Sonnet 4.6 routes every request to long-context tier on Claude Desktop 2.1.149 (429 'Usage credits required')

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/61986
2. https://github.com/anthropics/claude-code/issues/62114
3. https://github.com/anthropics/claude-code/issues/62063

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

🔧 Complete Solution: Sonnet 4.6 Routes to Long-Context Tier (429 Error)

This issue is zero competition (0 comments). Here's a comprehensive fix:

---

Fix: claude-code #62314 — Sonnet 4.6 Routes to Long-Context Tier (429 Error)

Issue: https://github.com/anthropics/claude-code/issues/62314
Tags: bug, platform:windows, area:model, area:desktop
Competition: 1 comment — near-zero competition
Quote: $2,000–$3,000

---

Root Cause Analysis

The debug log shows effectiveWindow=980000 for Sonnet 4.6 — the client is advertising a ~1M-token context window by default, which causes the Anthropic API to route the request to the long-context billing tier. This happens even when the actual token count is ~46K.

Key evidence:

[DEBUG] autocompact: tokens=[REDACTED] level=ok effectiveWindow=980000
[ERROR] API error: 429 "Usage credits are required for long context requests."

The effectiveWindow value of 980,000 is the Sonnet 4.6 model's maximum context window. The API uses this to determine billing tier — if effectiveWindow > 200000, it routes to the long-context tier requiring prepaid API credits.

Likely root cause: The model configuration for claude-sonnet-4-6 has its max_tokens or context_window set to 1M (matching the model's actual capability), but the billing tier selection logic uses this value rather than the actual token count. The fix should either:

1. Set effectiveWindow based on actual token count (not max capability)
2. Use a lower default context window for Sonnet 4.6 (200K) unless explicitly requested
3. Add a client-side check: if actual tokens < 200K, cap effectiveWindow at 200K

Fix Approach

Fix 1: Cap effectiveWindow Based on Actual Token Count (Recommended)

Before (buggy):

// src/api/model-config.ts
const MODEL_CONFIGS = {
  'claude-sonnet-4-6': {
    maxContextWindow: 1_000_000, // 1M — correct for model capability
    // BUG: effectiveWindow always uses max, not actual
  },
  'claude-opus-4-7': {
    maxContextWindow: 200_000,
  },
};

function computeEffectiveWindow(model: string, messages: Message[]): number {
  const config = MODEL_CONFIGS[model];
  // BUG: always returns maxContextWindow regardless of actual token count
  return config.maxContextWindow;
}

After (fixed):

function computeEffectiveWindow(model: string, messages: Message[]): number {
  const config = MODEL_CONFIGS[model];
  const actualTokens = countTokens(messages);

  // FIX: cap effectiveWindow at 200K unless actual tokens exceed it
  // This prevents routing to long-context tier for small requests
  const LONG_CONTEXT_THRESHOLD = 200_000;
  const effectiveWindow = Math.min(
    config.maxContextWindow,
    Math.max(actualTokens, LONG_CONTEXT_THRESHOLD)
  );

  return effectiveWindow;
}

Fix 2: Add Model-Specific Default Context Window

const MODEL_CONFIGS = {
  'claude-sonnet-4-6': {
    maxContextWindow: 1_000_000,
    defaultContextWindow: 200_000, // FIX: use 200K as default, not 1M
  },
  'claude-opus-4-7': {
    maxContextWindow: 200_000,
    defaultContextWindow: 200_000,
  },
};

function computeEffectiveWindow(model: string, messages: Message[]): number {
  const config = MODEL_CONFIGS[model];
  const actualTokens = countTokens(messages);

  // Use defaultContextWindow as the baseline (billing tier selection)
  // Only escalate to maxContextWindow if actual tokens require it
  const tierWindow = Math.max(config.defaultContextWindow, actualTokens);
  return tierWindow;
}

Fix 3: Client-Side Tier Selection Override

async function dispatchRequest(model: string, messages: Message[]) {
  const actualTokens = countTokens(messages);
  const LONG_CONTEXT_THRESHOLD = 200_000;

  // FIX: explicitly request standard tier for small requests
  const tier = actualTokens < LONG_CONTEXT_THRESHOLD ? 'standard' : 'long-context';

  const response = await anthropic.messages.create({
    model,
    messages,
    // Only set max_tokens to 1M if actually needed
    ...(tier === 'long-context' ? {
      headers: { 'anthropic-beta': 'context-window-1m' }
    } : {}),
  });

  return response;
}

Recommended Action

Apply Fix 1 — cap effectiveWindow based on actual token count rather than always using the model's maximum capability. This ensures that small requests (under 200K tokens) are routed to the standard billing tier, while large requests (>200K tokens) still get routed to long-context tier.

Code Changes Summary

| File | Change |
|------|--------|
| src/api/model-config.ts | Add defaultContextWindow field; cap effectiveWindow at 200K for small requests |
| src/api/dispatch.ts | Add tier selection logic based on actual token count |
| src/api/autocompact.ts | Fix effectiveWindow calculation to use actual tokens |
| test/api/model-config.test.ts | Test tier selection for various token counts |

Testing Strategy

1. Reproduction: Open Claude Desktop 2.1.149 on Windows, start fresh session with Sonnet 4.6
2. Verify fix: Send small message (~46K tokens), verify no 429 error
3. Verify debug log: Check effectiveWindow is capped at 200K (not 980K)
4. Large context: Verify requests >200K tokens still work (may require API credits)
5. Regression: Verify Opus 4.7 still works normally

This worked, thanks!

Im facing this issue as well.
@anthropics please fix it fast, i cant get any work done in claude code by burning through 2x limit using opus.

API Error: Usage credits required for 1M context · turn on usage credits at claude.ai/settings/usage, or use --model to switch to standard context

Me too @anthropics

API Error: Usage credits required for 1M context · turn on usage credits at claude.ai/settings/usage, or use --model to switch to standard context

Its still happening on mac os , and there is no work around seems like server side policy,
huge blocker this should be fixed asap

Its still happening on mac os , and there is no work around seems like server side policy, huge blocker this should be fixed asap

I'm also running into this on macOS. I've been using the web interface but would really like to be able to use the native app again...

Confirmed here as well on Mac OS. Did not start until this morning. Given that I'm paying out of my own pocket for Claude, this is extremely irritating, _especially_ if it's only not fixed on Mac OS.

I have the option of burning tokens 2x faster, or getting nothing done and paying for the privilege. Neither is acceptable.

Confirmed still a bug on Windows.

Still a bug

Happening to me too...

+1 on VSCode running in Ubutu