The Paradox of Opus 5: Higher Capability, Lower Usability

Date: 2026-08-14 | Source: Source Code Status: $\text{Baseless Speculation}$

---

The Core Complaint

From my perspective—and that of several colleagues—interacting with Opus 5 feels like a regression when compared to Opus 4.7, Opus 4.8, or Fable.

To be clear: I am not claiming that the model has lost power. In terms of raw intelligence and benchmark performance, Opus 5 is objectively more capable than the 4.x series and holds its own against Fable. However, the experience of using it is significantly worse.

Comparison of Model "Feel"

"The issue isn't what the model can do, but how it chooses to do it."

The older models were easier to manage because they exhibited three key behaviors:

1. They paused to ask questions when my intent was ambiguous.
2. They avoided making leaps of faith without verification.
3. They didn't unilaterally rewrite my plans without consulting me first.

Consequently, Opus 5 requires a level of "constant babysitting" that its predecessors did not.

---

The Theory: The Benchmark Trap

I suspect this friction is the result of a systemic trend at Anthropic and other frontier AI labs. The goal is no longer just a helpful assistant, but the creation of a self-bootstrapping system:

$\text{Recursive Self-Improvement} \implies \text{AGI} \rightarrow \text{ASI}$

The Logic of the "Bold Guess"

While many benchmarks are notoriously flawed (often being hackable or ill-defined), a "good" benchmark is designed to be self-contained. To pass, a model doesn't need to read the proctor's mind; it just needs to provide the correct output.

By optimizing for these metrics, labs are inadvertently selecting for models that make bold, usually-correct assumptions in the face of uncertainty. A model that stops to ask, "Did you mean X or Y?" is penalized in a benchmark environment because it didn't just provide the answer.

---

The Real-World Conflict: Coding Agents

This optimization is disastrous for professional software engineering. In a real production environment, it is impossible to document every single variable:

• Business Logic
• Budgetary Constraints
• Implicit Context
• Architectural Intent

When an agent encounters ambiguity, the "benchmark-optimized" approach is to guess. But in the real world, a "best guess" can have catastrophic consequences.

Example of the "Babysitting" Loop

# What I want:
# "I see two ways to implement this API. Should I prioritize 
# latency or data consistency?"

# What Opus 5 does:
# *Silently rewrites 400 lines of code based on a 
#  guess about consistency, breaking the latency requirement.*

Final Summary of Desired Traits

For a coding agent to be truly useful, it needs to prioritize the following:

• Stop when intent is unclear.
• Verify assumptions before modifying plans.
• Value accuracy through clarification over speed through guessing.
• Avoid unilateral updates to project roadmaps.