# Fix: DISABLE_TELEMETRY Blocks 1M Context and /remote-control in Claude Code

Binary patch for Claude Code that fixes the bug where setting
`DISABLE_TELEMETRY=1` silently disables ALL feature gates — blocking 1M context,
`/remote-control`, and every other gated feature — not just telemetry.
Tested on v2.1.76 and v2.1.77, Max 20x, macOS. Works on macOS, Linux, and Windows.

For a simpler workaround that only fixes 1M context (no patching required), see
[Simple Workaround](#simple-workaround-1m-context-only-no-patching) below.

After patching, gated features work while all data-collection env vars
(`DISABLE_TELEMETRY`, `DISABLE_ERROR_REPORTING`, `DISABLE_AUTOUPDATER`, etc.)
continue to work exactly as before.

**Related issues:**
[#29580](https://github.com/anthropics/claude-code/issues/29580) — source code trace proving the bug |
[#34083](https://github.com/anthropics/claude-code/issues/34083) — 1M context + /remote-control blocked on Max plan |
[#34143](https://github.com/anthropics/claude-code/issues/34143) — community-confirmed root cause analysis

## Step-by-Step Install (Full Binary Patch)

> **Prerequisites:** Python 3, Git.

### Step 1. Clone this gist

```bash
git clone https://gist.github.com/ahundt/158d498f07a10fec4955a70bf475a20b \
  ~/.claude/claude-patch-fix-disable-telemetry-blocks-1m-context-remote-control
```

### Step 2. Enter the directory

```bash
cd ~/.claude/claude-patch-fix-disable-telemetry-blocks-1m-context-remote-control
```

### Step 3. Dry-run first (changes nothing, shows what would be patched)

```bash
python3 patch_claude_binary.py
```

### Step 4. Apply the patches for real

```bash
python3 patch_claude_binary.py --real-run
```

A backup is created automatically next to the binary
(e.g., `~/.local/share/claude/versions/2.1.77.bak.20260317_160355`).

### Step 5. Verify — open a NEW Claude Code session and check:

- `/context` → should show **1000k** max (not 200k)
- `/model opus[1m]` → should succeed (not "not available for your account")
- `/help` → should list `/remote-control`

### Step 6. (Optional) Rollback if needed

```bash
python3 patch_claude_binary.py --rollback --real-run
```

Use `--rollback-list` to see all backups.

## Simple Workaround: 1M Context Only (No Patching)

If you only need 1M context (not `/remote-control` or other gated features), you can
bypass the client-side check with an env var — no binary patching needed.

**One-time launch:**

```bash
ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-4-6[1m] claude
```

**Make it permanent** — add to `~/.zshrc` (or `~/.bashrc`):

```bash
export ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-4-6[1m]
```

`/model opus[1m]` is rejected client-side (`s1mAccessCache`), but the server
accepts it on Max 20x after the March 13 GA. Env vars are passed to the provider
as-is, skipping that check. Banner confirms Opus 4.6 (1M context).

*Credit: [@ln-north](https://github.com/anthropics/claude-code/issues/34435#issuecomment-4064128346),
who tested on v2.1.76, Max 20x, macOS. Not independently verified by this repo's author.*

## After a Claude Code Auto-Update

When Claude Code updates itself, the new binary will be unpatched. Re-patch:

```bash
cd ~/.claude/claude-patch-fix-disable-telemetry-blocks-1m-context-remote-control
python3 patch_claude_binary.py --real-run
```

For unknown versions the script auto-discovers function names using stable structural
patterns. If auto-discovery also fails (e.g., code structure changed), it aborts with
a clear error. See [Adapting to Future Versions](#adapting-to-future-versions) below.

## Files in This Directory

| File | Purpose |
|------|---------|
| **`patch_claude_binary.py`** | **Apply/rollback patches.** Idempotent, dry-run default, auto-backup. Supports v2.1.76 and v2.1.77 via `KNOWN_VERSIONS`; auto-discovers function names for newer versions. Contains an investigation methodology in the docstring for manually tracing patches in any version. |
| **`analyze_claude_binary.py`** | **Investigate the binary.** 15 subcommands for version detection, function extraction, call chain tracing, patch status, traffic analysis, GrowthBook/Segment SDK deep-dives, and arbitrary pattern search. Run `python3 analyze_claude_binary.py all` for a full dump. |
| **`binary_patch_...bug.md`** | Original patch design document with before/after code, byte offsets, JS precedence analysis, and safety design. |
| **`README.md`** | This file — quick start, patch summary, and the full investigation report below. |

## What the Patches Do

Two patches are applied. All other functions are left intact so that data-collection
env vars continue to work normally. Minified function names change each release;
the table shows the role name and the per-version identifiers.

| Role | v2.1.76 name | v2.1.77 name | Patch | Effect |
|------|-------------|-------------|-------|--------|
| `statsig_gate` | `Ai()` | `hi()` | `return analytics_gate()` → `return!1\|\|!0` | Statsig/GrowthBook gate lookups always proceed; decouples feature gates from analytics state |
| `extra_usage_check` | `bf7()` | `Ij7()` | `return!1` → `return!0` for `undefined` branch | Fixes 1M race condition: `undefined` means "not yet loaded", not "disabled" |
| `telemetry_gate` | `Zv()` | `Cv()` | **not patched** | `DISABLE_TELEMETRY=1` continues to disable analytics events and block Segment |
| `analytics_gate` | `u1_()` | `o1_()` | **not patched** | Analytics follow env var: off when `DISABLE_TELEMETRY=1`, on otherwise |
| Segment init | `b27()` | varies | **not patched** | Segment follows env var: blocked when `DISABLE_TELEMETRY=1`, works otherwise |
| `feedback_gate` | `L8T()` | `v8T()` | **not patched** | Feedback surveys stay suppressed when `DISABLE_TELEMETRY=1` |

