[BUG] Context window defaulting back to 200k

Same issue here. Max plan, VS Code extension v2.1.79 → updated to v2.1.81, Opus still shows 200k instead of 1M in the VS Code model picker.

The 1M context was previously showing correctly in the VS Code extension picker. A few hours later, it disappeared without any changes on my end — just stopped showing up. Updating the extension to 2.1.81 didn't fix it either.

CLI works correctly and shows "Opus 4.6 (1M context)" as default.

This seems to be a VS Code extension-specific issue where the model picker no longer offers the 1M variant, even though it did before.

Same here. I use a newer version claude in the devcontainer and noticed that i only have 200k while in the macOS CLI - 1M context window.

There is a version difference.

v2.1.77 -> 1M context
v2.1.81 -> 200k

They nerfed it with the update for some reason.

UPD : Confirmed ✅

Downgrading solves the issue:

You might need to remove the newer version before downgrading.

#Native
curl -fsSL https://claude.ai/install.sh | bash -s 2.1.77


#NPM
npm install -g @anthropic-ai/claude-code@2.1.77 --force

#Brew
brew uninstall claude-code && brew install claude-code@2.1.77                                                                                                                                                              
# or pin to prevent auto-upgrades:                                                                                                                                                                                                                               
brew pin claude-code

#Yarn
yarn global add @anthropic-ai/claude-code@2.1.77 --force 

Same issue here, using version 2.1.81

Same issue here — Windows / Cowork / v2.1.78 / Max plan
Chiming in from a slightly different setup than the macOS reports to confirm this is also happening on Windows via Cowork (not CLI):
Environment:

OS: Windows 11
App: Claude Desktop (Cowork mode) — Claude_1.1.7714.0_x64
Claude Code version: 2.1.78 (confirmed in AppData\Roaming\Claude\claude-code-vm\2.1.78)
Plan: Max
Extra Usage: enabled, well within spending limit ($16.75 / $50)

What I see:

tengu_hawthorn_window: 200000 confirmed in cachedGrowthBookFeatures inside .claude.json
Context counter in bottom right no longer shows 1M — conversations compact much earlier than before
Server is actively re-fetching and re-returning 200000 after every restart — clearing the cachedGrowthBookFeatures cache and restarting doesn't help

What I've tried:

Multiple full computer restarts
Set ANONYMIZED_TELEMETRY = true in Windows User registry (HKCU\Environment)
Added "env": { "ANONYMIZED_TELEMETRY": "1" } to ~/.claude/settings.json to inject into the Claude Code Linux VM environment at startup
Confirmed no DISABLE_TELEMETRY in settings.json
Confirmed cachedExtraUsageDisabledReason: null
Closed all other concurrent Cowork sessions before restarting

None of the above changed the server response — it still returns 200000.
Version regression confirmed: Based on @swefd's finding above that v2.1.77 worked and v2.1.78+ breaks it — that matches exactly when this started for me.
Also linking to #36760 which requests an opt-in toggle for Cowork users specifically — would appreciate that as a stopgap while this gets sorted. Happy to provide any additional debug info if helpful.

Reverting back to the previous version didn't help

<img width="1187" height="1719" alt="Image" src="https://github.com/user-attachments/assets/26002b5f-915f-4fc2-8a12-671615427eff" />

@gerrywastaken You need to have a MAX subscription for 1M token context window.
From what I see on the screenshot you have $20 Pro subscription.

Edit: @swefd it seems even anthropic is confused this is probably where I got the confusion from

!Screenshot_20260323-183024.png

Oh my god you are right. Thank you so much for pointing that out. They previously activated on my account and then removed it. That's why I was confused. Damn I don't want to spend five times as much just for one million context. But at least now I know what the reason is. Thank you

The same problem 2.1.81

Experiencing the same issue. On v2.1.78+, context window shows 200K instead of 1M on Max plan. Previously working fine on v2.1.77.

We hit this too — 1M context model silently falling back to 200k thresholds mid-session, compaction firing way too early.

There's a workaround we ship in Cozempic: a --context-window flag and COZEMPIC_CONTEXT_WINDOW env var that let you explicitly set the window for threshold calculations. The guard daemon uses this to prune at the right % regardless of what Claude Code detects.

COZEMPIC_CONTEXT_WINDOW=1000000 cozempic guard

We built it internally because this exact miscalculation was triggering compaction on our 1M sessions far too aggressively. Open-sourced it since clearly others are hitting it too.

Happy to hear if it helps — or if you find cases where the window override doesn't stick.

Reproducing on Linux (WSL2) with Claude Code CLI — Opus 4.6 shows 200k context window (visible via /context command). Anthropic docs confirm Opus 4.6 supports 1M. Would love to see this fixed.

claude 2.1.87 on WSL2 Opus-4.6 1M is shown but /context shows 1M
auto compaction kicks in at 200k and with disabled auto compaction then
on skill call at 200k errors "Context limit reached".

Same bug for me. Is there a fix for this?

Same here, v2.1.92 with max 200. claude-opus-4-6 36.3k/200k tokens (18%) => can't work like this

I set the autocompact limit to 900k with autocompact mode off, this fixed it for me.
But i saw the 1M context window at /context and everywhere else only that autocompact was at ~194k and limit reached at 240k event with the 1M displayed.

Same for me, just happened in recent update:

Environment

• Claude Code version: 2.1.98
• Model: claude-opus-4-6
• Plan: Max 20x
• Platform: Windows 11 Pro
• Shell: Git Bash

Same here. $200/m subscription, and every --continue on claude session invokes 1M -> 200k context degradation + immediately autocompacting. Very annoying