You've hit your session limit · resets 3:50pm (America/Los_Angeles)

Bug Description
I have no idea what just happened but 100% of 3 different claude accounts (2x $200 MAX plans and 1x $100 Enterprise plan) just got absolutely WRECKED within 45 minutes after the Fable 5.1 update and version 2.1.257 - This is the same workflow I've had in my suite of work for the past two weeks while refactoring a very large infrastructure project and I don't even understand. I've checked my config and my ultracode / workflows limit was set to medium as always and supposed to be <15.

Please find below a summary baked from ChatGPT (since I have zero available usage in Claude now):

Environment

• Date observed: 2026-09-01
• Platform: macOS on Apple Silicon
• Claude Code versions before incident: 2.1.251 and 2.1.252
• Incident version: 2.1.257
• Authentication: Claude subscriptions, not API-key billing
• Accounts affected: two Max 20× subscriptions and one premium organizational 5× subscription
• Effective incident model: claude-fable-5-1
• Effort: xhigh
• Workload: established AGENTS-based, multi-agent repository workflows
• Repository, account, prompt, task, and session content are intentionally omitted

Summary

Immediately after the active Claude Code process moved to version 2.1.257, an established multi-agent workflow entered a substantially different usage regime.

Within approximately 42 minutes:

• 1,659 successful model requests were recorded.
• 282.3 million processed tokens were recorded, predominantly cached context.
• Approximately 21.4 million cache-write tokens were generated.
• More than 96% of processed tokens came from subagents.
• Three subscription accounts appeared exhausted or unusable after sequential account switching.
• The server returned 43 distinct quota-related HTTP 429 responses.

A comparable earlier fan-out period using the pre-update clients ran at approximately $2.31/minute in API-list-price-equivalent usage. The post-update interval ran at approximately $8.97/minute: a 3.88× increase, composed of approximately 1.80× more requests per minute and 2.15× more equivalent cost per request.

API-list-price equivalents are used only to compare workload shape. They do not represent subscription charges or reproduce Anthropic’s undisclosed subscription-quota calculation.

Expected behavior

Updating Claude Code and resuming the same class of workflow should not silently produce a several-fold usage-rate increase, materially expand runtime fan-out, or exhaust several subscription entitlements in minutes.

A documented one-time cache rebuild after an update may be expected. It should not repeatedly rebuild large contexts across an entire subagent fleet without clear warning or visibility.

Actual behavior

After 2.1.257 became active:

1. A resumed long-running session routed its first response to Fable 5.1 at xhigh, without a --model command-line override.
2. Diagnostics recorded a model_changed cache miss involving approximately 630,000 tokens.
3. Multiple established workflow definitions were started.
4. Actual runtime fan-out and cache writes increased substantially.
5. Thirty-six agent executions failed.
6. Sequential account switching was followed by reported exhaustion across all three accounts.
7. Server responses changed from successful inference to quota and organization-level blocking errors.

The workflow definitions were not intentionally changed to increase scope or concurrency. However, the actual post-update execution did expand: more workers ran, and cache writes per worker increased approximately 2.5×. Determining why the runtime behavior changed is part of this report.

Timeline

| Time, PDT | Observation |
|---|---|
| 11:18 | Claude Code 2.1.257 installed |
| 11:19 | Client recorded a Fable-to-Fable-alias migration timestamp |
| 13:46 | Final response from the older active client; model was Opus 4.8 |
| 13:50 | New 2.1.257 process started by resuming the existing session |
| 13:55 | First 2.1.257 response; model was Fable 5.1 at xhigh |
| 13:59 | First large post-update workflow wave began |
| 14:08 | Additional workflow wave began |
| 14:11 | Settings file was rewritten with Fable 5.1 and xhigh; Fable 5.1 had already been used before this rewrite |
| 14:34 | Another workflow wave began |
| 14:37 | First quota-related 429 response |
| 14:42 | Final recorded 429 response |
| 15:21 | /feedback report submitted |

