Model behaviour issues on Linux
Issues the maintainers labelled both platform:linux and area:model.
2,108 issues · 444 open · 1,664 resolved (79%) · first seen May 14, 2025
Is this getting better or worse?
This class of problem is converging. 664 new reports in the last 90 days vs 1247 in the 90 before — -47%. The open backlog peaked at 908 in 2026-07 and sits at 444 today.
Derived by counting each matching issue from its creation date until it closed. First-party
computation over the anthropics/claude-code tracker — the trend is not published
anywhere upstream.
Which builds report this
923 of these reports name the Claude Code build they were running, spanning 184 releases. Heaviest reporters:
- Claude Code v2.1.201
- Claude Code v2.1.170
- Claude Code v2.1.207
- Claude Code v2.1.198
- Claude Code v2.1.296
- Claude Code v2.1.263
- Claude Code v2.1.251
- Claude Code v2.1.283
Report volume, not a defect count: an issue can mention a build without that build having caused the failure. Versions come from the version string in the bug-report template, matched against real published tags.
Do these get fixed?
Median time to close is 47 days across 1,664 closures. Of the 1,664 closures with a recorded reason, 6% were closed as completed and 1,562 as not-planned or duplicate. The most recent completed fix landed in 2026-10.
Workarounds reported by the community
66 of these issues post a workaround someone says worked. The highest-engagement ones:
- [BUG] Can't see Claude's thinking anymore
- …problem as you. Looking through other issues, I found this workaround: Hopefully they fix it in general soon. In many cases, Claude is unusable when the thinking is not displayed. It runs for minutes, burns tokens, and…
Found in the thread of #24285 · still open - [BUG] max_tokens must be greater than thinking.budget_tokens
- workaround is possible by editing settings.json (in my case in ~/.claude/settings.json), by setting the "alwaysThinkingEnabled" to false.
Found in the thread of #8756 · still open - Tool calls intermittently emitted as literal text (stray 'court' + raw <invoke>) instead of executing — mostly Edit/Read (Opus, CLI)
- …k is now access-restricted for us — leaving no model-switch workaround — please consider raising severity above cosmetic. Happy to share a sanitized JSONL excerpt of the dropped-call events. Detector we used (reproduces…
Found in the thread of #64108 · still open - [Bug] Anthropic API Error: Usage credits required for 1M context window with opus-plan model
- …sn't work, and just stops working at 1M Context. A feasible workaround would be to periodically use the /compact command manually.
Found in the thread of #61869 · resolved - [BUG] Settings effortLevel "max" is silently downgraded when user interacts with effort selection UI
- …ot above. Using only one of the variables doesn't fix that. Workaround: - then it shows the effort in the welcome text, in the footer, and as the selected effort in .
Found in the thread of #30726 · still open
Excerpts are quoted from the linked issue so you can judge relevance before opening it. Threads that only ask for a workaround, or report that there is none, are excluded. These are community suggestions, not official guidance from Anthropic — read the full thread before applying one.
Matching issues 1–100
Ranked by community engagement (reactions weighted over comments). Listing the top 300 of 2,108 matches; the statistics above cover all 2,108.
- [BUG] Severe performance degradation
- [BUG] Can't see Claude's thinking anymore
- [BUG] max_tokens must be greater than thinking.budget_tokens
- Tool calls intermittently emitted as literal text (stray 'court' + raw <invoke>) instead of executing — mostly Edit/Read (Opus, CLI)
- [Bug] Anthropic API Error: Usage credits required for 1M context window with opus-plan model
- [BUG] Settings effortLevel "max" is silently downgraded when user interacts with effort selection UI
- Fable 5 model behavior: 16 min of silent tool calls, then AskUserQuestion about never-shared findings during incident response
- [BUG] Replies drift into English right after reading English tool output, despite a CLAUDE.md rule requiring Japanese; corrections do not hold
- [BUG] Fable 5: mid-turn assistant text blocks intermittently delivered as summarized thinking blocks (turn appears silent)
- Invalid API Model Name in Request
- [BUG] Claude Pro users are not currently able to use Opus in Claude Code
- [BUG] CLAUDE_CODE_MAX_OUTPUT_TOKENS has no effect on Opus 4.6 — output stays capped at 32000
- [Bug] Anthropic API Error: Long context beta unavailable for Max plan with Claude Opus 4.6 1M
- [Bug] Fable 5 safeguards overly broad, flagging routine coding tasks
- [Bug] Claude refuses to fill in password fields in testing environment
- [BUG] Thinking and not coding for more than 10 minutes for simple tasks with oppus 4.6 - then freezes
- [BUG] Server-side experiment (`x-cc-atis`) silently removed Opus 4.8 thinking summaries, and the CLI silently self-updated even with `autoUpdates: false`. No notice, no opt-in, settings silently overridden.