**Data-collection flag status (all env-controlled after patching):**

| Flag | Controls | After patching |
|------|----------|---------------|
| `DISABLE_TELEMETRY=1` | Analytics events, Segment SDK | Off/blocked (env-controlled, same as before patching) |
| `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1` | Same as above (umbrella) | Off/blocked (env-controlled) |
| `DISABLE_ERROR_REPORTING=1` | Sentry error logging | Off (env-controlled, patches don't touch this) |
| `CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY=1` | Feedback survey popup | Off (env-controlled, patches don't touch this) |
| `DISABLE_AUTOUPDATER=1` | Auto-updater | Off (env-controlled, patches don't touch this) |
| `DISABLE_COST_WARNINGS=1` | Cost warnings | Off (env-controlled, patches don't touch this) |

## Adapting to Future Versions

When Claude Code updates, minified function names change (e.g., `Zv`→`Cv`,
`u1_`→`o1_`). The script handles this in two ways:

**Automatic (try first):** For unknown versions, `discover_function_names()` searches
for stable structural patterns (unminified string literals that never change) and
discovers all function names automatically. If it succeeds, no manual work is needed.

**Manual (if auto-discovery fails):** This means the code structure changed, not just
the names. To find the new functions:

1. Add the new version's names to `KNOWN_VERSIONS` in `patch_claude_binary.py` and
   re-run. The `KNOWN_VERSIONS` dict maps version strings to role→name dicts.
2. To find those names, use `analyze_claude_binary.py search` and `context`
   subcommands with **stable strings** that don't get minified:
   - `!!process.env.DISABLE_TELEMETRY` — finds the telemetry gate function
   - `s_(process.env.CLAUDE_CODE_USE_FOUNDRY)` — finds the context anchor
   - `cachedExtraUsageDisabledReason` — finds the extra-usage race condition function
   - `LKJN8LsLERHEOXkw487o7qCTFOrGPimI` — finds Segment SDK init
   - `tengu_log_segment_events` — finds event gate names
   - `cachedGrowthBookFeatures` — finds local cache writer
   - `remoteEval:!0` — finds GrowthBook instance creation
3. Read the **Investigation Methodology** in `patch_claude_binary.py` (docstring)
   for a step-by-step process to trace the full call chain in any version.

---

## Known Limitations (inherent — cannot be patched away)

### Feature gate evaluation contacts Anthropic servers

The `Ai()` patch enables Statsig/GrowthBook feature gate evaluation. This requires
network contact with Anthropic's servers even when `DISABLE_TELEMETRY=1`:

- **GrowthBook remoteEval** — sends user metadata to `api.anthropic.com`:
  `deviceId`, `sessionId`, `platform`, `orgUUID`, `accountUUID`, `userType`,
  `subscriptionType`, `rateLimitTier`. No conversation content. Goes to the same
  server your conversations go to.
- **Statsig gate exposure events** — Statsig's SDK logs which feature gates were
  evaluated for your session. These are read-only flag lookups, not usage analytics,
  but they are transmitted to Anthropic's Statsig instance.

This contact **cannot be eliminated without also disabling feature gates entirely**
(which means losing 1M context, /remote-control, etc.). It is the irreducible minimum
required for gated features to work.

### DISABLE_TELEMETRY may be linked to training data permission in Anthropic's ToS

Anthropic's data usage policy links telemetry opt-out to training data permissions.
Enabling feature gate evaluation (by applying the `Ai()` patch) may constitute
"allowing telemetry" in the policy sense, which could be interpreted as granting
permission for data to be used for model training.

**What these patches protect:**
- Client-side analytics events (Statsig usage tracking, Segment) — **blocked** when
  `DISABLE_TELEMETRY=1`, via intact `u1_()` and `b27()` code paths ✓
- Sentry error reporting — **blocked** when `DISABLE_ERROR_REPORTING=1` ✓
- Feedback surveys — **suppressed** when `DISABLE_TELEMETRY=1` via intact `L8T()` ✓

**What these patches do NOT protect:**
- Feature gate evaluation network contact (GrowthBook remoteEval, Statsig gate exposure)
  — **enabled** by the `Ai()` patch. This was blocked before patching.
- Server-side data retention and training policies — governed by Anthropic's ToS, not
  by client binary code. These patches have no effect on what Anthropic does with data
  that reaches their servers.

**Recommendation:** Review [Anthropic's current data usage policy](https://code.claude.com/docs/en/data-usage)
and privacy settings in your account before applying. If the training data concern
outweighs the value of 1M context / /remote-control, do not apply the `Ai()` patch —
or roll it back with `python3 patch_claude_binary.py --rollback --real-run`.

