Claude Outage — Pricing Impact & What It Means (Sep 2026)
Anthropic reports elevated errors across Claude models, its API, Claude Code, and Cowork. See the live status, cost impact, and fallback checklist.
By AI Pricing Guru Editorial Team
AI Pricing Guru articles are maintained by the editorial workflow behind the site: daily pricing snapshots, provider source checks, and review passes for model launches, subscription limits, and billing changes.
TL;DR
- Claude suffered a major-impact incident involving elevated errors across claude.ai, the API, Claude Code, and Claude Cowork.
- Anthropic resolved the incident at 02:35 UTC and says customer impact ran from 00:50 to 02:10 UTC.
- No API or subscription price changed. The immediate cost is failed requests, retries, interrupted agent runs, and staff time.
- Resume Opus 5 batch work gradually with canaries, safe retries, and checks for duplicated tool actions rather than releasing the entire backlog at once.
Current API price scale—this incident did not change rates
USD per 1M tokens. Input and output rates are charted separately.
Estimate the budget for a temporary API fallback
Assumes 75% input tokens and 25% output tokens using current per-million rates.
Claude Sonnet 5
anthropic
$40.00
- Input share
- $15.00
- Output share
- $25.00
GPT-5.6 Sol
openai
$80.00
- Input share
- $30.00
- Output share
- $50.00
Claude Opus 5
anthropic
$100.00
- Input share
- $37.50
- Output share
- $62.50
Affected Claude models and two possible fallback routes
| Model | Provider | Input / 1M | Cached / 1M | Output / 1M |
|---|---|---|---|---|
| Claude Fable 5.1 | anthropic | $10.00 | $0.25 | $50.00 |
| Claude Mythos 5.1 | anthropic | $10.00 | $20.00 | $0.25 |
| Claude Opus 5 | anthropic | $5.00 | $0.5 | $25.00 |
| Claude Sonnet 5 | anthropic | $2.00 | $0.2 | $10.00 |
| GPT-5.6 Sol | openai | $4.00 | $0.4 | $20.00 |
Built from pricing.json at publish time.
Anthropic has resolved a major-impact Claude incident involving elevated model-request errors. The incident record was opened at 00:57 UTC on September 22, 2026 and affected claude.ai, the Claude API, Claude Code, and Claude Cowork.
Anthropic first said Fable 5, Fable 5.1, Mythos 5, and Mythos 5.1 had returned to normal success rates while Opus 5 remained affected. At 02:11 UTC, it reported normal success rates across all affected models and marked every affected service operational. At 02:35 UTC, Anthropic closed the incident and said customer impact ran from 00:50 to 02:10 UTC.
Claude outage status and timeline
| Time (UTC) | Official update | Operational meaning |
|---|---|---|
| 00:57 | Anthropic began investigating elevated errors for Mythos 5.1, Fable 5.1, and Opus 5 | Requests could fail across Claude’s consumer and developer surfaces |
| 01:17 | The cause was identified and a fix was in progress | Recovery work had started, but no model was declared recovered |
| 01:35 | Fable 5/5.1 and Mythos 5/5.1 success rates returned to normal | Opus 5 remained the active model problem |
| 02:11 | Success rates returned to normal across affected models | Anthropic moved the incident to monitoring; all four services are operational |
| 02:35 | Incident resolved; official impact window reported as 00:50–02:10 UTC | Keep post-incident canaries and reconcile retries, usage, and duplicated side effects |
Claude’s status record now marks the incident resolved and all four services operational. Production users should still confirm recovery with a small canary before releasing queued traffic and reconcile work attempted during the official impact window.
Pricing impact: rates are unchanged, failed work is not free
Anthropic has not announced a price increase, refund, service credit, or billing change. Its official rate card still lists Fable 5.1 and Mythos 5.1 at $10 input, $0.25 cache reads, and $50 output per million tokens; Fable 5 and Mythos 5 at $10/$1/$50; and Opus 5 at $5/$0.50/$25. The live chart, calculator, and table above use the maintained Anthropic pricing data and show current rates.
Our public pricing JSON and API pricing endpoint remain byte-aligned. Their model statuses describe lifecycle state—active, preview, legacy, or retired—not transient service health, so no canonical model row changes for this outage.
The real cost comes from incomplete work:
| Cost source | What to watch during the incident |
|---|---|
| Failed API requests | Confirm whether failed calls appear in usage records before estimating the bill |
| Automatic retries | A retry storm can increase load, latency, and successful-token spend after recovery begins |
| Agent runs | Tool actions may complete even when the final model response fails |
| Staff time | Developers and operators lose time diagnosing what is actually a provider incident |
| Emergency switching | Another model may require prompt changes, fresh evaluation, and compliance approval |
Do not assume an HTTP failure was billed or unbilled without checking Anthropic’s usage console and the final invoice. For workload-specific estimates, use the token cost calculator and include retry volume rather than only successful requests.
What API and Claude Code users should do now
Resume non-urgent Opus 5 batch jobs gradually rather than releasing the entire backlog at once. For interactive traffic, keep exponential backoff with jitter, a retry ceiling, and a circuit breaker during the recovery ramp. Preserve original request IDs and log model, timestamp, status code, token usage, and whether any external tool action occurred.
Never blindly replay a tool-using agent step. A database write, deployment, message, or purchase may have succeeded before the response failed. Check the external system first, then retry with an idempotency key or an explicit recovery step.
Claude Sonnet 5 and Haiku 4.5 were not named among the affected models, and Anthropic now marks the API and related services operational. Test a small canary before rerouting. A cross-provider route adds more independence, but only switch after checking quality, context limits, tool schemas, data residency, and policy requirements. Compare current OpenAI API pricing before activating a budget.
Teams that want a separately billed coding fallback can also evaluate the Z.ai coding plan. Test it on a non-critical repository before relying on it during an incident.
Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored link above at no extra cost to you. It does not affect this analysis.
Labs coverage decision
AI Pricing Guru Labs already includes exact-route results for Claude Fable 5 and Claude Opus 5. Those fixed 49-task results measure model quality and cost under completed runs; they are not uptime guarantees and are not being relabeled as incident-period scores.
The coverage note records an explicit reliability-measurement boundary. A defensible comparison would need synchronized requests across pinned routes and regions, identical concurrency, timeouts and retry rules, raw status codes, latency distributions, provider request IDs, accepted-task grading, and billed-usage reconciliation. The public incident timeline does not provide that dataset.
Bottom line
This was a confirmed multi-surface Claude incident, not a rumor. Anthropic resolved it at 02:35 UTC after reporting customer impact from 00:50 to 02:10 UTC.
There is no announced pricing change. Treat retries, duplicated tool actions, delayed work, and emergency migration as the financial impact. Ramp Opus 5 traffic back up with canaries, make recovery idempotent, and reconcile usage from the impact window.
Sources: Anthropic’s official incident record, Claude status page, and Claude API pricing documentation, plus AI Pricing Guru’s live pricing dataset. Incident state, affected products, model rates, and canonical data checked September 22, 2026 at 02:51 UTC.