When Claude's answers feel worse than they used to, there's no way to verify from the user side whether overall model performance has actually changed. What you can check directly, though, is a set of concrete factors on your own screen: how long the conversation has gotten, which model and settings are currently answering, personal customizations you saved earlier, and how much material is in your project knowledge. Working through this checklist takes about 5 to 10 minutes and narrows down the likely cause.
🟢 Current lineup (September 2026): Claude Opus 5.5 / Claude Sonnet 5 / Claude Haiku 4.5 (higher tier: Claude Fable 5.1). · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (
Before you start, it helps to know three things:
- Which plan you're on — Free, Pro, Max, or Team
- Whether you're chatting on the web/app (claude.ai) or using Claude Code (the terminal-based coding tool)
- Whether the problem shows up in one specific conversation, or persists even in a brand-new one
Quick terms
- Context window — the maximum span of conversation and documents Claude can "hold in mind" and reference at once.
- Compaction — the automatic process that turns older messages into a summary to free up space once a conversation nears the context window limit.
- Effort — a setting that controls how much work Claude puts into a single answer (how much it double-checks and verifies).
- Extended thinking — a toggle that has Claude lay out its reasoning process before giving a final answer.
- RAG (retrieval-augmented generation) — instead of re-reading everything you uploaded every time, Claude searches for and pulls in only the parts that look relevant to your question.
If the conversation has gotten long, suspect context compaction first
If answers feel worse in a long-running conversation, the fastest thing to check is whether the context filled up and older content got automatically summarized (compacted). On claude.ai, if you're on a paid plan (Pro, Max, Team, or Enterprise) with code execution enabled, Claude automatically summarizes older messages once a conversation nears its limit, so the conversation doesn't get cut off. You can still scroll back and see the full chat history, but the key point is that the "working context" Claude actually uses to generate answers has been compacted. Once that happens, detailed instructions or examples from earlier in the conversation may no longer be reflected well in new answers.
How to check: Ask the same question in a brand-new conversation. If the answer noticeably improves, context is likely the cause. If you're using Claude Code (the terminal tool), you can check directly with this command:
/contextThis shows what's taking up space in your current context. To free up space while keeping the thread of an ongoing task, use /compact; to start a completely new task, reset with /clear.
Caution: /clear wipes the entire conversation context. If there's important context from earlier you still need, /compact (which keeps a summary) is the safer choice.
Check which model is actually answering right now
If answers suddenly feel thin, check whether a different model — or a lower setting — is answering instead of the one you meant to use. On claude.ai, the model selector menu in the chat window controls three things together: which model you're using, what level of effort it puts in, and whether extended thinking is on. If any of these has quietly dropped to a lower setting, answers can feel simpler than before.
How to check: Open the model selector at the bottom of the chat window and confirm the currently selected model and effort level. The exact menu labels and level names are best confirmed directly on screen, since they can vary.
There's one easy-to-miss case worth knowing. According to official support documentation, Opus 5 runs an automated safety check on every request, and for a narrow set of higher-risk requests — such as offensive cybersecurity techniques — it automatically falls back to a less capable model (Opus 4.8) and answers within the same conversation instead of refusing outright. When this happens, you'll see an on-screen notice explaining the switch, along with the name of the model that actually answered. The catch is that the model picker then stays on the less capable model for the rest of that conversation — so if answers keep feeling off afterward, open the model picker and switch back manually. If you were doing security-related coding work when quality suddenly changed, this is worth checking first.
If you're using Claude Code, you can check and change the current model directly:
/modelThis shows the currently selected model; appending a name, like /model opus, switches immediately. /model default reverts to whatever default is configured in your settings file.
Check whether custom instructions, styles, or memory are shaping the answer
If the model and context both check out, the next thing to look at is whether a personal customization you set up earlier is now working against the task at hand. The claude.ai settings screen has a field for personal instructions that apply automatically to every new conversation (check the exact menu label on screen, since it can vary). If you once added something like "keep answers short," that instruction can make a currently detailed task feel underserved. The same goes for the Style feature in the chat window — if a terse style is still switched on, answers can come across shorter and less thorough than before.