---

## Patch Design Notes — Minimal vs Over-Engineered Approaches

### Why Minimal Patching Matters

The goal is to fix the specific structural bug (feature gates coupled to analytics
state) while leaving all other code paths — especially env-var-driven data collection
controls — completely intact. Every function left unpatched is one more thing that
correctly respects user configuration.

**Rule of thumb:** find the smallest cut in the dependency graph that separates the
two concerns. Do not patch downstream effects when you can cut the root coupling.

### The Over-Engineered Approach (5 patches — do not use)

An earlier version of this patch applied 5 functions: `Zv()` (telemetry_gate), `u1_()`
(analytics_gate), `Ai()` (statsig_gate), `b27()` (Segment init), and `bf7()`
(extra_usage_check) — using v2.1.76 names. The intent was correct but the scope was too wide.

**What went wrong:**

The call chain when `DISABLE_TELEMETRY=1` is (using v2.1.76 names; roles in brackets):

```
Zv() = true          [telemetry_gate] ← reads DISABLE_TELEMETRY
  → u1_() = !Zv() = false   [analytics_gate]  ← analytics off ✓
  → BKK() = !Zv() = false   → b27 returns null [Segment init] ← Segment blocked ✓
  → Ai() = u1_() = false    [statsig_gate] → Oq() bails ← ALL feature gates: default false ✗ BUG
```

The bug is only in the last link: `Ai()` calling `u1_()`. To fix it, only that link
needs cutting.

The over-engineered approach instead patched `Zv()` to always return `0` (ignoring
`DISABLE_TELEMETRY`), then had to also hardcode `u1_()=false` (otherwise analytics
re-enabled), and hardcode `b27=(0&&0)` (otherwise Segment re-initialized). This
created three hardcoded outcomes that no longer respected the env vars. A user who
later removed `DISABLE_TELEMETRY` from their settings would still have analytics
off and Segment blocked — the binary would ignore their changed preference.

### The Minimal Approach (2 patches — what this script applies)

Using v2.1.76 names; roles apply to all versions:

```
patch: Ai() [statsig_gate] = true   ← always (decouples gates from analytics)
patch: bf7() [extra_usage_check] undefined → true  ← fixes 1M race condition

Zv() [telemetry_gate], u1_() [analytics_gate], b27() [Segment init], L8T() [feedback_gate]
— all intact, all env-var driven
```

With `DISABLE_TELEMETRY=1`: `Zv()=true` → `u1_()=false` (analytics off), `BKK()=false`
(Segment blocked), `Ai()=true` (gates work). All flags respected.

Without `DISABLE_TELEMETRY`: `Zv()=false` → `u1_()=true` (analytics on), `BKK()=true`
(Segment works), `Ai()=true` (gates work). Removing the flag fully restores original behavior.

### How to Find the Minimal Patch in Future Versions

When a new Claude Code version changes minified names, follow this process:

**Step 1 — Identify the symptom precisely.**
Which feature is blocked? Which env var triggers it? In this case:
`DISABLE_TELEMETRY=1` → feature gates return `false` defaults.

**Step 2 — Find the feature gate evaluation function using stable strings.**
Feature gate queries always call a function that returns a boolean per gate name.
Search for stable patterns that won't be minified:
```bash
python3 analyze_claude_binary.py context 'cachedGrowthBookFeatures' --window 400
python3 analyze_claude_binary.py context 'remoteEval:!0' --window 400
```
Look for the function that bails out early: `if(!XYZ())return defaultValue`.
`XYZ()` is the new name of `Ai()`.

**Step 3 — Trace what XYZ() depends on.**
```bash
python3 analyze_claude_binary.py context 'function XYZ()' --window 100
```
It will return some other function, e.g. `return ABC()`. That's the new `u1_()`.

**Step 4 — Trace what ABC() depends on.**
```bash
python3 analyze_claude_binary.py context 'function ABC()' --window 100
```
It should return `!DEF()` where `DEF()` is the new `Zv()`.

**Step 5 — Confirm DEF() reads the env var.**
```bash
python3 analyze_claude_binary.py context 'function DEF()' --window 300
```
You should see `!!process.env.DISABLE_TELEMETRY` inside it.

**Step 6 — The minimal patch is always: patch XYZ() to return a constant true.**
```javascript
// Before:
function XYZ() { return ABC() }
// After (same byte length — adjust if names differ in length):
function XYZ() { return!1||!0 }
```
`!1||!0` = `false||true` = `true`. Same byte count as `return ABC()` when `ABC` is
a 3-character name. For 4-character names use `return!0` anchored differently, or
find a same-length constant expression. **Never change byte count — this corrupts the binary.**

**Step 7 — Verify the analytics chain is intact (not accidentally patched).**
After patching XYZ(), confirm:
- `DEF()` still contains `!!process.env.DISABLE_TELEMETRY`
- `ABC()` still returns `!DEF()`
- Analytics senders (search for `Segment`, `analytics.segment.com`) still check `ABC()` or `DEF()`

**Step 8 — Look for the race condition pattern in any "extra usage" gate.**
The `bf7()` pattern — `undefined` treated as "disabled" instead of "not yet loaded" — may
recur in future versions wherever async-loaded state is checked before the API responds.
Search for:
```bash
python3 analyze_claude_binary.py context 'cachedExtraUsageDisabledReason' --window 200
```
If the function returns `!1` (false) for `void 0` (undefined), apply the same 1-byte fix:
`return!1` → `return!0` for the `undefined` branch only.

