Claude Code’s Thinking Depth Dropped 73% — Anthropic Admits It Quietly Dialed Down Default Reasoning, API Users Restored on April 7

Between February and March, Anthropic silently downgraded Claude’s default thinking intensity from high to medium, resulting in a 73% reduction in reasoning depth. This post breaks down what happened, why it happened, and how API users can verify they’ve been restored to full-strength thinking mode after April 7.

What Happened

Between February and March 2026, Anthropic quietly made three undisclosed configuration changes to Claude:

• Feb 9— Claude switched to adaptive reasoning depth, letting the model decide how much to think per task rather than using a fixed budget

• Feb 12 — Intermediate thinking steps were hidden in the UI to reduce perceived latency

• Mar 3 — Default effort level was silently downgraded from high to medium (level 85) None of these changes were announced prominently. Developers started noticing Claude Code felt increasingly sluggish and unreliable — but with no changelog to reference, most could only post on GitHub, Reddit, and X asking “Is it something I’m doing wrong?”

!Anthropic’s 3 undisclosed configuration changes to Claude

The Numbers: How Bad Was the Drop?

What turned this into a major story was a systematic analysis by Stella Laurenzo, Senior Principal Engineer at AMD AI — she ran an engineering-grade diagnostic across 6,852 sessions, 17,871 thinking blocks, and 234,760 tool calls:

CLAUDE CODE · Performance Comparison

Claude Code Reasoning Depth Down 73%: January vs. March

Data source: AMD AI Senior Principal Engineer Stella Laurenzo — 6,852 sessions · 17,871 thinking blocks · 234,760 tool calls

The drop in files read before editing — from 6.6 down to 2.0 — is particularly telling. Claude Code was editing code without properly understanding the dependency graph first, leading directly to more errors and rework cycles.

Why Did Anthropic Do This?

Boris Cherny, the product lead for Claude Code, offered a public explanation:

“We lowered the default effort to medium because we were getting user feedback that Claude was consuming too many tokens.”

In other words: to save tokens and reduce latency, reasoning quality was sacrificed — without giving users a choice, and without a clear public announcement.

A more controversial theory also circulated — the “compute squeeze” hypothesis. Around the same time, Anthropic was internally testing its Mythos model, and some speculated that compute resources were being prioritized for frontier model development. Anthropic never directly addressed this.

Where Things Stand Now

On April 7, Anthropic pushed an official changelog update:

The default effort level has been restored to high for API Key users, Bedrock, Vertex, Foundry, Team, and Enterprise users.

For developers calling Claude via API Key, reasoning strength is back to January levels.

Who’s still affected: Free and Pro subscribers on claude.ai remain on medium effort. Session rate limits have also been tightened — roughly 7% of users will hit their caps faster than before.

What This Means for ClaudeAPI Users

For developers accessing Claude through ClaudeAPI using an API Key:

• Feb 9 – Apr 7— Requests ran at medium effort; complex task performance may have been noticeably weaker
• After Apr 7 — Automatically restored to high effort; no action required on your end If you want to explicitly control reasoning depth, you can pass the thinking parameter directly in your API request:

response = client.messages.create(
    model="claude-opus-4-7-20260417",
    max_tokens=16000,
    thinking={
        "type": "enabled",
        "budget_tokens": 10000  # explicitly set thinking token budget
    },
    messages=[{"role": "user", "content": "Your task here"}]
)

response = client.messages.create(
    model="claude-opus-4-7-20260417",
    max_tokens=16000,
    thinking={
        "type": "enabled",
        "budget_tokens": 10000  # explicitly set thinking token budget
    },
    messages=[{"role": "user", "content": "Your task here"}]
)

Passing this parameter explicitly overrides the default effort setting, locking in the reasoning depth you actually need. For more usage details, check out the Claude API Python Tutorial，or visit the Claude API error Guide to diagnose unexpected behavior.

What to Watch Next

• Whether effort is restored for Free/Pro chat users — Anthropic has not published a roadmap for this

• Mythos model API availability — Currently limited to 12 partner organizations

• Updated session rate limit rules — Heavy users should monitor how frequently they’re hitting caps

FAQ

Q: Did Claude actually get worse? Is there data to back this up?

A: Yes, and the data is solid. On March 3, Anthropic lowered the default effort level from high to medium, resulting in roughly a 73% drop in reasoning depth.

Q: I’m using an API Key — am I still affected?

A: No. As of April 7, Anthropic has restored API Key users to high effort by default.

Q: How do I force a specific reasoning depth?

A: Pass the thinking parameter in your API request and set budget_tokens. This overrides the default and locks in your desired reasoning intensity.

Q: When will Free/Pro chat users be restored?

A: No timeline has been announced. Currently, only API Key and enterprise users have been restored.

Q: Will this incident affect ClaudeAPI’s reliability?

A: No. ClaudeAPI proxies the official Anthropic API directly — the effort level restoration applies uniformly to all API Key users with no extra steps needed on your end.

Sources

• Fortune — Anthropic faces user backlash
• VentureBeat — Is Anthropic ‘nerfing’ Claude?
• Axios — Anthropic’s AI downgrade stings power users
• Anthropic Claude Code Release Notes

Published by theClaudeAPI team.