Your Claude ‘Opus’ Alias Might Still Be Opus 5, Not 5.5 — Here’s the Fix (2026)

A one-line pricing announcement is easy to skim past. What’s less obvious is that switching to the cheaper model isn’t automatic just because Anthropic shipped it. If your app, agent, or Claude Code setup calls the opus alias instead of a pinned version string, you might still be paying Opus 5 rates — or getting Opus 5 quality — weeks after the new model launched.

Contents

• What Actually Changed With Opus 5.5 on September 22, 2026?
• Why Is Your Opus Alias Still Running the Old Model?
• How Do You Check Which Model Actually Answered You?
• How Do You Force Claude Code to Use Opus 5.5?
• Opus 5 vs Opus 5.5: Full Pricing Comparison
• Where the Alias Trick Still Bites People
• How Much Does Switching Actually Save on a Typical Workload?
• Tips
• Warnings
• FAQ
• Sources

What Actually Changed With Opus 5.5 on September 22, 2026?

Anthropic released Claude Opus 5.5 on September 22, 2026, at $4 per million input tokens and $20 per million output tokens — a 20% cut from Opus 5’s $5/$25 rate (anthropic.com). Cache reads dropped further, from $0.50 to $0.20 per million tokens, a 60% reduction. Anthropic says the model performs at roughly Fable 5.1 level on most everyday work, runs more than 30% faster, and — because it uses fewer tokens to finish the same task — ends up costing about 40% less on typical workloads, not just 20%.

When I compared the raw per-token numbers against Anthropic’s own workload-cost claim, the gap made sense: a faster model that needs fewer output tokens saves money twice — once on the lower rate, once on the smaller bill.

Why Is Your Opus Alias Still Running the Old Model?

The opus alias is supposed to track “the newest Opus model,” but in practice it tracks the newest Opus model your specific provider has turned on — and that’s not the same everywhere. According to Anthropic’s model configuration docs (code.claude.com), here’s how the alias resolved right after the September 22 launch:

If your stack runs through Microsoft Foundry, calling opus quietly gets you a model that’s two full versions behind — and you’d never see an error telling you so. Third-party routers add another failure mode. OpenRouter’s claude-opus-latest alias, when a request shape isn’t supported by Opus 5.5, falls back down the family tree to Opus 5, and if that fails too, all the way to 4.8 — each step billed at that older model’s rate, not the one you thought you were calling.

How Do You Check Which Model Actually Answered You?

You don’t have to guess. Anthropic’s own support documentation confirms that model substitutions are supposed to be visible, not silent, inside the first-party apps — you’ll see a notice, and the response is labeled with whichever model actually generated it (support.claude.com). For API and third-party routing setups, check it directly instead of trusting the alias name.

1. Read the response metadata. Every API response includes a model field with the exact model ID that served the request — not the alias you called.
2. Check your Claude Code version. Opus 5.5 as the default requires Claude Code v2.1.280 or later; older builds keep resolving opus to whatever was current when they shipped.
3. Look for a routing layer. If you’re going through OpenRouter, a proxy, or an enterprise gateway, check that tool’s logs for the resolved model name, not just the alias in your own config.
4. Compare token pricing on your invoice. A quick way to sanity-check: if your last invoice shows $5/$25 per million tokens instead of $4/$20, you were billed at Opus 5 rates regardless of which alias you thought you were using.

How Do You Force Claude Code to Use Opus 5.5?

Pin the exact model instead of relying on the alias to catch up. Three ways to do it, from fastest to most permanent:

1. Pass the model flag directly. Run claude --model claude-opus-5-5 for a one-off session that skips alias resolution entirely.
2. Set the environment variable. Export ANTHROPIC_DEFAULT_OPUS_MODEL='claude-opus-5-5' so every session in that shell or deployment resolves to the exact version, regardless of what the provider’s alias currently points to.
3. Edit your settings file. Add "model": "claude-opus-5-5" to your Claude Code settings JSON if you want the pin to persist across machines and shell sessions.

