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

[consdevs]

Preflight Checklist

• Searched existing issues
• Single focused report
• 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

[Yaakov-Belch]

Adding a dimension to this report: the cost regression is worse than the tokenizer inflation alone suggests.

Opus 4.6 was paradoxically cheaper per task than Sonnet for many workflows, despite higher per-token pricing, because it was so token-efficient — fewer turns, less backtracking, less verbose output. That made it a genuine sweet spot as a default model. Opus 4.7 breaks this in three compounding ways: the new tokenizer inflates input tokens (1.0–1.35×), xhigh-default thinking inflates output tokens, and the more literal instruction-following means more prompt-engineering effort (and often more correction turns) to get equivalent results.

But cost is a symptom. The deeper issue is a personality discontinuity. Every Anthropic model from early Haiku through Opus 4.6 felt like the same personality maturing — getting smarter while retaining an interpretive, collaborative style. 4.7 is a different personality: more literal, less willing to infer intent, more verbose. It may win coding benchmarks, but for users who relied on the interpretive style for production work (documentation, domain-specific code, collaborative iteration), the "improvements" are regressions.

Anthropic's own prompting docs acknowledge this indirectly: "Opus 4.7 takes the instructions literally... where previous models interpreted instructions loosely." That's framed as a feature, but for many of us it means a model that requires more handholding — more junior, not more capable.

I cannot use Opus 4.7 for my work. Removing 4.6 will force me to switch away.

[github-actions]

Closing for now — inactive for too long. Please open a new issue if this is still relevant.