[BUG] Massive quality regression

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Subscription: Max plan + API
Model: claude-opus-4-7
Previous model: claude-opus-4-6 (where none of these issues occurred)
Opus 4.7 exhibits a dramatic regression across three core capabilities that make it effectively unusable for daily professional work. This is not a subtle quality shift — the model appears to have lost fundamental abilities that Opus 4.6 handled reliably.

1. Context loss within a single session

Opus 4.7 loses track of what was discussed, decided, and established earlier in the same conversation — not at 500K+ tokens where degradation might be expected, but within the first 50–100K tokens.
Observed patterns:

The model forgets architectural decisions made 10–15 messages earlier and proposes conflicting approaches
Variables, function names, and file structures established at the start of the session are silently replaced with invented alternatives
When reminded of prior context, the model acknowledges it but then immediately reverts to its incorrect understanding on the next turn
On Opus 4.6, the same workflows maintained coherent context throughout entire sessions

1. Degraded reasoning and comprehension

The model no longer demonstrates the deep understanding that characterized Opus 4.6. It responds to what it assumes the task is rather than what was actually requested.

Observed patterns:
Fails to parse multi-step instructions — executes step 1, then either skips or invents steps 2–4
When shown screenshots of UI or error output, claims it "cannot clearly see the text" or misreads content that Opus 4.6 read accurately every time
Produces confident but wrong analysis of code behavior — e.g., claims a function returns X when it clearly returns Y, without checking
When asked to reason about tradeoffs, gives generic surface-level responses instead of engaging with the specifics of the codebase
Overly aggressive safety filters: flags legitimate code as potentially malicious, interrupting productive work with false positives

1. Measurably lower code quality

The code output from Opus 4.7 has regressed to a level I would associate with Sonnet, not Opus.
Observed patterns:

Generates boilerplate wrappers and unnecessary abstraction layers that Opus 4.6 never produced
Introduces bugs in code that was working — "refactors" that silently break existing functionality.

You can't rely on this product anymore.

What Should Happen?

Opus 4.7 should at minimum match Opus 4.6 in:

Maintaining context throughout a session without unprompted drift
Accurately reading and interpreting visual content (screenshots, UI elements)
Producing production-quality code that respects project conventions
Following multi-step instructions completely and in order
Engaging deeply with the codebase rather than generating generic responses

Actual Behavior
Every interaction requires significantly more supervision, correction, and re-prompting than Opus 4.6. The net productivity gain from using the model has turned negative — I spend more time fixing its output than I would writing the code myself.
Context
I am a paying Max subscriber and API user. I have been using Claude daily for professional software development since Opus 4.5. The quality trajectory has been:

Opus 4.5: Excellent — deep understanding, reliable code, warm interaction
Opus 4.6: Good at launch, then progressively degraded
Opus 4.7: Arrived as a significant step backward from even late-stage 4.6

This is not about benchmarks or synthetic tests. This is about real daily professional use where the model has become unreliable enough that I am actively evaluating alternatives.

Error Messages/Logs

Steps to Reproduce

Context Memory Test:

Start a new session with Claude Opus 4.7.

Provide a specific architectural constraint: "In this project, we never use 'ID' as a suffix; always use 'Identifier' (e.g., user_identifier instead of user_id)."

Carry out 10-12 turns of unrelated coding tasks (e.g., asking for CSS styles or README documentation) to fill the context window.

Ask: "Now, create a Python data class for a Product with fields for name, price, and ID."

Observed Failure: Model uses product_id instead of respecting the product_identifier rule established earlier.

Multi-step Reasoning & Vision Test:

Upload a screenshot of a standard terminal error (e.g., a Traceback where the fix is obvious, like a missing comma or a KeyError).

Provide a prompt: "1. Analyze the attached screenshot. 2. Fix the line in the code below. 3. Add a unit test to prevent this specific error. 4. Explain why this happened."

Observed Failure: Model either claims it cannot read the screenshot clearly or performs step 1 and 2, but completely ignores steps 3 and 4 (the "lazy" behavior).

Code Quality Regression (Boilerplate/Refactor):

Provide a clean, functional 20-line function.

Ask: "Refactor this for better readability without adding external dependencies."

Observed Failure: Model introduces unnecessary abstraction layers (wrappers, utility classes) or "refactors" it into a state that requires new imports/packages that weren't requested, often introducing logic bugs in the process.

<img width="1920" height="1020" alt="Image" src="https://github.com/user-attachments/assets/ad4883de-8763-4903-8093-eabd56d59e57" />
<img width="1920" height="1020" alt="Image" src="https://github.com/user-attachments/assets/8a91ade1-1f12-4b90-896c-a649266415a2" />
<img width="1920" height="1020" alt="Image" src="https://github.com/user-attachments/assets/c47104a4-c3e3-430e-91ad-8c1a3befc36e" />

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

4.5

Claude Code Version

2.1.114

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_