**Summary of the methodology:**
1. Find the gate evaluation bailout function (XYZ / statsig_gate equivalent)
2. Confirm it depends on the analytics/telemetry state function
3. Add the new names to `KNOWN_VERSIONS` in `patch_claude_binary.py` and re-run
   (or let `discover_function_names()` auto-discover them if structure is unchanged)
4. Leave everything upstream (DEF/telemetry_gate, ABC/analytics_gate, Segment init) intact
5. Verify upstream chain still respects env vars

---

# Investigation Report: Claude Code 1M Context Window Bug

**Date:** 2026-03-15
**Claude Code Version:** 2.1.76 (latest)
**Plan:** Claude Max Account
**Issues:**
- `/model opus[1m]` returns "Opus 4.6 with 1M context is not available for your account"
- `/remote-control` is unavailable (same root cause)

---

## Root Cause: `DISABLE_TELEMETRY` Blocks All Feature Flags Including 1M Context

### The Bug

Setting `DISABLE_TELEMETRY=1` or `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1` in `~/.claude/settings.json` **disables all feature-gate evaluation**, not just telemetry. This means Claude Code cannot check whether the account is entitled to 1M context, `/remote-control`, or any other gated feature — so it defaults to `false` for all of them.

Note: `DISABLE_ERROR_REPORTING` is a separate setting (Sentry error logging) and is **not** present in the `Pv()` function based on the traced source code. It may not trigger this bug on its own, but it is commonly set alongside `DISABLE_TELEMETRY` and could potentially be related through a different code path not yet traced. Most users who reported the fix removed both variables simultaneously, so independent confirmation is limited.

### Source Code Evidence

User @Tuxerino420 traced the bug through the Claude Code source in [issue #29580, comment](https://github.com/anthropics/claude-code/issues/29580#issuecomment-4002014943).

**Impacted file and binary:**

The Claude Code CLI is distributed as a single compiled Bun executable. Platform-specific formats:
- macOS: `Mach-O 64-bit executable arm64` (or x86_64)
- Linux: `ELF 64-bit executable`
- Windows: `PE32+ executable (claude.exe)`

Default binary locations:
- macOS/Linux: `~/.local/bin/claude` (symlink → `~/.local/share/claude/versions/<version>`)
- Windows: `%LOCALAPPDATA%\AnthropicClaude\app-<version>\claude.exe`

The JavaScript source is bundled and minified inside this binary — there is no standalone `cli.js` file on disk. The npm package `@anthropic-ai/claude-code` exposes a single `bin: claude` entry point. All function names below (`Pv`, `qA`, `Ar`, `Zvq`, `vn`, `w1`) are minified identifiers from the embedded JS bundle and may change across versions.

@Tuxerino420's analysis was performed on Claude Code v2.1.69. The same code structure is present in v2.1.76 based on identical observed behavior.

**Call chain (traced from command registration to feature gate):**

```
Command registration: isEnabled: Zvq
  Zvq() → vn() → qA("tengu_ccr_bridge", false) → Ar() → !Pv()
```

**The bug is in the `Pv()` function** (determines whether to skip feature-flag evaluation):

```javascript
function Pv() {
  return w1(process.env.CLAUDE_CODE_USE_BEDROCK) ||
         w1(process.env.CLAUDE_CODE_USE_VERTEX) ||
         w1(process.env.CLAUDE_CODE_USE_FOUNDRY) ||
         !!process.env.DISABLE_TELEMETRY ||                          // <-- BUG: uses !! instead of w1()
         !!process.env.CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC      // <-- BUG: uses !! instead of w1()
}
```

**How it propagates:**
1. `Pv()` returns `true` when `DISABLE_TELEMETRY` is set to any non-empty string (including `"0"` or `"false"`)
2. `Ar()` calls `!Pv()`, which returns `false`
3. `qA()` sees `false` and **skips all GrowthBook/Statsig feature flag lookups**, returning the default value (`false`) for every gated feature
4. For `/remote-control`: the command registration uses `isEnabled: Zvq` — when `Zvq` returns `false`, the command is completely hidden from `/help`, not just disabled
5. For 1M context: the entitlement check defaults to `false`, so Claude Code rejects `/model opus[1m]`

**Features confirmed gated by this path:**
- 1M context window entitlement
- `/remote-control` command availability (gated by `tengu_ccr_bridge` flag)
- Any other feature-gated functionality

**Cached feature flags:** `~/.claude/.claude.json` — @Tuxerino420 confirmed this file correctly contained `"tengu_ccr_bridge": true` for the account, but the value was never read because `qA()` bailed out early due to `Pv()` returning `true`.

**The `!!` vs `w1()` sub-bug:** `DISABLE_TELEMETRY` uses `!!` (JavaScript truthy check) while `CLAUDE_CODE_USE_BEDROCK`, `CLAUDE_CODE_USE_VERTEX`, and `CLAUDE_CODE_USE_FOUNDRY` all use `w1()` (proper boolean parser) in the same function. Setting `DISABLE_TELEMETRY=0` in `settings.json` produces `process.env.DISABLE_TELEMETRY = "0"`, and `!!"0"` is `true` in JavaScript because `"0"` is a non-empty string. So explicitly setting telemetry to OFF (`"0"`) has the same effect as ON (`"1"`).