- [Bug] Agent spams no-op echo probe commands to flush shell output
- [BUG] Sonnet 4.6 blocked by false "usage credits required" error
- [Critical] Max Plan subscriber: Opus 4.6 instruction-following regression breaks production workflows — 24-hook enforcement system cannot compensate for model-level degradation
- [BUG] Claude using AWS provided inference profiles when application inference profiles are set
- [BUG] Cyber safeguards false positive: routine sysadmin audit commands blocked, write-only reporting blocked in new session, context poisoning breaks session recovery
- [BUG] degraded thinking performance since Dec/Feb still persisting (stop ignoring your paying customers pls)
- [BUG] /compact fails with cyber safeguards false positive during legitimate defensive security session
- [BUG] CLAUDE.md language/dialect instructions drift during long sessions (Spanish voseo leaks into neutral-Spanish output)
- [Bug] OPUS 4.7 IS DOWNGRADED JOKE
- [BUG] Task tool `model` parameter returns 404 - Blocks cost-optimized workflows (10-30x unnecessary spending)
- [BUG] 503 Model is getting throttled. Try your request again in opus 4 model
- [Bug] Anthropic API Error: Usage credits required mid-session
- [BUG] Opus 4.8 confabulates user messages, a fake "prompt injection attack" narrative, and fabricated tool/host facts in long sessions (2 sessions, JSONL-verified)
- [Bug] Anthropic API Error: Assistant message thinking block must be first
- Doing useless tool call wasting token
- [MODEL] Quantified evidence: Sonnet 4.6 quality regression since March 9 — 1400+ frustration events across 50 sessions
- API stream hangs mid-response on complex generation tasks (all models: Opus, Sonnet, Haiku)
- Fable 5.1 unreachable in Claude Code despite unused plan-included Fable allocation; credit-block dialog shows "not included in plan" copy with no affirmative action
- [BUG] Claude ignores instructions and violates rules consistently
- [Bug] Claude Opus 4.7 model instability requiring downgrade to 4.6
- [BUG] CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1 bypasses model capability checks, breaks Haiku subagents
- [Bug] Model Alias Not Supported in Custom Slash Command Configuration
- Max 20x plan drained in ~22 hours after Fable 5.1; 5.1 emits 2x the output per turn of Fable 5
- [BUG] Fable 5.1 burns tokens faster than 5.0 100% of limit in 20 mins.
- [BUG] [SUGGESTION] Fable classifier broken completely
- [Bug] Usage Policy classifier stuck in blocked state for legitimate kernel security work
- [BUG] Claude repeatedly mistakes current year as 2024 despite environment showing 2025
- OpenAI API: Unexpected Opus 4 Availability Restriction in Pro Plan
- [BUG] 2.1.63 silently defaults to 1M context model without UI indication, causing excessive usage consumption and a BILLING issue.
- [BUG] Extremely slow performance / near-zero token generation since Feb 10, 2026
- [BUG] Verified Evidence: Claude Code Delivers 10% of Requested Thinking Budget While Charging Full Price
- [BUG] haiku subagent delegates to 3.5
- [BUG] Claude Code is displaying ludicrous behavior. Guardrails are failing.
- [Bug] Claude 4.8 model stuck in infinite loop executing spurious bash commands
- [Bug] VERY sensitive cyber-safeguards
- [Bug] Extended Thinking causes 7–8 minute stalls with no output and silent token consumption
- Sudden reasoning quality degradation — model behavior change April 9-10, 2026?
