I’ve been using Claude Pro for months as a software engineer and loved it for complex algorithm development and debugging large applications.
Recently upgraded to Claude Max to access Claude Code features, even though I rarely hit Pro limits on the web interface. The higher tier seemed worth it for the additional tools.
Since switching plans, I’m experiencing a dramatic drop in response quality. What used to be maybe 2% incorrect answers on Pro has become roughly 85% unreliable responses on Max.
Sonnet 4 was exceptional before the upgrade but now produces basic errors. Opus 4 seems completely unreliable for development work.
Example of current behavior:
Claude: “Error found! Your code let result = value * 50; uses an undefined variable ‘value’.” Me: “But ‘value’ is clearly declared three lines above this statement.” Claude: “My mistake, you’re correct. The variable is properly declared earlier.”
These fundamental mistakes were almost nonexistent before. Now accurate responses are uncommon.
I’m wondering if Max subscribers get routed to different infrastructure where performance is sacrificed for throughput, even though the models should be identical.
Has anyone else noticed quality differences between subscription tiers? I’m considering downgrading back to Pro if this continues.
Looking for input from other developers who can evaluate model reasoning capabilities. Non-technical users might not notice these subtle but critical differences in code analysis quality.
Same here, but with a twist. I’ve been on Max for six months and the quality problems aren’t the same across all models. Sonnet 4 mostly works fine, but Opus 4 is garbage for multi-file codebases. It loses context way too often. Here’s what’s weird - switching conversation threads actually helps. Bad responses in one thread? Start fresh with the exact same prompt and it works better. Makes me think it’s session problems, not the models getting worse. Have you tried the API instead of the web interface? I switched to API calls for important dev work and it’s way more consistent. The web interface probably has different routing or caching that screws with performance. Extra work but worth testing before you downgrade.
Been dealing with this for years across different platforms. What you’re describing sounds like API rate limiting or request queuing issues, not model degradation.
Max subscribers probably get dumped into higher capacity pools that share resources differently. I’ve seen this when companies scale - they optimize for concurrent users but individual response quality takes a hit.
What works for me: add explicit context reminders in every prompt. Don’t assume Claude remembers that variable from earlier - repeat key stuff like “Given that ‘value’ is declared as let value = 10 above”. Annoying but it forces the model to focus.
Also restructure your prompts. I put the actual question at the very end after all context. Models lose focus parsing long conversations, especially on shared infrastructure.
Those debugging issues scream context window problems under load. When this happens, I step through code analysis systematically instead of asking for broad reviews.
This covers debugging approaches that work regardless of which AI you’re using. Check out the systematic troubleshooting methods.
If nothing helps after a week, downgrade and test side by side. Max features aren’t worth unreliable responses for critical dev work.
Yes, I faced a similar issue during my initial month on Max. It seems that the problem is less about the subscription tiers themselves and more about managing conversation length. Max struggles with extensive chat histories in a way that Pro did not. To improve response quality, I’ve found that starting new conversations every 15 to 20 messages can make a significant difference, especially during debugging sessions. Shorter conversations tend to yield responses comparable to what I experienced on Pro.
Additionally, I’ve noticed that the web interface for Max processes information differently, possibly due to extra middleware, as prompts sent via the API provide more reliable outputs. Before you consider downgrading, it would be wise to experiment with keeping conversations concise; if the earlier responses seem reminiscent of your time on Pro, it suggests that this is indeed a conversation handling issue rather than a degradation in the model itself. The features of Max might still offer value if you can adjust to frequent new threads.
This sounds backwards, but you’re probably hitting server load issues during busy hours, not actual model problems. I had the same thing happen about two months after upgrading to Max - responses got inconsistent and full of errors. What fixed it was switching when I used Claude. Early morning EST and late evening worked way better than midday sessions. The infrastructure thing makes sense since Max users probably get priority routing, but that doesn’t mean better hardware. I also noticed complex prompts get hit harder than simple ones when things are running slow. Try breaking your debugging requests into smaller, focused questions instead of dumping entire code blocks at once. Same context window, but processing seems shakier under load. Before you downgrade, test this for a week and see if timing matches quality. If evening sessions consistently work better, you’ll know it’s infrastructure, not your subscription tier.
This sounds like model versioning issues, not infrastructure problems. Anthropic pushes updates constantly and max users probably get experimental builds first. I’ve noticed quality swings every few weeks that stabilize after a couple days. Your timing with the upgrade might just be bad luck - you switched right when they deployed a buggy version. Give it another week before downgrading. These things get patched quick once enough people complain.