Memory is worth checking too. According to official support documentation, Claude records things like your role, feedback you've given, and ongoing work in real time during conversations, and draws on that in later ones. If something outdated is stored there, Claude may reference it and steer answers in the wrong direction.
How to check: Open the Memory item inside Settings to review everything saved so far, and edit or delete anything outdated or no longer relevant. If you use Projects, memory is stored separately per project, so check the memory for the specific project where the problem is happening.
If your project knowledge is large, retrieval may have switched modes
If you've uploaded a lot of material to a Project, the way Claude reads that material may have changed automatically. According to official support documentation, once project knowledge grows large enough, it automatically switches to RAG (retrieval-augmented generation) mode. This turns on by itself with no setup required, and instead of referencing the entire document set at once, Claude searches for and pulls in only the parts that look relevant to your question. As a result, details it used to find reliably can sometimes get missed.
How to check: Think back on whether you've added a lot of new material to project knowledge recently. If so, phrasing questions more specifically — referencing a particular document or section by name — helps the search surface the right part.
Checklist at a glance
| What to check | Where to look | How to check |
|---|---|---|
| Conversation length / compaction | Same conversation vs. a new one | Retry the same question in a new chat; Claude Code: /context |
| Model / effort / extended thinking | Model selector menu | Confirm current selection; Claude Code: /model |
| Model switched by a safety classifier | Notice above the answer | Check for a switch notice and whether the picker is stuck on a lower model |
| Custom instructions / style | Settings · chat window menu | Review saved instructions and any active style |
| Memory | Settings > Memory | Review saved entries; delete outdated ones |
| Project knowledge size | Project knowledge list | Check recent upload volume; ask more specific questions |
Claude Code users should also check usage limits
If you're using Claude Code (the terminal-based coding tool), it's worth checking usage limits separately from answer quality. According to official documentation, Claude Code shares its usage pool with claude.ai chat and Cowork, and both a 5-hour limit and a weekly limit apply together. As you approach a limit, the interface displays how much usage remains — so if your workflow keeps getting interrupted, that may be a usage-limit issue rather than a quality issue. Exact limit numbers vary by plan and can change over time, so treat the on-screen notice as the source of truth.
If you've checked everything above and still can't find the cause
If none of the above explains what you're seeing, the issue may be on the service side rather than your settings. In that case, the most reliable step is checking the official status page for any ongoing incident or degradation notice. If the problem only shows up for a specific kind of task — a long code review, or a complex research question — try that task again with effort or extended thinking turned up.
Frequently asked questions
Q. Is it actually true that Claude's overall performance has gotten worse?
There's no official way for a user to confirm an overall performance change from their side. What actually helps is working through checkable factors like this checklist covers — context, model and settings, and personal customizations.
Q. Does starting a new conversation really help?
If context compaction was the cause, a new conversation should show improvement immediately. If the same problem persists in a fresh conversation, look at model settings or personal customizations instead.
Q. How can I tell if the model switched without me noticing?
In one officially documented case, Opus 5 automatically falls back to Opus 4.8 for a narrow set of higher-risk requests, and displays an on-screen notice along with the name of the model that actually answered. The model picker then stays on the lower model for the rest of that conversation, so open it and check directly if something feels off.
Q. How do I check which model Claude Code is currently using?
Run /model to see the current model, and append a model name, like /model opus, to switch immediately.
Related articles
- Claude Not Working? Fix Login, Access & Errors (2026)
- Claude File Upload FAQ: Formats, 500MB/20-file Limits, PDF Handling
- Fixing Claude's 'File Limit Exceeded' Error: Upload Limits and How to Fix Them
- Troubleshooting Claude Error Messages: Limits, Capacity, Login
- Fix "Your Previous Message Wasn't Sent" in Claude — Causes & Steps (2026)