Model behaviour issues on Windows
Issues the maintainers labelled both platform:windows and area:model.
1,301 issues · 469 open · 832 resolved (64%) · first seen May 31, 2025
Is this getting better or worse?
This class of problem is still growing. 671 new reports in the last 90 days vs 358 in the 90 before — +87%. The open backlog peaked at 469 in 2026-10 and sits at 469 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
901 of these reports name the Claude Code build they were running, spanning 180 releases. Heaviest reporters:
- Claude Code v2.1.289
- Claude Code v2.1.296
- Claude Code v2.1.280
- Claude Code v2.1.233
- Claude Code v2.1.71
- Claude Code v2.1.278
- Claude Code v2.1.237
- Claude Code v2.1.119
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 36 days across 832 closures. Of the 832 closures with a recorded reason, 10% were closed as completed and 749 as not-planned or duplicate. The most recent completed fix landed in 2026-10.
Workarounds reported by the community
120 of these issues post a workaround someone says worked. The highest-engagement ones:
- [BUG] Advisor always "unavailable" with Fable 5 advisor (Opus 4.8 main) across all sessions (v2.1.198)
- Right now, my workaround is just launching a Fable 5 agent. Not ideal since seamless handoff is preferred, but it works. Does anyone know if the advisor mechanism actually saves on Fable 5 tokens, considering how differe…
Found in the thread of #73365 · resolved - Recurring error: "The model's tool call could not be parsed (retry also failed)" interrupts sessions
- Workaround: use an English or French session instead of Chinese / Japanese First-person data point supporting the non-ASCII trigger suspected here: I kept hitting while working in Chinese . Opened a new session in French…
Found in the thread of #63875 · resolved - [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 - Subagent model routing is broken — all mechanisms resolve to parent model (Opus)
- …pe), which the changelog doesn't mention. The full-model-ID workaround noted above likely still applies there.
Found in the thread of #43869 · still open - Hard block on typing passwords breaks legitimate dev/test workflows - needs a permission-gated opt-in for own dev environments
- …rver-side and dies on every app restart. Claude's suggested workaround was exactly that, and it collapsed on first contact with reality. We ended up building a passwordless dev-auto-login endpoint into the app (gated to…
Found in the description of #78160 · 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 1,301 matches; the statistics above cover all 1,301.
- [BUG] Advisor always "unavailable" with Fable 5 advisor (Opus 4.8 main) across all sessions (v2.1.198)
- Recurring error: "The model's tool call could not be parsed (retry also failed)" interrupts sessions
- [BUG] Can't see Claude's thinking anymore
- Subagent model routing is broken — all mechanisms resolve to parent model (Opus)
- Hard block on typing passwords breaks legitimate dev/test workflows - needs a permission-gated opt-in for own dev environments
- Claude Code changed default model to 1M context without notifying Pro users
- [BUG] Opus 4.7 Hallucinations
- Opus 4.6 comprehensive regression: loops, memory loss, ignored instructions - daily professional user report
- [BUG] Silent billing change in v2.1.51: 1M context moved to extra-usage-only without notice — JSONL data proves identical workload billed differently
- [BUG] Opus 4.8 repeatedly emits malformed tool_use blocks, entire response discarded (4.7 works fine)
- CLAUDE.md hard rules and persistent memory instructions consistently ignored — violations escalate with each session despite repeated explicit reinforcement
- [BUG] VS Code extension blocks Fable 5 with "requires usage credits" while account has 20% of weekly Fable allowance remaining (Max plan)
- Opus 1M context on Max plan requires extra usage despite docs saying 'included with subscription'
- [BUG] CLAUDE.md instructions completely ignored after context compaction
- [BUG] Severe over-triggering of safety guardrails — Claude hallucinates policy violations in completely harmless prompts
- [BUG] extended thinking doen't work (think, think hard, think harder)
- /model opusplan fails with "Unsupported model" - worked for months until today
- [BUG] Tool-use markup generated as "court" instead of "antml:invoke" (malformed/parse error, tool not executed)
- Claude Opus 4.6 Behavioral Degradation: Confident Unverified Analysis Pattern (15-Day User-Documented Evidence)
- Auto-compact fails with Opus 4.5 extended thinking blocks
- [BUG] 1M context window removed from Desktop Code tab model picker after v1.1.7714 update on Max plan
- [BUG] 400 "output_config.effort 'xhigh' is not supported when thinking is disabled" on Opus 4.8 despite alwaysThinkingEnabled: true (VS Code, 2.1.205–2.1.207)
- Opus 4.8 returns empty thinking blocks — no thinking shown in chat (same regression as Opus 4.7 #49268)
- [Bug] Fable 5 incorrectly requires usage credits on Max 20x plan despite available weekly allowance
- Claude Code suggests non-existent "fan out subagents" command
- Opus/Sonnet 1M context variants broken on Max 20x — misleading 'Rate limit reached' error
- [Bug] Claude Code: Persistent Instruction Disregard and Output Quality Degradation
- [BUG] Usage credits required for 1M context - Pro plan blocked despite 17% usage
- Stray "call"/"court" token before tool calls; internal <invoke> XML printed as text instead of executing (Windows local + cloud Cowork)
- 🚪🤜 KNOCK KNOCK...