When I checked the provider-specific notes in Anthropic’s docs, Bedrock, Google Cloud, and Microsoft Foundry each expect their own prefixed model ID rather than the bare Anthropic string — pinning the wrong format is a common reason people think the fix “didn’t work.”

Opus 5 vs Opus 5.5: Full Pricing Comparison

Where the Alias Trick Still Bites People

Most of the confusion isn’t about the price cut itself — it’s about assuming the alias already caught up.

Enterprise deployments pinned to Microsoft Foundry are the biggest blind spot. Teams that budgeted for the new 40%-lower cost, then didn’t see the savings show up, are often still calling Opus 4.6 without realizing the alias never moved on that provider.

Agent frameworks that route through OpenRouter are the second trap. A request shape that Opus 5.5 doesn’t support — an unusual tool schema, for instance — doesn’t error out. It just quietly serves an older model instead, at that model’s price.

This part trips up a lot of teams that treat “opus” as a fixed name instead of a moving target.

The third trap is subtler: safety-classifier fallbacks. Certain prompt categories — cybersecurity exploit generation, specialized biology, or (new to 5.5) frontier ML accelerator work — can trigger an automatic downgrade mid-conversation. Anthropic says this is supposed to be labeled, not silent, in the first-party apps, but it’s still worth knowing it exists if a response suddenly reads like an older model wrote it.

How Much Does Switching Actually Save on a Typical Workload?

On a workload of 10 million input tokens and 2 million output tokens a month, Opus 5 rates come out to $50 (input) plus $50 (output) — $100 total. At Opus 5.5’s per-token rate alone, that same split drops to $40 plus $40, or $80 — a 20% cut. Factor in Anthropic’s claim that Opus 5.5 needs fewer output tokens to do the same job, and the real-world bill for that workload lands closer to $60, matching the ~40% figure Anthropic advertises rather than the 20% you’d get from the rate card alone.

Honestly, that gap between “20% cheaper per token” and “40% cheaper in practice” is exactly why checking your actual invoice matters more than trusting the announcement.

Tips

• Pin the full model ID (claude-opus-5-5) in CI/CD and production configs — don’t rely on aliases in anything you can’t watch closely.
• If you’re on Microsoft Foundry, budget for Opus 4.6 pricing until you explicitly update your pinned model ID; the alias won’t do it for you.
• Check your provider’s own docs page for alias resolution before assuming it matches Anthropic’s direct API behavior.

Warnings

• Don’t assume a lower invoice means you’re on Opus 5.5 — check the model field in the response, since some routing layers report the alias you called rather than the model that served it.
• Extended 1M-context pinning syntax (appending [1m]) is provider-specific; using it on a provider that doesn’t support it will fail the request rather than silently ignoring the flag.

FAQ

Does the opus alias eventually update on Microsoft Foundry?

Anthropic’s docs describe aliases as updating over time as each provider turns on new versions, but they don’t give a fixed timetable. Until Microsoft Foundry flips it, you need the pinned model ID to get Opus 5.5 there.

Will pinning claude-opus-5-5 break when Anthropic ships Opus 6?

No — a pinned full model ID stays fixed to that exact version. You’d need to update the pin yourself to move to a future release, which is the point: it trades automatic upgrades for predictable behavior and pricing.

Is the 40% savings figure guaranteed for every workload?

No. Anthropic frames it as typical, driven by faster output and fewer tokens per task. Short, simple prompts may see closer to the flat 20% per-token cut, while long agentic tasks that previously needed lots of output tokens tend to see the bigger drop.

Sources

• Anthropic — Introducing Claude Opus 5.5
• Anthropic — Claude Code Model Configuration Docs
• Anthropic Support — Why Claude Switched Models in Your Conversation

Fact-checked based on public sources as of September 29, 2026. Pricing and provider behavior reflect official Anthropic documentation at time of publication and may change — always confirm current alias resolution and rates on Anthropic’s own docs before budgeting.