- [Bug] Opus 4.6 default model alias not using 1M context window
- [BUG] Pipe operator (|) in quoted strings incorrectly interpreted as shell pipe in Bash tool
- [BUG] Fable 5 [1m] prompts 'usage credits required' during active 1M promo window (Jul 17, promo extended to Jul 19)
- [MODEL] Artificial degradation, Acquisition Bias, and unacceptable compute throttling for paid users
- [Feature Request] ultrathink should default to MAX or XHIGH effort level
- [BUG] German output: the model coins compound nouns that do not exist and uses real words in meanings they do not have; corrections do not survive the turn, the session, or a written rule (Opus, Fable, Sonnet)
- [Bug] Opus 4.7 xHigh performance degradation post-Colossus-1 rollout
- Significant quality regression — unauthorized actions and unreliable instruction following
- [Bug] Anthropic API Error: Opus model inaccessible from Claude Code despite Team plan access
- [BUG] since a couple hours ago, the claude code agent often gets stuck while working. downgrading doesn't help.
- [BUG] Opus 4.8 and 5.0 is consistently broken and unusable.
- Multi-minute first-byte stalls on claude-opus-4-8[1m], compounded by auto-mode classifier timeouts — sessions near-unusable
- [BUG] Severe quality regressions
- [BUG] [MODEL] Opus 4.6 and 4.5 systematic failure to read before acting — four+ incidents in one day
- [BUG] Memory Leaks with Opus 4.6 CLI - massive sudden spikes in RAM/SWAP usage
- [BUG] Opus usage charged when using Haiku model
- [BUG] Sonnet 4 is completely unsuable because of 429s
- Claude Code AI ignored direct user instructions, resulting in financial loss of $112.77
- [Bug] Claude hangs indefinitely during execution across all 4.6 models when executing a particular sequence
- [BUG] @agent-explore "Out of Credits" on $100 Max Plan
- [BUG] claude with model sonnet[1m] doesn't accept single messages > 200.000 tokens
- [BUG] Fable 5 routes assistant prose before tool calls into non-rendered 'narration' blocks — summaries preceding AskUserQuestion are invisible
- [Bug] Anthropic API Error: False positive cyber content block on legitimate local security review
- [BUG] Agent fabricated tool output and reported a non-existent prompt-injection "incident"
- [BUG] erroneous API Error: Claude Code is unable to respond to this request
- [BUG] VSCode extension forces 1M context on Sonnet 4.6 for Pro users with no way to opt out**
- [BUG] [Opus 4.7] Korean lexical fixation: repeated insertion of "영역" in unrelated output
- [Bug] Claude ignores --model flag in prompt mode, uses Sonnet instead of selected model
- [Bug Report: Degraded code generation quality and reasoning accuracy in recent days]
- Claude ignores explicit instructions in skills/commands when they conflict with default safety behaviors
- Instruction-following regression after recent Claude Code update - agents skip verification steps
- [BUG] Frequent crashes when running agents requiring repeated --resume cycles (Sonnet 4.5)
- [BUG] general-purpose agent sets model as 'sonnet' rather than 'inherit'
- [Bug] Model downgrade despite authorized work and CVP status
- [MODEL] Prose hard-wrapping and max-width capping recur across sessions despite CLAUDE.md and saved memory
- [BUG] Cyber safeguards being hit despite being enrolled in CVP/silently being removed from CVP?
- [Bug] Auto-downgrade from Fable to Opus triggered by false-positive safety flag on legitimate OSS repository
- [Bug] Claude Code incorrectly downgrades to Opus for legitimate privacy/compliance scanning tool
- [MODEL] Opus 4.8 (xhigh): claims work verified when it isn't, self-contradicts, and over-parallelizes contradictory tool calls — on a low-complexity codebase
- [BUG] Claude Code with Opus 4.7 just started having a conversation with itself, and began unilaterally taking action on my workstation
- [Bug] Model classifier rejects claude-opus-4-7 as invalid model
- [BUG] Silent model switch to Opus 4.7 [1M] mid-session caused ~4× quota burn
- [BUG] Claude Code for VS Code: cannot choose Sonnet 4.6
- [BUG] Claude Code resists re-analysis, delivers shallow work, admits to avoiding effort
- False positive Usage Policy violation on scientific chemistry codebase
- CLAUDE_CODE_EFFORT_LEVEL environment variable not reflected in /model UI