Anthropic’s Claude Max plans promise up to 20× more usage than Pro. But developers paying $100 or $200 a month are discovering the same frustrating reality: their session still drains in 90 minutes. The price went up 5-10×. The fundamental problem didn’t change.
🔍 THE BOTTOM LINE: Claude Max doesn’t fix the token waste problem — it just gives you a bigger budget to waste. 60-80% of every prompt’s tokens go to re-reading files the agent has no memory of. Paying more doesn’t solve that.
💸 The Pricing
| Plan | Monthly Cost | Usage vs Pro | Effective Session |
|---|---|---|---|
| Pro | $20 | 1× | ~45 messages / ~10-40 Claude Code prompts |
| Max 5× | $100 | 5× | ~225 messages / ~50-200 prompts |
| Max 20× | $200 | 20× | ~900 messages / ~200-800 prompts |
Those numbers look generous until you understand what actually consumes tokens.
🔥 The 90-Minute Problem
A single Claude Code prompt asking for a multi-file refactor can consume 150,000+ tokens:
- Auth middleware file: 4K tokens
- Grep for auth imports: 5K tokens
- 8 dependency files: 32K tokens
- 5 test files: 20K tokens
- Config and environment files: 8K tokens
- Full conversation history from earlier turns: ~80K tokens
- Generated output: 5K tokens
Total for one prompt: ~154K input tokens.
On a Max 20× plan with an estimated 500K session budget, that single prompt consumed 31%. Two more like it and you’re done.
Switch to Opus instead of Sonnet, and your effective session is cut in half — Opus costs more per token, so the same dollar budget buys fewer tokens.
⏰ Peak Hours Make It Worse
Since March 26, 2026, Anthropic confirmed that session limits burn faster during peak hours (5am-11am PT weekdays). The same prompt that’s fine at 8pm might drain your limit at 9am. The weekly total hasn’t changed — but the distribution has.
Developers who do their heavy work in the morning — which is most of them — get hit hardest.
📊 The 7× Math
The DataChaz tweet that kicked this off pointed out a simple calculation:
- Pro plan: $20/month for baseline usage
- Max 5×: $100/month for 5× usage
- Max 20×: $200/month for 20× usage
That’s 5× the cost for 5× the usage on Max 5× — which at least scales linearly. But Max 20× is 10× the cost for only 20× the usage — a 2× improvement in value per dollar.
The real problem isn’t the price-per-token ratio. It’s that 60-80% of tokens are wasted on context the agent re-reads every session because it has no persistent memory. Buying 20× more tokens means buying 20× more waste.
Upgrading from Pro to Max 20× is like buying a bigger gas tank for a car with a leak. You drive further before you stall — but you’re still leaking 70% of your fuel.
🇳🇿 NZ Impact
For NZ developers paying in NZD:
- Pro ($20 USD): ~$34 NZD/month
- Max 5× ($100 USD): ~$170 NZD/month
- Max 20× ($200 USD): ~$340 NZD/month
At $340/month, you’re paying more than most NZ cloud hosting bills for a tool that still runs out during a normal work session. For comparison, direct API access with prepaid credits is more predictable per-token, though it requires more setup.
🛠️ The Real Fix
The structural problem is context re-reading. Every session, the agent starts fresh and has to re-ingest your entire codebase to understand what it’s working on. No plan tier fixes this.
Possible solutions:
- Persistent context — Agents that remember file structures between sessions
- Pre-computed context graphs — Only send relevant file metadata, not full files
- Smarter context management — Agents that know what they’ve already read and don’t re-read it
- DESIGN.md-style specs — Structured context that replaces brute-force file reading (see our coverage of Google’s open-source DESIGN.md)
Until one of these is solved at the platform level, no amount of money solves the drain.