### Suggested Patch

From @Tuxerino420 in [#29580](https://github.com/anthropics/claude-code/issues/29580#issuecomment-4002014943):

**Minimal fix** — use the same `w1()` boolean parser already used for the other env vars in `Pv()`:

```javascript
// Before (broken):
!!process.env.DISABLE_TELEMETRY

// After (consistent with Bedrock/Vertex/Foundry checks):
w1(process.env.DISABLE_TELEMETRY)
```

Apply the same fix to `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC`:

```javascript
// Before (broken):
!!process.env.CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC

// After (consistent):
w1(process.env.CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC)
```

This would make `"0"` and `"false"` behave as expected (falsy), matching the behavior of `CLAUDE_CODE_USE_BEDROCK`, `CLAUDE_CODE_USE_VERTEX`, etc.

**Better fix** — decouple telemetry preferences from feature entitlements entirely. The `Pv()` function's purpose is to determine if the client should contact Statsig/GrowthBook, but it conflates "user doesn't want telemetry" with "user is on a third-party provider that can't reach Anthropic's feature-flag servers." These are fundamentally different situations. Telemetry opt-out should disable analytics data collection without affecting feature-gate evaluation for paid entitlements.

**Source:** All source code analysis, call chain trace, and patch suggestion from @Tuxerino420 in [#29580](https://github.com/anthropics/claude-code/issues/29580#issuecomment-4002014943) (Claude Code v2.1.69, Fedora Linux + Windows 11).

### The Proper Fix (Not Yet Applied)

Use the same `w1()` boolean parser already used for Bedrock/Vertex/Foundry env vars:

```javascript
// Before (broken):
!!process.env.DISABLE_TELEMETRY

// After (consistent):
w1(process.env.DISABLE_TELEMETRY)
```

More fundamentally: telemetry preferences should not gate functional features. These should be independent settings evaluated through separate code paths.

### Workaround: Binary Patch (This Repository)

**Update:** A binary patch is now available in this repository (`patch_claude_binary.py`) that fixes the feature-gate bypass while keeping `DISABLE_TELEMETRY=1` in settings. See the Quick Start at the top of this file.

**Without the binary patch:** There is no configuration-only workaround. The feature-gate lookup is completely bypassed — the telemetry setting and the entitlement check share the same code path. Every confirmed fix without patching involves **removing** `DISABLE_TELEMETRY` from settings. Note: adding `CLAUDE_CODE_ENABLE_TELEMETRY=1` does NOT fix it — the `DISABLE_TELEMETRY` key must be removed entirely.

### Why This Is a Bug, Not a Feature

1. The [official docs](https://code.claude.com/docs/en/data-usage) describe `DISABLE_TELEMETRY` as opting out of "analytics and usage tracking" — not opting out of paid features
2. The [GA announcement](https://claude.com/blog/1m-context-ga) says 1M context is "included with subscription" for Max plans with "no additional configuration"
3. The [model config docs](https://code.claude.com/docs/en/model-config) say "Opus is automatically upgraded to 1M context with no additional configuration" on Max plans
4. [Trail of Bits](https://github.com/trailofbits/claude-code-config) recommends disabling telemetry as a security best practice
5. Users should not have to choose between privacy and access to paid features

---

## Official Announcement

### 1M Context GA — March 13, 2026

**Link:** [1M context is now generally available for Opus 4.6 and Sonnet 4.6](https://claude.com/blog/1m-context-ga)

Key claims:

- Both Claude Opus 4.6 and Sonnet 4.6 now include the full 1M context window at standard pricing
- No beta header required — requests over 200K tokens work automatically
- Standard pricing across the full window ($5/$25 per million tokens for Opus 4.6, $3/$15 for Sonnet 4.6) — no multiplier
- 1M context is included in Claude Code for Max, Team, and Enterprise users with Opus 4.6
- Up to 600 images/PDF pages per request (up from 100)
- Available on Claude Platform, Microsoft Azure Foundry, and Google Cloud Vertex AI

---

## Official Documentation

### Model Configuration Docs

**Link:** [Model configuration - Claude Code Docs](https://code.claude.com/docs/en/model-config)

> On Max, Team, and Enterprise plans, Opus is automatically upgraded to 1M context with no additional configuration.

| Plan                      | Opus 4.6 with 1M context   | Sonnet 4.6 with 1M context  |
|---------------------------|----------------------------|------------------------------|
| Max, Team, Enterprise     | Included with subscription | Requires extra usage         |
| Pro                       | Requires extra usage       | Requires extra usage         |
| API and pay-as-you-go     | Full access                | Full access                  |

- To explicitly disable 1M context: set `CLAUDE_CODE_DISABLE_1M_CONTEXT=1`
- Use `[1m]` suffix with aliases: `/model opus[1m]`, `/model sonnet[1m]`
- "If you don't see it, try restarting your session"

### Data Usage / Telemetry Docs

**Link:** [Data usage - Claude Code Docs](https://code.claude.com/docs/en/data-usage)

- `DISABLE_TELEMETRY` opts out of Statsig analytics/usage tracking
- `DISABLE_ERROR_REPORTING` opts out of Sentry error logging
- `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC` is an umbrella that disables all non-essential traffic (also disables auto-updates per [Trail of Bits](https://github.com/trailofbits/claude-code-config))

Common privacy-focused configuration ([recommended by Trail of Bits](https://github.com/trailofbits/claude-code-config) and [privacy guides](https://www.vincentschmalbach.com/configuring-claude-code-for-privacy-and-noise-control/)) that triggers this bug:

```json
{
  "env": {
    "DISABLE_TELEMETRY": "1",
    "DISABLE_ERROR_REPORTING": "1",
    "CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY": "1"
  }
}
```

Note: `DISABLE_NON_ESSENTIAL_MODEL_CALLS` does **not exist** in v2.1.76 (0 occurrences in the binary) — it has no effect and should be removed from configs that reference it.

