn’t degraded during the March incident, only the effort it was applying by default. The April 7 reversal restored the behavior customers had built workflows around. But the “Claude is the reliable one” narrative took a real hit.

Laurenzo’s GitHub issue — a forensic audit from a technical director at a Tier 1 hardware partner — is the kind of evidence that circulates in engineering Slack channels and architecture review meetings. The reversal helps. It doesn’t erase the fact that the change went undisclosed for five weeks.

Our Take

The throttle decision itself was defensible. AI companies cannot run infinitely scalable compute at fixed prices, and adjusting effort defaults to manage costs isn’t inherently wrong. The transparency failure is what turned this into a story.

For users relying on Claude for professional work, the practical lesson is clear: treat AI performance as something that can change without warning, build verification into your workflows, and stay close enough to your usage metrics to notice shifts before they cost you real work. Don’t assume what worked last month is what you’re getting this month.

For Anthropic, the harder lesson is about the asset they’ve spent years building. “Helpful, harmless, and honest” is worth protecting. Silence during product changes is an integrity expense that doesn’t show up on a P&L — until something like this happens, and then it shows up everywhere.

The ARR growth is real. The compute pressure is real. The path to sustainable unit economics is roughly what they’ve announced. But users who built their workflows on “the AI you can trust” are watching more carefully now.

Frequently Asked Questions

What is Claude’s “effort” setting?

Effort is a parameter that controls how many tokens Claude dedicates to its reasoning process before generating a response. Higher effort produces deeper analysis and better performance on complex tasks — at higher compute cost. Anthropic set the default to “high” at launch, reduced it to “medium” on March 3 to cut per-request costs, and restored it to “high” on April 7 following user backlash. You can control it manually with /effort high, /effort medium, or /effort low.

How do I set Claude’s effort level manually?

Use the /effort command in Claude Code or pass the effort parameter via the API. /effort high restores maximum performance. /effort medium is the previous default. /effort low is optimized for speed on simple tasks. If you’re running complex agentic workflows, explicitly setting /effort high insulates you from future default changes.

Did Anthropic acknowledge the change publicly?

Not proactively. The reversal appears in the Claude Code CHANGELOG as a one-line note in v2.1.94. Claude Code head Boris Cherny commented on the original reduction in response to user questions, but Anthropic published no formal explanation of the March 3 change, the degradation complaints, or the April 7 reversal.

Is Claude still reliable for complex engineering tasks?

With effort restored to “high,” yes — the underlying model capability was never degraded. The March issue was specifically the reduced default, which meant Claude was applying less reasoning capacity per request than users expected. For current benchmarks, our Claude Opus 4.6 review covers where the model stands today.

How does the new usage-based pricing affect enterprise customers?

The shift from flat fees (up to $200/user/month) to a $20/seat base plus standard API usage rates will significantly change costs for heavy users. Customers running agent loops or high-volume workflows may see 3–10x cost increases. Run the math before renewal — specifically, pull your token consumption from the last 30 days and multiply by current API rates.

Why does this matter for Anthropic’s IPO?

Anthropic’s enterprise valuation depends partly on its reputation for reliability and transparency — two differentiators from OpenAI that show up in every enterprise sales pitch. An incident combining undisclosed performance degradation with a simultaneous pricing restructure creates a credibility narrative. Fortune’s coverage directly connected the two.

---

Last updated: April 16, 2026. Sources: Fortune, The Register, PC Gamer, GitHub issue #42796, VentureBeat, Gizmodo, Bloomberg.

Related reading: Anthropic vs OpenAI 2026 · Claude Review 2026 · AI Pricing Comparison 2026