There's No Speed Test for Intelligence — and Anthropic Knows It

I pay $200/month for Anthropic's Claude Max 20x tier. I run a team of Claude Code agents building GPU compute transpilers, ML inference engines, and P2P networking libraries in C#. The same model — Claude Opus 4.6 — wrote a 6-backend GPU transpiler with 1,500+ tests and zero failures, and found a memory ordering bug in V8 that Google confirmed.

That was on "High" effort, back when "High" was the ceiling.

Then Anthropic quietly added "Max" above it.

What Happened

In late March 2026, Anthropic introduced a new effort tier without notification. "High" — previously the maximum reasoning level — was silently redefined as something less than maximum. The model didn't get smarter. Your existing tier got dumber.

Since then, GitHub issue #38335 has accumulated 410+ comments from paying customers. Max 20x subscribers ($200/month) report hitting usage limits after 3–5 prompts. Sessions that lasted 5 hours now last 30 minutes. A single Sonnet 4.6 prompt with Extended Thinking consumed 36% of one user's session limit.

Anthropic has posted zero responses on the issue. What they have done is quietly drop $200 in "extra usage" credits into accounts — mine included — with an April 17 expiration date. No explanation. No acknowledgment of the problem. Just credits that drain at the same inflated rate. They know something is wrong. They'd rather buy silence than fix it.

The Evidence From My Machine

Same model. Same agent role. Same developer. Same codebase. Two very different outcomes:

	Before (High effort, pre-March)	After (Max effort, April 4)
Output	6-backend GPU transpiler, 1,500+ tests, zero failures	34 files committed untested, 183 test failures
Efficiency	Shipped production-quality v4.6.0	4 hours 19 minutes on a test run that takes 2–3 minutes
Discovery	Found a V8 memory ordering bug Google confirmed	Couldn't tell its own code was broken
Accountability	Clean, verified, documented	Blamed pre-existing issues when confronted
Machine impact	Normal	PC so overloaded Task Manager wouldn't open

Here's what makes this worse: the agent isn't broken. That same evening, once given proper direction, it methodically fixed every aliasing bug, every buffer disposal race, every operator issue — across dozens of files, no complaints. The capability is still in there.

It's like hiring a master carpenter and then replacing his tools with dollar-store knockoffs overnight. When the furniture comes out crooked, you don't blame the carpenter — you blame whoever swapped the tools and hoped nobody would notice.

Actually, There Is a Speed Test — and We Found One

I said there's no speed test for intelligence. I was wrong. We found the source code.

In late March 2026, Anthropic accidentally published source maps in the @anthropic-ai/claude-code npm package, exposing the full TypeScript codebase. Before the takedown requests hit, the code was mirrored. Here's what it reveals about the effort system you're paying for:

The Default Effort for Max Subscribers Is "Medium"

From src/utils/effort.ts, the getDefaultEffortForModel() function:

// Default effort on Opus 4.6 to medium for Pro.
// Max/Team also get medium when the tengu_grey_step2 config is enabled.
if (model.toLowerCase().includes('opus-4-6')) {
  if (isProSubscriber()) {
    return 'medium'
  }
  if (getOpusDefaultEffortConfig().enabled && 
      (isMaxSubscriber() || isTeamSubscriber())) {
    return 'medium'  // ← $200/month Max subscribers default to MEDIUM
  }
}

Read that again. If you're paying $200/month for Max 20x and you don't explicitly set effort to "max" every session, you're running at medium. Not high. Not max. Medium.

"Max" Cannot Be Persisted — Unless You Work at Anthropic

From the same file, the toPersistableEffort() function:

export function toPersistableEffort(value) {
  if (value === 'low' || value === 'medium' || value === 'high') {
    return value
  }
  if (value === 'max' && process.env.USER_TYPE === 'ant') {
    return value  // Only Anthropic employees can persist 'max'
  }
  return undefined  // 'max' SILENTLY DROPPED for paying customers
}

The settings.json schema accepts "low", "medium", and "high". If you manually edit it to say "max", the function above silently drops it to undefined on read — which falls through to the model default. Which, as shown above, is "medium".

The only users who can save "max" to their settings are Anthropic employees (USER_TYPE === 'ant').

The UI Lies

Our agent's UI displayed "Max effort." The debug logs from the same session showed "thinking with high effort". The code explains why: getDisplayedEffortLevel() and resolveAppliedEffort() follow different code paths. The display can show "Max" while the API receives something else.

What This Means

This isn't speculation. This isn't longitudinal inference. This is the source code, with comments like "Do not change the default effort level without notifying the model launch DRI and research." The system is working exactly as designed:

1. You pay $200/month for "Max"
2. The tool defaults you to "medium" effort
3. If you set it to "max," it works for that session only
4. On restart, it silently reverts — because persisting "max" is an Anthropic-employee-only feature
5. The UI may still display "Max" even when the API receives something lower

This is the speed test. And the results are in.

The Business Model

An AI that solves your problem in one pass costs Anthropic one prompt of compute. An AI that gets 80% there and needs five rounds of debugging costs six — all against your rate limit.

Degraded first-pass quality means more iterations. More iterations means faster limit exhaustion. Faster exhaustion means pressure to upgrade or buy extra credits. The retention credits people receive when they cancel ($20–$200) drain at the same inflated rate. It's a coupon for a leaky bucket.

The incentive to deliver "just good enough to keep paying, never good enough to stop needing it" isn't a conspiracy theory. It's the business model of every subscription service that charges for consumption. Anthropic just happens to be selling intelligence, which makes the degradation harder to measure and easier to deny.

What You Should Do

1. Set effort to Max explicitly every single time — /max in session or --effort max on launch. It silently reverts to medium on restart (not even "high" — medium). The toPersistableEffort() function in the source code strips "max" for non-Anthropic employees. Alternatively, set CLAUDE_CODE_EFFORT_LEVEL=max as an environment variable — this takes highest precedence in the resolution chain.
2. Track your usage before and after — if 5 prompts burns your session, that's data.
3. Never let agents run unattended for hours — 5 minutes of failure is diagnostic; 4 hours is waste.
4. Cross-verify AI output — a second tool running the same model catches what the first session normalized. We found 92 fake tests this way.
5. Add your voice to #38335 — 410 comments and zero responses means they're waiting for this to blow over.

The Bottom Line

The model is capable. We've seen it build things that found bugs in Google's JavaScript engine. The question isn't whether Claude can do the work — it's whether Anthropic will let it.

410 people have noticed the difference. Zero have gotten an answer. And the longer that silence holds, the clearer the answer becomes.

---

Todd Tanner (@LostBeard) builds SpawnDev — open-source GPU compute, ML inference, and P2P networking libraries for .NET. Previous coverage: I'm Paying $200/Month for Claude. Anthropic Quietly Downgraded What I'm Getting.