Controlled usage comparison

Comparable fan-out intervals

| Metric | Earlier fan-out, pre-update | Post-update incident |
|---|---:|---:|
| Duration | 60 minutes | 42 minutes |
| Requests | 1,313 | 1,659 |
| Requests/minute | 21.88 | 39.44 |
| API-list equivalent/request | $0.1056 | $0.2274 |
| API-list equivalent/minute | $2.31 | $8.97 |
| Relative rate | 1.00× | 3.88× |

Same resumed root session

| Client period | Duration | Requests | Processed input context | API-list equivalent |
|---|---:|---:|---:|---:|
| 2.1.251 | approximately 9 hours | 2,274 | 388.0M | $306.16 |
| 2.1.257 | approximately 46 minutes | 1,584 | 277.3M | $361.45 |

The 2.1.257 portion generated a greater API-list-price-equivalent workload in 46 minutes than the same session generated during approximately nine hours under 2.1.251.

Server-side account evidence

The client recorded 43 distinct quota-related 429 responses:

• 31 org_spend_cap_reached responses between 14:37:43 and 14:38:54
• 12 org_level_disabled responses between 14:41:19 and 14:42:03

Two OAuth account switches occurred during this period.

Client logs do not include a reliable organization/account identifier on each request, so requests cannot be authoritatively assigned to individual subscriptions. The client’s cached usage display also appeared stale across an account switch.

Consequently:

• Server-side blocking is confirmed for the original organizational account and at least one subsequently active account.
• The user interface reported exhaustion on all three accounts.
• Local traffic after one brief Max-account switch does not appear sufficient by itself to explain that account showing complete exhaustion.
• Anthropic must use server-side records to determine whether the remaining observation was stale client state, cross-account state leakage, or incorrect quota metering.

Update-specific observations

• Version 2.1.257 is the first published release after 2.1.252.
• It changed the default fable alias to Fable 5.1.
• The local installation recorded an automatic Fable alias migration immediately after installation.
• No CLAUDE_CODE_SUBAGENT_MODEL or force override was configured.
• Fable 5.1 requests occurred before the settings file was rewritten to name Fable 5.1 explicitly.
• Version, model, and effort changes can invalidate prompt-cache keys.
• Workflow and subagent requests use five-minute cache lifetimes, making a fleet-wide cold-cache transition particularly expensive.
• Version 2.1.257 also changed interrupted subagents so they can continue after sleep, dropped connections, or server errors.
• Fable 5.1’s published base pricing does not independently explain the increase. The measured difference came from request volume, cache-write volume, model mix, and work performed per request.

Evidence limitations

• Subscription-quota weighting is not published and cannot be reconstructed from client logs.
• The client does not identify the charged organization on every JSONL usage record.
• Cached account-utilization state was not reliable enough to apportion usage.
• Client evidence cannot independently distinguish changed model behavior from a server-side metering defect.
• Repository content and account identities cannot be posted publicly, but sanitized timestamps, request IDs, model routing, cache categories, and usage totals can be supplied privately.

Requested investigation

Please correlate the private originating session and /feedback submission timestamp with server-side records and:

1. Reconcile each request against the affected accounts’ five-hour, weekly, and Fable-specific quota debits.
2. Confirm the effective model and effort for the root session and every subagent.
3. Determine why Fable 5.1 was selected before the settings rewrite.
4. Review the 2.1.257 Fable-alias migration for unintended model or account-state changes.
5. Check for repeated full-context cache invalidation across subagents.
6. Compare workflow worker counts, continuation behavior, and cache writes between 2.1.251/2.1.252 and 2.1.257.
7. Check retry, fallback, and automatic-continuation paths for duplicate inference or quota debit.
8. Determine whether cached utilization or entitlement state leaked across OAuth account switches.
9. Restore affected usage if the server-side debit was incorrect.
10. Publish a mitigation, warning, rollback, or hotfix if this is a client regression.

Potentially related reports

• Issue #91289: limit exhausted in under 20 minutes
• Issue #91318: five-hour usage consumed in approximately 10 minutes
• Issue #91315: suspected incorrect Fable safeguard or usage trigger

These reports corroborate the timing and symptom class but do not establish a common root cause.

Environment Info

• Platform: darwin
• Terminal: ghostty
• Version: 2.1.257
• Feedback ID: 2e07738c-eaf7-4e18-b24c-794fb6b3fc3f

Errors

[]