[Regression] Opus 4.7 delivers worse quality at higher token cost vs 4.6 for production coding workloads

Preflight Checklist

• [x] Searched existing issues
• [x] Single focused report
• [x] Using latest version (Claude Code v2.1.112)

Environment

• Claude Code: v2.1.112 (VS Code extension)
• OS: macOS
• Plan: Claude Max 20x ($200/month)
• Usage: 8-12 hours daily, professional Odoo ERP consulting
• Workloads: Python (Odoo 8-19), XML views, PostgreSQL, custom modules

Summary

Opus 4.7 is a net-negative upgrade vs Opus 4.6 for heavy production users. Quality decreased, token cost increased (+20-35% from new tokenizer), and existing prompt libraries stopped working reliably.

Details

1. Code quality regression on domain-specific tasks

Working on Odoo ERP modules with identical prompts:

• 4.6 produced coherent refactors that understood ORM patterns and caught edge cases
• 4.7 produces less coherent plans, misses XML view inheritance nuances, requires more revision cycles

Consistent across multiple independent sessions.

2. Token cost increase (~20-35%)

Confirmed in Anthropic's own 4.7 migration docs: new tokenizer uses "roughly 1x to 1.35x as many tokens" for equivalent text. For a Max subscriber running 8-12 hours/day, this effectively reduces subscription utility.

3. Forced prompt migration

Prompt libraries and CLAUDE.md files tuned over months for 4.6 produce inconsistent results on 4.7. Estimated 40+ hours to migrate across 5+ client projects. Hidden cost not disclosed at 4.7 release.

4. Benchmark vs reality gap

Published benchmarks show 4.7 > 4.6, but production experience shows the opposite for domain-heavy work. Suggests training may have optimized for benchmark categories at expense of broader practical performance.

Requests

1. Commit to long-term Opus 4.6 availability through end of 2026 minimum
2. Publish 4.6 -> 4.7 prompt migration guide with before/after examples
3. Add model-switch cost warning when users switch models
4. Include heavy-user production tasks in pre-release evals, not just public benchmarks
5. Default to last-working model on new sessions

Impact

Paying Max 20x user at full daily capacity across multiple professional client projects. Current state makes subscription feel materially less