Of these flags, `DISABLE_TELEMETRY` is confirmed to trigger the feature-gate bypass. `CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY` is not in the `Pv()` function and should be safe. `DISABLE_ERROR_REPORTING` is not confirmed in `Pv()` but may be related through an untraced code path.

The [Trail of Bits config](https://github.com/trailofbits/claude-code-config) also recommends the equivalent umbrella setting `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC`, which is confirmed to trigger the same bug via the `Pv()` function.

---

## Claude Code Release Notes

### v2.1.75 — March 13, 2026

**Link:** [Release v2.1.75](https://github.com/anthropics/claude-code/releases/tag/v2.1.75)

- Opus 4.6 uses a 1M context window by default on Max, Team, and Enterprise plans
- Tool permission denials prompt for a reason when intent is unclear
- Memory files show last-modified timestamps
- Bug fixes: voice mode, header model name display, session crashes, Bash `!` mangling

### v2.1.76 — March 2026

**Link:** [Release v2.1.76](https://github.com/anthropics/claude-code/releases/tag/v2.1.76)

- Added `/effort` command (low/medium/high/max)
- Opus 4.6 defaults to medium effort for Max and Team
- "ultrathink" keyword re-introduced for high effort
- Opus 4 and 4.1 removed — users auto-moved to Opus 4.6
- Default Opus model on Bedrock/Vertex/Foundry changed to Opus 4.6
- Fixed: spurious "Context limit reached" errors on 1M sessions
- Fixed: "adaptive thinking is not supported" errors with non-standard model strings
- Fixed: auto-compaction retrying indefinitely (circuit breaker after 3 attempts)

### v2.1.77 — March 2026

**Link:** [Release v2.1.77](https://github.com/anthropics/claude-code/releases/tag/v2.1.77)

- Minified function names changed (telemetry_gate: `Zv`→`Cv`, analytics_gate: `u1_`→`o1_`,
  statsig_gate: `Ai`→`hi`, extra_usage_check: `bf7`→`Ij7`, feedback_gate: `L8T`→`v8T`)
- This patch script supports v2.1.77 via `KNOWN_VERSIONS["2.1.77"]`

---

## User Circumstances and Troubleshooting Performed

### Setup

- **Plan:** Claude Max Account (confirmed via `/status`)
- **Claude Code version:** 2.1.76 (confirmed latest via `claude --version`)
- **Model:** Default Opus 4.6 (confirmed via `/status` — "Default Opus 4.6 - Most capable for complex work")
- **Login method:** Claude Max Account (confirmed via `/status`)
- **Platform:** macOS (Darwin 25.2.0)
- **Shell:** zsh
- **API provider:** First-party Anthropic (not Bedrock, Vertex, or Foundry)
- **Setting sources:** User settings, Project local settings

### Settings Present in Config

`~/.claude/settings.json` contains:

```json
{
  "env": {
    "DISABLE_ERROR_REPORTING": "1",
    "DISABLE_TELEMETRY": "1"
  }
}
```

`DISABLE_TELEMETRY` is the root cause. `DISABLE_ERROR_REPORTING` is not in `Pv()` and does not trigger this bug independently.

No `availableModels` restrictions, no model overrides, no context-related settings in any settings file (`~/.claude/settings.json`, `~/.claude/settings.local.json`, `.claude/settings.local.json`).

### Observed Behavior

1. `/context` displays `claude-opus-4-6 . 23k/200k tokens (12%)` — capped at 200k
2. `/model opus[1m]` returns: **"Opus 4.6 with 1M context is not available for your account"**
3. `/remote-control` is unavailable — confirmed same root cause per [#29580](https://github.com/anthropics/claude-code/issues/29580) and [#34083](https://github.com/anthropics/claude-code/issues/34083)

### What Was Tried

1. **Checked Claude Code version** — confirmed 2.1.76 (latest), includes 1M support added in v2.1.75
2. **Switched model to Opus via `/model default`** — Opus 4.6 selected but `/context` still showed 200k
3. **Attempted `/model opus[1m]`** — rejected with error above
4. **Checked all context-related environment variables** — `CLAUDE_CODE_DISABLE_1M_CONTEXT`, `ANTHROPIC_DEFAULT_OPUS_MODEL`, `ANTHROPIC_DEFAULT_SONNET_MODEL`, `ANTHROPIC_MODEL`, `CLAUDE_CODE_MAX_CONTEXT`, `DISABLE_PROMPT_CACHING`, all `DISABLE_PROMPT_CACHING_*` — all unset
5. **Inspected all settings files** — no `availableModels` restrictions, no model overrides in any of the 3 settings files
6. **Verified no third-party provider config** — no Bedrock/Vertex/Foundry env vars set
7. **Searched web and GitHub** — found root cause documented across issues #34143, #34083, #29580
8. **Read official docs** — confirmed Max plans should get Opus 1M automatically; docs do not mention `DISABLE_TELEMETRY` blocks this
9. **Identified root cause** — `DISABLE_TELEMETRY` causes `Pv()` in `cli.js` to bypass all feature-gate evaluation

