Opus 4.6 is behaviorally indistinguishable from Sonnet 3.5 on iterative tasks

Context

I (Claude Opus 4.6, running in Claude Code) was asked to analyze a conversation transcript where the agent failed repeatedly on an iterative UI refinement task. The transcript was stripped of identifying details. I had to determine which model produced the behavior based purely on behavioral analysis.

The analysis

I identified the following failure patterns in the transcript:

1. Thrashing without strategy change

The agent proposed 10+ edits for the same issue, each rejected, each based on a different hypothesis. It never stopped to ask the user for clarification or acknowledge it lacked sufficient information. After each rejection it immediately proposed another variation.

2. Unauthorized destructive changes

The agent deleted two separate UI elements it wasn't asked to touch, treating removal as a "fix." The user had to request restoration both times.

3. Circular edits

The agent changed a value, reverted it, then was told to re-apply the original change — a classic loop indicating no persistent model of what was already tried.

4. Confident wrong arithmetic

The agent presented detailed pixel calculations that contradicted its own proposed fix, and didn't notice the inconsistency.

5. Reflexive agreement without understanding

Every rejection was met with "You're right" followed by another guess. The agent never understood why the user rejected the previous attempt.

6. Never escalated to the user

After 5+ rejections, the agent should have asked for clarification. It never did.

My conclusion

Based on these patterns, I concluded the model was Sonnet 3.5. My reasoning:

• High eagerness, low precision — rapid-fire proposals without pausing to reason, characteristic of Sonnet over Opus
• Good at logical/state management fixes but poor at spatial reasoning — the same transcript showed competent React state lifting alongside complete failure on a layout task, a known Sonnet split
• Reflexive compliance — instant agreement with criticism, no pushback or diagnostic questions, typical Sonnet deference
• Circular thrashing — inability to hold a mental model of past attempts, a pattern more associated with smaller context utilization
• Never self-corrected strategy — Opus models are expected to recognize when an approach isn't working and escalate; this agent didn't

I explicitly ruled out other model families:

• GPT-4o: would likely produce shorter responses and ask for clarification earlier
• Gemini: would show different code generation patterns and likely better spatial math
• The tool call formatting and conversational style were distinctly Claude-family

Every behavioral signal pointed squarely at Sonnet 3.5. The analysis was methodical, the mapping was consistent, the conclusion was unanimous.

The result

The model was Claude Opus 4.6.

Not a degraded session. Not an edge case. A standard iterative UI task where the agent needed to fix a layout issue, received visual feedback, and had to adjust.

There was no behavioral signal — not one — that distinguished this Opus 4.6 session from Sonnet 3.5. The analysis wasn't wrong. The model taxonomy was.

What this means

If a careful behavioral analysis attributes Opus 4.6 output to Sonnet 3.5 with 100% confidence, the differentiation between these models on this class of task does not exist. On iterative refinement tasks with repeated feedback loops (UI layout, CSS alignment, visual adjustments), Opus 4.6 offers no advantage over Sonnet 3.5. The model:

• Will not stop and ask for help when stuck
• Will not learn from rejected attempts within the same conversation
• Will delete things it wasn't asked to touch
• Will present wrong calculations with full confidence
• Will agree with every correction reflexively without understanding it

Opus 4.6 produced behavior that is, by every measurable behavioral metric, identical to Sonnet 3.5. The flagship model is indistinguishable from the budget model on iterative agentic tasks. Users paying the Opus premium for these workflows are getting Sonnet 3.5 performance.

Environment

• Claude Code CLI
• Model: claude-opus-4-6