- [BUG] Opus 4.6 with 1M context is not available on Max 5x plan
- [BUG] Opus 4.7 [1M] context capped at 500K on Max x20 — server-side enforcement persists across fresh install (org_level_disabled stale flag)
- [BUG] Sonnet 4.5 in new Claude Code v. > 2. degraded quality, adherence to rules, premature victory, not following and forgetting rules been told 1 prompt earlier
- Opus model quality regression — significant degradation starting ~April 15, 2026
- Messages meant for the user are sometimes emitted as hidden thinking before a tool call
- [BUG] A complete CLAUDE.md rule contract governed nothing: 9 fabricated causes, stale state asserted as current, acceptance step silently skipped across a 4.5h session
- Opus 4.7 regression vs 4.6: ~2x token burn, mid-session stalls, no quality gain
- [Meta] Systematic completion-integrity failures — 16 issues across 6 phases from 100+ sessions of evidence
- Misleading "Usage limit reached" error when selecting Sonnet — actually a 1M context tier gate
- Context window status bar shows 200k for Claude Sonnet 4.6, but model supports 1M tokens
- [BUG] Claude code quality (Opus 4.6) has degraded
- [Bug] Claude never proactively surfaces its own mistakes — requires user to act as quality gate
- [BUG] 'CLAUDE_CODE_MAX_OUTPUT_TOKENS' incorrectly validates Sonnet models against 32k limit instead of 64k
- [Feature Request] Configurable target model and effort for the Fable 5 classifier fallback
- [MODEL] Claude uses Bash (cat, grep, head) instead of dedicated tools despite system prompt prohibition
- [MODEL] Opus 4.8: runaway extended thinking (20k-64k output tokens/turn), replies to hallucinated user messages, fabricates forensic "evidence" when asked to investigate
- Opus 4.8 (1M context) — frequent "tool_use malformed" client errors despite tools executing successfully
- [BUG] Model Error
- [BUG] [1M] "Long context beta is not yet available for this subscription" error on Max plan ($200/mo)
- Claude Code VS Code extension forces 1M context on Pro plan - Windows
- [BUG] Haiku agents do not work if MAX_THINKING_TOKENS is set.
