Production Testing • Published Feb 7, 2025

I Tested Every Major Claude Opus 4.6 Feature — Here's What Actually Matters

15 min read By Alireza Rezvani ⚡ Claude Code Mastery

24 hours of real production testing — not benchmark summaries. The effort dial that's saving money daily, the 1M context window that actually works (with caveats), Agent Teams that cut task time in half but cost 4x more, and the SWE-bench regression nobody's talking about.

150+
Articles Published
2,800+
Medium Followers
22+
Years in Tech
12+
Months Claude Code

TL;DR — What You Need to Know About Claude Opus 4.6

🎚️

The Effort Dial Is a Game-Changer

Four reasoning levels (low → max) let you trade speed for depth. Medium matches Opus 4.5 at 76% fewer tokens. The author now toggles 10+ times per day — low for scaffolding, max for bugs that stumped three senior engineers.

📐

1M Context Window Actually Works

76% MRCR v2 score (vs 18.5% for Sonnet 4.5) — a 4x improvement in deep-context reasoning. Cross-file debugging that took 12+ manual references now works in a single prompt. Precision dips past ~800K tokens.

🤖

Agent Teams: Promising but Rough

Parallel agent execution cut a 3-4 hour task to ~90 minutes. But it costs 4x the tokens ($12-15 per session), has coordination failures, and experiences silent stalls. Research preview — not production-ready.

⚠️

Regressions Nobody's Mentioning

Small dip on SWE-bench Verified and MCP Atlas. 128K output tokens + Agent Teams = surprise bills. Context compaction is useful but lossy. US-only inference at 10% premium. Real production trade-offs to consider.

Key Findings from 24 Hours of Production Testing

Every claim tested against real codebases, real costs, and real deadlines — not synthetic benchmarks.

Finding 01

The Effort Dial Saves Money AND Improves Quality

At high effort (default), Opus 4.6 anticipated edge cases unprompted — rate limiting, graceful degradation, even a potential race condition the author hadn't considered. At max effort, it traced a concurrency bug in a single pass that three senior engineers spent 18 combined hours on. Medium effort matches Opus 4.5 quality at dramatically lower cost. This is the feature that changed daily workflow more than anything in the past year.

✓ Workflow-changing — use immediately
Finding 02

1 Million Token Context Removes the Multi-File Ceiling

A 47-file authentication system loaded in one shot. The model traced a permission bug from API gateway through middleware to database — identifying a stale cache issue in the third middleware layer. Previously required 12+ manual file references. The 15-file accuracy threshold that plagued Opus 4.5 effectively no longer exists.

✓ Major improvement — real production impact
Finding 03

Agent Teams Cut Time but Multiply Cost

A full microservice (API + DB + tests + docs) built in ~90 minutes instead of 3-4 hours. The lead agent caught interface inconsistencies between teammates before the author noticed. But: 4x token consumption ($12-15/session), coordination failures on environment variable naming, and a silent 8-minute stall with no error or timeout.

⚡ Wait 2-4 weeks for stability
Finding 04

The Regressions Are Real and Under-Reported

SWE-bench Verified score dipped below Opus 4.5 — relevant if your workflow is heavily issue-resolution focused. MCP Atlas benchmark also regressed slightly (matters for MCP-heavy setups). Context compaction loses specificity about early-session decisions. First full day of testing burned ~$45 in tokens.

⚠ Monitor before committing fully

Claude Opus 4.6 Effort Level Comparison

Real-world performance data from generating a Node.js service layer with error handling, validation, and database queries.

Level Speed Quality Best For
Low ~3 sec Structurally correct, generic Scaffolding, boilerplate
Medium ~8 sec ≈ Opus 4.5 at 76% fewer tokens Routine code, standard tasks
High (default) ~15 sec Anticipates edge cases unprompted Production-bound code
Max ~40 sec Deep reasoning, can over-explore Hard bugs, complex architecture

Claude Code Ecosystem by Alireza Rezvani

Production-tested tools built from 12+ months of daily Claude Code usage. MIT licensed.

Frequently Asked Questions About Claude Opus 4.6

The effort dial gives you four reasoning levels — low, medium, high, and max. Medium effort matches Opus 4.5 performance at 76% fewer output tokens, dramatically reducing costs for routine tasks. You can toggle between levels based on task complexity, using low for scaffolding, medium for standard work, high for production code, and max for the hardest problems.
Yes, with caveats. Opus 4.6 scores 76% on MRCR v2 (vs 18.5% for Sonnet 4.5) — a 4x improvement in deep-context reasoning. A 47-file authentication system was loaded and debugged in a single prompt. However, precision can decrease past ~800K tokens. It's the best available, but not infinite perfect memory.
Not yet. Agent Teams is in research preview and shows promising parallel execution — cutting a 3-4 hour task to ~90 minutes — but costs roughly 4x more tokens, has coordination failures between teammates, and can experience silent stalls with no error messages. Best to wait 2-4 weeks for stability improvements before building workflows around it.
Upgrade now if you're a Claude Code power user running complex multi-file tasks daily. The effort dial and 1M context window provide immediate workflow improvements. Wait if you rely heavily on SWE-bench performance (small regression) or plan to depend on Agent Teams (still in beta). Skip if you primarily use Sonnet for cost efficiency — the 10x price gap is still significant.
Opus 4.6 shows a small dip on SWE-bench Verified (GitHub issue resolution) and MCP Atlas (tool usage). Context compaction, while useful for extending sessions, can lose specificity about early-session decisions. Combined with 128K output tokens and Agent Teams, token costs can scale faster than expected — one full day of testing cost ~$45.
Alireza Rezvani — CTO and Claude Code Expert

About Alireza Rezvani — CTO & Claude Code Expert

CTO at LINDERA, a Berlin-based AI HealthTech startup. 22+ years of engineering leadership, 150+ published articles, and 12+ months of daily Claude Code production work. Creator of open-source Claude Code tools used by developers worldwide including claude-code-tresor, claude-skills, claude-code-skill-factory, and ClaudeForge.

I write about what actually works in AI-augmented development — production-tested insights, not theoretical takes. Every article is backed by real codebases, real costs, and real deadlines.