How Anthropic Nerfed Opus 4.6 Before the 4.7 Launch

Straight to the meat:

Anthropic shipped Opus 4.6 on February 5, 2026 [4]. Four days later, adaptive thinking became the default [1] [7]. The model now chose its own reasoning budget per turn.

On March 3, the default effort dropped from high to medium. Boris Cherny, the Claude Code lead [6], said the change balanced intelligence, latency, and cost [1] [6].

A dialog showed up when you opened Claude Code [1]. Most people clicked through [9].

Between March 5 and March 12, a beta header called redact-thinking-2026-02-12 rolled out [1]. By March 8, 58% of responses had their reasoning hidden from the UI [1]. By March 12, effectively all of them [1].

None of these were weight changes [1]. All three were product decisions [1].

Stella Laurenzo, Senior Director of AI at AMD [5], noticed her Claude Code sessions getting worse [1] [5]. She exported 6,852 of them and ran the numbers [1] [5].

Median visible thinking per turn dropped 73% [1] [5]. In late January it was about 2,200 characters [1]. By mid-March it was 600 [1].

The read-to-edit ratio fell from 6.6 to 2.0 [1]. The model stopped looking at files before changing them [1].

The user interruption rate per 1,000 tool calls rose twelvefold, from 0.9 to 11.4 [1].

Early stopping hooks fired close to zero times per day before March 8, and about 10 per day after [1].

Edits to files the model had not recently read climbed from 6.2% to 33.7% [1]. That is more than five times worse.

Laurenzo posted the analysis as GitHub issue #42796 on April 2 [1] [5].

On April 6, Cherny responded [1] [6]. He confirmed the adaptive thinking default on February 9 [1].

He confirmed the effort=85 default on March 3 [1]. He said the redact-thinking header was a UI-only change that did not affect thinking budgets [1].

Then he said this:

The specific turns where it fabricated had zero reasoning emitted, while the turns with deep reasoning were correct. We’re investigating with the model team. [1]

Adaptive thinking sometimes allocates zero reasoning to a turn [1] [7]. When that happens, the model fabricates [1]. Cherny’s interim workaround was CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 [1] [7].

Ten days later, on April 16, Anthropic released Opus 4.7 [3] [8]. The default effort on 4.7 was a new level called xhigh [3] [8].

Higher than the 4.6 default of medium [1] [3]. Higher than the 4.6 launch default of high [1] [4].

Users who upgraded felt the improvement [8]. Some of it was new weights [3]. Some of it was the effort default moving back up the curve [3] [8].

You cannot separate the two from outside the company [11].

In the same weeks, a second community story went viral [2]. BridgeMind posted a chart claiming Opus 4.6 had fallen from rank #2 to rank #10 on its benchmark, 83.3% down to 68.3% [2].

An independent AI researcher checked [2]. Yesterday’s score was from 30 tasks [2]. Last week’s was from 6 [2].

On the six common tasks the score was 87.6% then and 85.4% now [2]. The chart was comparing different test sets [2].

That piece did not hold up [2]. The Laurenzo analysis did [1] [5].

The version string did not change between February and April [3] [4]. The price did not change [3] [4]. What you got changed [1] [11].

I am not calling this deliberate planned obsolescence. The evidence does not establish intent. The structure produces the effect whether anyone plans it or not [10].

Three defaults changes ship quietly [1]. A UI redaction removes the main diagnostic signal [1] [11]. The new version launches with the defaults moved back up [3] [8].

Each decision defends on its own [1] [6]. The aggregate is a 70-day downgrade at a constant price [3] [4], followed by a release that feels like restoration [3] [8].

The version string is not the product. In an opaque platform, the defaults are.