[Bug] /context command reports 200K max for Claude Opus 4.6 instead of 1M

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/11281
2. https://github.com/anthropics/claude-code/issues/8678
3. https://github.com/anthropics/claude-code/issues/12006

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

I can confirm this bug

I confirm this bug. And it should not be closed considering latest Claude Opus 4.6 extended release, as it is not a duplicate issue

I confirm this bug. And it should not be closed considering latest Claude Opus 4.6 extended release, as it is not a duplicate issue

Also confirm and it seems to compact and <200K tokens as well.

_1M token context (beta). Opus 4.6 is our first Opus-class model with 1M token context. Premium pricing applies for prompts exceeding 200k tokens ($10/$37.50 per million input/output tokens)._
No 1m context windows for $100 & $200 MAX plans - only for API users

New convo started on 2/6 after code reboot:

<img width="280" height="319" alt="Image" src="https://github.com/user-attachments/assets/fb9fd983-e23d-4a17-8116-8b606e75acc4" />

Findings on Opus 4.6 1M context window availability in Claude Code:

1. Using /model opus[1m] in Claude Code does switch the model — /context correctly shows claude-opus-4-6[1m] · 41k/1000k tokens
2. However, the first API call immediately fails with: API Error: 400 - "The long context beta is not yet available for this subscription."
3. The session then falls back to claude-opus-4-6 · 41k/200k tokens

This was tested on a Max subscription (not API pay-as-you-go).

The https://platform.claude.com/docs/en/build-with-claude/context-windows#1m-token-context-window state the 1M beta requires "organizations in usage tier 4", but the https://claude.ai list 1M context as a feature of Max
$100/$200 plans. These are contradictory.

TL;DR: Claude Code correctly supports the [1m] model suffix, but the API rejects 1M requests from Max subscription users. Either the docs/marketing need to be updated to clarify this limitation, or 1M access should be
enabled for Max subscribers as advertised.

I just noticed this (stil on max subscription) when near compact: _'Context left until auto-compact: 3% · /model opus[1m] for more context · Billed as extra usage'_ so I tried and it works, but note the 'billed as extra usage' though...

/context
⎿ Context Usage
⛀ ⛁ ⛁ ⛁ ⛁ ⛀ ⛀ ⛀ ⛀ ⛁ ⛁ ⛁ ⛁ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ claude-opus-4-6[1m] · 40k/1000k tokens
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ (4%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ Estimated usage by category
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ System prompt: 2.6k tokens (0.3%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ System tools: 18.6k tokens (1.9%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ MCP tools: 295 tokens (0.0%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Custom agents: 1.5k tokens (0.1%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Memory files: 993 tokens (0.1%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Skills: 79 tokens (0.0%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛁ Messages: 19.1k tokens (1.9%)
⛶ Free space: 924k (92.4%)
⛝ Autocompact buffer: 33k tokens
(3.3%)

FYI, my first experiences with using the 1m model in current experimental stage: becomes increasingly slow (or there are other issues going on by coincidence, but it seems to be the 1m, when reaching more then 250k tokens used in 1 session) and also it seems to show mistakes more apparent at around 350k. So for now, i will stick with 200k standard and annoying but seemingly better working method of compacting around 190k tokens.

+1 — Max 20x subscriber, extra usage enabled, Claude Code 2.1.49, same issue: 'The long context beta is not yet available for this subscription.'

+1 — Max 20x subscriber, extra usage enabled, Claude Code 2.1.49, same issue: 'The long context beta is not yet available for this subscription.'

I think you need to enable 'extra usage' first

1M Beta Context is available ONLY for API user with Tier 3 PLAN!

I am using Max and have the 1m context since 2 days. I also have API, but that is not what I am using it with right now. Or rather was using it with. After 2 days, in current state of things, I'ld rather use the 200k limit and compact it turns out.

for the past few days, on max 20x I have seen the 1000k context.
now suddenly, it went back to 200k... why?

for the past few days, on max 20x I have seen the 1000k context. now suddenly, it went back to 200k... why?

I think it was during a test period when that happened, and now they’re back to the usual way. But it was a good experience for me, without the headache of context windows and finishing the conversation compactly.

Workaround for statusline hooks

If you have a custom statusline script, you can recalculate used_percentage from the actual token counts in current_usage instead of relying on the pre-calculated value (which is based on the wrong 200k window).

Root cause: In the binary, the DY() function correctly returns 1e6 when the context-1m-2025-08-07 beta is active, but lw() (which provides the betas list to the statusline payload builder) doesn't include it — so the statusline path falls through to the 200k default.

Fix: Add a config flag and recalculate from token counts:

# In your statusline script, after reading ctx_data:
context_pct = ctx_data.get("used_percentage", 0) or 0

# Recalculate % from actual tokens if on 1M plan
CONTEXT_WINDOW_OVERRIDE = 1000000  # set to 0 to disable
reported_window = ctx_data.get("context_window_size", 200000)
if CONTEXT_WINDOW_OVERRIDE and CONTEXT_WINDOW_OVERRIDE != reported_window:
    cur = ctx_data.get("current_usage") or {}
    actual_tokens = (
        (cur.get("cache_read_input_tokens") or 0)
        + (cur.get("cache_creation_input_tokens") or 0)
        + (cur.get("input_tokens") or 0)
    )
    context_pct = actual_tokens / CONTEXT_WINDOW_OVERRIDE * 100

This computes the real percentage from cache_read_input_tokens + cache_creation_input_tokens + input_tokens divided by the actual 1M window. Verified to match /context output exactly.

Why not just scale used_percentage * (200k / 1M)? Because if used_percentage is capped at 100% internally (once usage exceeds 200k tokens), the scaled value would max out at 20% and your threshold would never fire.

For me, I had CLAUDE_CODE_BLOCKING_LIMIT_OVERRIDE=200000 set in my env. Once I removed it and restarted my Opus 4.6 1M session it essentially lifted the context block.

I can switch model for opus 1m tokens (opus[1m] or eu.anthropic.claude-opus-4-6-v1[1m]) in console but even like this it have anyway same bugs (I can switch by command /model opus[1m] and it works but I do not see it in models list).
but for extension it doesn't work absolutely it looks like it ignores [1m] suffix

+1 — Max plan subscriber ($200/month), Opus 4.6 on macOS Darwin 25.3.0, Claude Code v2.1.96. Still showing 200k cap in /context as of 2026-04-08. This also causes premature autocompaction, which is especially painful when sessions crash and need to be restored from JSONL history — context gets compacted way earlier than it should.

Boooo...super unhappy, derailed me :(

Duplicate from #52537 — adding root-cause analysis from the minified source.

The bug is in the context window detection function (zG in cli.js):

function WQK(A) { return A.toLowerCase().includes("claude-sonnet-4") }

function zG(A, K) {
  if (A.includes("[1m]") || K?.includes(FVA) && WQK(A))
    return 1e6;  // 1,000,000
  return _QK;    // 200,000
}

The 1M path is only reachable for:

1. Models with [1m] suffix (e.g., sonnet[1m])
2. Sonnet-4 models with the context-1m-2025-08-07 beta header

WQK explicitly checks for "claude-sonnet-4" — Opus models always fall through to the 200K default. The fix should either add an Opus check to WQK or default Opus to 1M directly.

Impact beyond display: autoCompactWindow and likely max_input_tokens are both derived from this function, so Opus users are effectively capped at 200K of their 1M window.