[BILLING] Medium effort default (effort=85) causing excessive token burn and unexpected extra usage costs

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Unexpected token consumption and cost impact from default effort=85

Context

Related to #42796 and @bcherny's response in https://github.com/anthropics/claude-code/issues/42796#issuecomment-4194007103 which confirmed:

1. Adaptive thinking became default Feb 9 (Opus 4.6)
2. Medium effort (85) became default Mar 3

This issue is specifically about the cost/quota impact of these changes on heavy users.

The Problem

I'm a Max 20x subscriber using Claude Code daily for complex engineering tasks (Playwright browser automation, accounting engine with 21 MCP tools, multi-file Python/TypeScript projects, Google Sheets API integrations).

Before the Feb/Mar changes: Weekly limits were sufficient. Extra usage was minimal.

After the changes:

• Weekly "All models" limit hits 100% consistently within days
• Extra usage: $361.37 spent in less than 2 days (of $400 monthly cap, 90% burned)
• Sonnet-only limit: only 6% used — suggesting almost all spend is on Opus
• Current balance: $34.20 remaining with 3+ weeks left in the billing cycle

Why Medium Effort Causes Higher Total Cost

The irony of effort=85 is that while each individual turn uses fewer tokens, the total session cost goes up because:

1. Shallower thinking → more mistakes — Claude makes errors that require correction loops
2. More retry turns — each correction is another Opus call
3. Compounding context — failed attempts accumulate in context, making subsequent turns more expensive (larger input)
4. User frustration → restart sessions — when corrections spiral, users start fresh conversations, losing all prior context investment

For complex engineering tasks (multi-step Playwright automation, form-filling across paginated wizards, Wicket AJAX event dispatching), medium effort regularly produces incorrect solutions that require 2-3 correction cycles. At high effort, these typically succeed on the first attempt.

Net effect: effort=85 saves tokens per turn but burns 2-3x more tokens per task.

Evidence from My Usage

| Metric | Value |
|--------|-------|
| Plan | Max 20x |
| Weekly limit | 100% used (all models) |
| Extra usage | $361.37 / $400 (90%) — burned in under 2 days |
| Sonnet usage | 6% (barely touched) |
| Primary workload | Claude Code (Opus) — Playwright, Python, MCP tools |
| Time to hit weekly cap | ~2 days of active use |
| Billing cycle remaining | 3+ weeks (resets May 1) |

The Second-Account Workaround Tells the Story

When I hit both the weekly cap AND the extra usage ceiling within 2 days, I contacted Anthropic support asking whether creating a second Max subscription under a different email was compliant. Support confirmed it is:

"Multiple accounts are permitted. You're welcome to create and maintain two separate Claude accounts using different email addresses... Your dual-account strategy effectively gives you additional capacity beyond these standard options... Your setup appears fully compliant with our terms." — Anthropic Support (April 9, 2026, Conversation ID: 215473842203754)

The fact that a Max 20x user needs a second $200/mo subscription just to get through the week suggests the default settings are not calibrated for the plan's intended audience. If effort=85 weren't causing retry spirals, one Max plan should be more than sufficient.

What I Expected

The Max 20x plan ($200/mo) should be sufficient for a single power user. Hitting the weekly cap AND burning $361 in extra usage within 2 days suggests the default settings are not optimized for the plan's intended audience.

Suggestions

1. Default Max plan users to effort=high — Boris mentioned this is being considered for Teams/Enterprise (#42796). Max 20x users are paying for heavy usage — they should get the full thinking depth by default.

1. Show estimated token cost per turn — Users have no visibility into how much each interaction costs. A simple "~X tokens" indicator would help users make informed choices about effort levels.

1. Effort-aware quota tracking — The weekly limit doesn't distinguish between a high-effort turn that succeeds once and three medium-effort turns that fail twice then succeed. Both count equally against the quota, but the user experience and cost are vastly different.

1. Warn when retry patterns suggest effort is too low — If Claude Code detects repeated corrections on the same task (e.g., 3+ edits to the same file/function), suggest bumping effort for that task.

Workarounds I'm Using

• /effort high (sticky, per Boris's recommendation in #42796)
• Switching to Sonnet for simple tasks (explains the 6% Sonnet usage)
• Second Max subscription under a different email (confirmed compliant by Anthropic support — but this shouldn't be necessary)

Claude Code Version

Latest (as of April 2026)

Platform

macOS (Apple Silicon)

Impact

High — Monthly cost approaching $561+ ($200 Max + $361 extra usage) for a single user within the first 2 days, primarily driven by the medium-effort default causing retry loops on complex tasks. A second Max subscription ($200) is being considered as a workaround, potentially bringing total monthly cost to $761+.