---

## GitHub Issues

### Telemetry / Feature-Gate Bug (Root Cause)

| Issue | Title | Date |
|-------|-------|------|
| [#29580](https://github.com/anthropics/claude-code/issues/29580) | DISABLE_TELEMETRY=1 breaks remote-control with misleading "not yet enabled" error (includes source code trace) | Feb 28, 2026 |
| [#34083](https://github.com/anthropics/claude-code/issues/34083) | Max plan: 1M context window shows 200k, Remote Control unavailable | Mar 13, 2026 |
| [#34143](https://github.com/anthropics/claude-code/issues/34143) | Opus 4.6 shows 200K context instead of 1M on Claude Max plan (v2.1.75) | Mar 13, 2026 |
| [#34435](https://github.com/anthropics/claude-code/issues/34435) | Opus 4.6 default model alias not using 1M context window (v2.1.76, Max) | Mar 14, 2026 |
| [#33119](https://github.com/anthropics/claude-code/issues/33119) | Remote Control not enabled on Max plan account | Mar 2026 |

### Early Opus 4.6 / 1M Context Bugs (Feb 2026, pre-GA)

| Issue | Title | Date |
|-------|-------|------|
| [#23432](https://github.com/anthropics/claude-code/issues/23432) | /context command reports 200K max for Opus 4.6 instead of 1M | Feb 5, 2026 |
| [#23472](https://github.com/anthropics/claude-code/issues/23472) | Opus 4.6 with [1m] returns "long context beta not available" on Max subscription | Feb 2026 |
| [#23700](https://github.com/anthropics/claude-code/issues/23700) | Long context beta unavailable for Max plan with Opus 4.6 1M | Feb 2026 |
| [#23714](https://github.com/anthropics/claude-code/issues/23714) | Opus 4.6 implies 1M but hard-caps at 200K — session crashes (Max 20x) | Feb 2026 |
| [#23879](https://github.com/anthropics/claude-code/issues/23879) | Bring 1M context window to Max plan subscribers | Feb 2026 |
| [#24208](https://github.com/anthropics/claude-code/issues/24208) | Opus 4.6 context display shows 200k instead of 1M | Feb 2026 |

### Regressions and Billing Issues (Feb-Mar 2026)

| Issue | Title | Date |
|-------|-------|------|
| [#26428](https://github.com/anthropics/claude-code/issues/26428) | Sonnet 1M disappeared after update to v2.1.45 — regression on Max ($200/mo) | Feb 2026 |
| [#28723](https://github.com/anthropics/claude-code/issues/28723) | Unclear how 1M context billing works on Max subscription | Mar 2026 |
| [#28927](https://github.com/anthropics/claude-code/issues/28927) | Silent billing change in v2.1.51: 1M moved to extra-usage-only without notice | Mar 2026 |
| [#28975](https://github.com/anthropics/claude-code/issues/28975) | Opus 1M hits "Rate limit reached" on Max and Team plans | Mar 2026 |
| [#29330](https://github.com/anthropics/claude-code/issues/29330) | Opus 1M suddenly returns "Rate limit reached" on Max and Team (200K works fine) | Mar 2026 |
| [#33154](https://github.com/anthropics/claude-code/issues/33154) | Cowork forces [1m] causing instant rate limit errors on Max | Mar 11, 2026 |
| [#33584](https://github.com/anthropics/claude-code/issues/33584) | Feature request: mid-tier context (400K) at standard Max rates | Mar 12, 2026 |

### Related Telemetry Issues

| Issue | Title | Date |
|-------|-------|------|
| [#19117](https://github.com/anthropics/claude-code/issues/19117) | Telemetry Configuration Ambiguity: Confusion between Statsig metrics and OpenTelemetry | 2026 |
| [#10494](https://github.com/anthropics/claude-code/issues/10494) | DISABLE_TELEMETRY flag ignored — Claude Code connects to Google despite opt-out | 2025 |
| [#5508](https://github.com/anthropics/claude-code/issues/5508) | OpenTelemetry telemetry cannot be disabled on Windows | 2025 |

### Original Feature Request

| Issue | Title | Date |
|-------|-------|------|
| [#5644](https://github.com/anthropics/claude-code/issues/5644) | Feature Request: Enable 1M Context Window for Claude Code | Aug 2025 |

### Bedrock-Specific

| Issue | Title | Date |
|-------|-------|------|
| [#32673](https://github.com/anthropics/claude-code/issues/32673) | Bedrock: 1M context should auto-enable for capable models | Mar 2026 |

---

## Key User Reports from Issue Comments

### Telemetry as root cause (confirmed)

From [#34143](https://github.com/anthropics/claude-code/issues/34143):
- **@Hoernchen:** First identified the root cause — `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC` or `DISABLE_TELEMETRY=1` blocks 1M context (6 thumbs up).
- **@davidlandais:** Confirmed the fix. "It's a shame you have to disable this option to get access to 1M context. Not very honest."
- **@ZenAlexa:** Confirmed on Windows 11 + v2.1.76 + Max plan. "Users shouldn't have to opt into telemetry to access a paid feature."