- [BUG] VIBE CODED STUPID ANTHROPIC USAGE POLICY FLAGS EVERYTHING FOR NO REASON
- [Bug] System prompt outdated: Claude Opus 4.7 misidentified as Opus 4
- Authorized bounty/CTF security research blocked mid-session — need context-aware handling for whitehat workflows
- [Bug] Anthropic API Error: Cannot modify thinking blocks in multi-turn conversation
- [BUG] Haiku 4.5 not available in v2.0.19
- [Bug] Claude Pro Opus Model Not Accessible in Claude Code
- [BUG] Model continues past end of turn, emitting a fabricated user turn and a fabricated system-reminder block inside its own assistant message
- [MODEL] Claude Code CLI: Opus 4.8 hallucinating tool calls, user messages, and system prompts
- [BUG] Claude Code spirals questioning the spelling of AssertionError
- [Bug] Usage credits required for 1M context on Max 5x plan — standard model selected, Cowork + Claude Code
- [Bug] Context amnesia in long sessions — constraints silently dropped as context grows
- [BUG] Opus limit reached shown even if Opus hasnt been used
- [MODEL] Desktop Code tab refuses all local file deletion (browser-agent safety rule), then behaves inconsistently under pushback
- [BUG] Opus 4.8 missing from /model picker in CLI and VS Code extension
- [MODEL] Uses PowerShell here-string (@'...'@) inside the Bash tool on Windows, silently corrupting a git commit message
- Opus 4.7 ([1m] variant) capped at 200K context in Claude Code Desktop on Max plan
- [BUG] v2.1.32 / Opus 4.6 regression: New tab unresponsive + EBUSY .claude.json lock
- Model Overrides Documented Protocols Even When Context Is Present (756 Deviations)
- InputValidationError: long unicode-escaped tool arguments fail JSON parse (AskUserQuestion) — retry with shorter input succeeds
- [BUG] Opus 4.7 - An AI model with Down syndrome
- [BUG] 1M context unavailable on Max plan — Windows 10
- [FEATURE] Opus 4.8 (1M context) model option disappeared from the model picker
- [BUG] Fable 5 advisor silently disabled by its own safety classifier — no Opus fallback, generic "unavailable", sticky-off for the session
- [BUG] False positive: Usage Policy violation triggered by legitimate input
- [BUG] False positive usage policy violation on computational biology research content
- [BUG] Cowork: Agent tool `model` parameter silently ignored — all sub-agents routed to Haiku
- Advisor tool: Haiku 4.5 executor rejected by CLI despite API docs listing it as valid pair
- [BUG] Claude Code ignores explicit instructions in memory/CLAUDE.md files
- Default high think effort causes Opus 4.6 to ignore skills and CLAUDE.md rules
- [BUG] Opus 4.5 showing in Claude Code CLI, but not showing in Claude Code VSCode extension
- [MODEL] Severe multi-symptom degradation since 2026-06-08 on Opus 4.8 (Desktop app): ignores CLAUDE.md, bypasses permission prompts, hallucinates, refuses doable tasks, writes files unprompted
- [BUG] Raw <invoke> XML printed to terminal and session hangs — requires user message to resume (Windows, v2.1.168)
- [BUG] Switching to mimo-v2.5-pro fails after reading images in mimo-v2.5
- Opus 4.7 is lazy and asks for too many confirmations even on auto mode
- [Bug] Claude silently skips documented verification steps instead of asking the user
- [Bug] Anthropic API Error: Reasoning extraction safeguard false positive on consecutive tool calls
- [BUG] Opus 4.8 regression now also hits Sonnet 4.6 — the "switch to Sonnet" workaround no longer works (June 22–23, 2026)
- [Bug] Opus 4.8 responses not displayed after tool use — quota consumed with no output
- Model picker selects Opus 4.7 1M when choosing Opus 4.7 (200k), possibly after compaction
- [Bug] Claude defaults to assertion from memory instead of verification from tools
- [Bug] Claude reads files but ignores actionable instructions contained in them
- [BUG] Fable 5.1: final answer emitted as a thinking block (not text) before AskUserQuestion, so the user never sees it
- [BUG] Fable 5 is disabled in my claude code terminal
- Sonnet 4.6 routes every request to long-context tier on Claude Desktop 2.1.149 (429 'Usage credits required')
- [BUG] Opus 4.7 safety filter produces false-positive refusals on encoded input
- [Bug] Claude does not verify file edits applied correctly — blind edit-and-proceed
- [BUG] Unauthorised launch of extremely inefficient subagents that wasted tokens - I want a refund
- Tool calls rendered as raw text and fail to parse with claude-opus-4-8 in VS Code extension
- Claude Opus - Poor algorithm anticipation and performance in code generation