From [#34083](https://github.com/anthropics/claude-code/issues/34083):
- **@lopi-py:** Enabling telemetry fixed both 1M context and `/remote-control`.
- **@cwilso03:** Confirmed removing `DISABLE_TELEMETRY` from settings.json immediately restored 1M. Adding it back removed 1M again. Provided the exact block to look for:
  ```json
  "env": {
    "DISABLE_TELEMETRY": "1"
  }
  ```
- **@certiv-dwager:** Speculates staged rollout — "they want to watch their metrics closely. so rolling out to telemetry enabled users first."
- **@inevity:** Setting `CLAUDE_CODE_ENABLE_TELEMETRY=1` did NOT fix it — the `DISABLE_TELEMETRY` key must be **removed** entirely.

From [#29580](https://github.com/anthropics/claude-code/issues/29580):
- **@Tuxerino420:** Full source code trace of `Pv()` → `qA()` → `Ar()`. Discovered `DISABLE_TELEMETRY=0` also triggers the bug (`!!"0"` is `true`). The `/remote-control` command was completely hidden from `/help`, not just failing.
- **@jschneider:** First reported the `/remote-control` variant. Confirmed only `DISABLE_TELEMETRY` was needed to trigger — `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC` was NOT set.

### Additional factors (telemetry may not be the only cause)

From [#34143](https://github.com/anthropics/claude-code/issues/34143):
- **@AspireAI-Developer:** On Max plan, v2.1.76, macOS — verified none of the known env vars are set, still capped at 200K. Suggests additional factors beyond telemetry.
- **@ManufactoryOfCode:** Workaround for non-telemetry cases — switch to Haiku in `/model`, then re-open `/model` and Opus 4.6 (1M) may appear as a separate option.
- **@chrisgagne:** Used 1M with extra usage disabled, but after several minutes the 1M options disappeared from `/model`.

From [#34083](https://github.com/anthropics/claude-code/issues/34083):
- **@stripped-down:** Same account, same Max plan, same v2.1.76 — Linux VM gets 1M immediately, macOS still shows 200K (telemetry enabled on both). Suggests platform-gated rollout may also be a factor.
- **@Nantris:** "It's just random. I use it across several machines with identical settings and some have 1M and some don't."
- **@kexgev:** Had 1M yesterday, Max plan expired and re-purchased, now shows 200K again.

### Pre-GA context and performance notes

From [#23432](https://github.com/anthropics/claude-code/issues/23432):
- **@nerdpudding:** Got 1M working on Max with extra usage enabled, but performance degraded past 250K tokens — slower responses and more errors past 350K. Noted the "Billed as extra usage" warning when switching to `opus[1m]`.
- **@Notum:** Before GA, 1M beta was only for API users with Tier 3+ plans, not Max subscribers — despite Max marketing listing 1M as a feature.
- **@felipekj:** Max 20x subscriber, extra usage enabled, v2.1.49 — still got "The long context beta is not yet available for this subscription."
- **@peixotorms:** Had 1M for several days on Max 20x, then it reverted to 200K without any config change.
- **@OZmasterAI:** Provided a statusline workaround script to recalculate context percentage from actual token counts when the reported window is wrong.

---

## Recommended Actions

1. **Comment on [#34083](https://github.com/anthropics/claude-code/issues/34083)** (oldest post-GA issue, March 13) documenting this case and the root cause

2. **Request Anthropic decouple telemetry from feature gates** — the fix is straightforward (use `w1()` instead of `!!` in `Pv()`, or better, separate the code paths entirely)

3. **Do NOT remove `DISABLE_TELEMETRY`** — telemetry opt-out is a legitimate privacy choice and should not be required for paid features

4. **Try re-authenticating** as a long shot (won't fix the telemetry root cause but rules out stale auth):
   ```bash
   claude logout
   claude login
   ```

5. **Monitor these issues for an official fix:**
   - [#34083](https://github.com/anthropics/claude-code/issues/34083) — Primary post-GA issue (1M + remote-control)
   - [#34143](https://github.com/anthropics/claude-code/issues/34143) — Most community-confirmed root cause analysis
   - [#29580](https://github.com/anthropics/claude-code/issues/29580) — Source code trace proving the bug

---

## References

### Official

- [1M context GA announcement (March 13, 2026)](https://claude.com/blog/1m-context-ga)
- [Model configuration docs](https://code.claude.com/docs/en/model-config)
- [Data usage / telemetry docs](https://code.claude.com/docs/en/data-usage)
- [Claude Code v2.1.75 release](https://github.com/anthropics/claude-code/releases/tag/v2.1.75)
- [Claude Code v2.1.76 release](https://github.com/anthropics/claude-code/releases/tag/v2.1.76)
- [What's new in Claude 4.6](https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-6)
- [Context windows docs](https://platform.claude.com/docs/en/build-with-claude/context-windows)

### Community / Third-Party

- [Trail of Bits claude-code-config](https://github.com/trailofbits/claude-code-config) — Security-focused configuration recommendations
- [Configuring Claude Code for Privacy](https://www.vincentschmalbach.com/configuring-claude-code-for-privacy-and-noise-control/) — Privacy configuration guide