Claude Opus 5.5: Cheaper at Default Effort, Pricier at Max

Claude Opus 5.5 shipped on September 22, 2026, sixty days after Opus 5, with a headline that makes two claims at once. Anthropic says it "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5."

Both halves hold up better than launch claims usually do. Both also depend on a setting most people will never look at.

The per-token price fell 20%. Most of the rest of the 40% comes from a change to the default effort level, which dropped from high to medium. Leave it there and the independent numbers say you get Opus 5’s best work for a little over a fifth of the money. Turn it up to max to chase the headline benchmark scores and Opus 5.5 costs more to run than Opus 5 did.

We covered what Opus 5 shipped in July and what Anthropic’s faster release cycle means for your stack a week later. This is the next turn of that cycle, two days old.

What Claude Opus 5.5 shipped, and where you can use it

One model, claude-opus-5-5, which Anthropic describes as the first model in a new Claude 5.5 family. There is no dated snapshot suffix. Anthropic’s versioning documentation says models from the 4.6 generation onward use a dateless format and are pinned snapshots, so the ID is the version.

Availability on day one, per Anthropic’s release notes:

• Claude API as claude-opus-5-5.
• Amazon Bedrock as anthropic.claude-opus-5-5, with global, US, EU, Australia and Japan inference profiles, and on Claude Platform on AWS.
• Google Cloud and Microsoft Foundry, both as claude-opus-5-5.
• GitHub Copilot, rolling out to Pro+, Max, Business and Enterprise plans, per GitHub’s changelog.

In the Claude apps it is available on Pro, Max, Team and Enterprise plans. The Free plan is not on that list, so staff on free accounts will not see it.

In Claude Code the switch has already happened. The opus alias and the default model setting both now resolve to Opus 5.5 for subscription plans, the API, Bedrock and Google Cloud, per the model configuration docs. Microsoft Foundry is the exception. If your team runs Claude Code on default settings, it has been running Opus 5.5 since Tuesday.

The context window is 1 million tokens with no long-context price premium. Maximum output is 128,000 tokens, or 300,000 through a Batch API beta. Anthropic gives a June 2026 knowledge cutoff; the AWS Bedrock model card says August 2026. We could not resolve that conflict, and it matters only if you rely on the model knowing recent events, which is a bad idea with any model.

The prices, set out in full

Per million tokens, from Anthropic’s pricing page.

Input and output are 20% below Opus 5. The cache read price is the bigger move: 60% lower, and now level with Sonnet 5. Any workload that sends the same long system prompt, style guide or document set on every call has most of its input billed at the cache rate, so the effective saving on that kind of work is well above 20%.

Cache writes are $5.00 per million for the five-minute cache and $8.00 for the one-hour cache. The minimum cacheable prompt is 512 tokens.

Fast mode, a research preview that charges more for quicker output, is $8.00 input and $40.00 output and is on the Claude API only. It is not available on Bedrock, Google Cloud or Foundry. Pinning processing to US infrastructure with inference_geo: "us" adds 10% to every line.

Against Fable 5.1, Opus 5.5 is 60% cheaper on input and output. If "the level of Claude Fable 5.1 on most work" holds for your tasks, that is the comparison that changes a budget.

Where the 40% actually comes from

Anthropic’s exact wording, from the launch page: "Our tests show that at default settings it will cost 40% less than Opus 5 on typical workloads."

"At default settings" is doing real work in that sentence. Opus 5’s default effort was high. Opus 5.5’s default is medium, per the what’s new page. Any request that does not set an effort level now runs at medium. Anthropic also says that at the same effort, Opus 5.5 "needs fewer steps and tool calls per task than Opus 5."

So the saving is three things stacked: a lower price per token, fewer steps per task, and a lower default setting. The third is the largest, and it is the easiest to undo.

Artificial Analysis tested Opus 5.5 at several effort levels and published what it cost to run its full evaluation suite at each. Two results matter:

• At default medium effort, Opus 5.5 scores 51 on their Intelligence Index, the same score Opus 5 reached at max effort. The suite cost $1,627. Opus 5 at max cost $7,275. Same score, about 22% of the cost.
• At max effort, Opus 5.5 scores 58, the highest on the index. It used about 119,000 tokens per task against Opus 5’s 73,000, and the suite cost $8,708. That is 20% more than Opus 5 at max.

The second result matches something Anthropic documents itself. At the same effort setting, Opus 5.5 "tends to think more per turn" than Opus 5, most noticeably at xhigh and max. A lower price per token does not help when the model writes roughly 60% more tokens.

Anthropic did publish some default-effort comparisons, and they are the ones worth reading. At medium, it says Opus 5.5 "beats GPT-6 Astra at max effort for about a fifth of the cost per task," and scores 52.5% on CursorBench against 51.8% for Fable 5.1 at max and 46.6% for Opus 5 at max. Those are self-reported, but they are at least measured at the setting most people will run.

Self-reported numbers against independent ones

Anthropic’s launch table, self-reported, max effort unless noted:

Anthropic’s own table includes two losses, and it deserves credit for leaving them in. GPT-6 Astra leads on AutomationBench, 41.4 to 40.0, and on Terminal-Bench-Science, 64.6 to 58.7. There is no SWE-bench figure at all.

Where independent testers ran the same benchmark, they got lower numbers:

A gap like that is normal and is not evidence of anything improper. Harness, tool access, effort setting and retries all move these scores, which is the argument we made in vendor benchmark scores are not leaderboard results. Anthropic’s Terminal-Bench figure was run at xhigh with a stated margin of plus or minus 2.6 points. GDPval-AA matches exactly at 1846 because Artificial Analysis runs that benchmark itself and Anthropic is quoting it.

On the overall ranking, Artificial Analysis places Opus 5.5 first on its Intelligence Index at 58, "five points clear of GPT-6 Astra and Claude Fable 5.1, which are tied at 53." That is the strongest independent result of the launch. One detail of method: their runs had Anthropic’s server-side fallback option switched on, which is worth knowing if you try to reproduce them.

What does not exist yet: an ARC Prize entry, an LMArena listing, or an independent SWE-bench run. METR reviewed the model before release, with about ten business days of access and Anthropic holding review rights over the text, and concluded it is "unlikely to be able to fully automate AI R&D." That is a safety finding, not a capability score.

Where it is weak

Vals AI scored Opus 5.5 at 3.75% on Harvey’s Legal Agent Benchmark, 31st of 64 models. That is one benchmark in one narrow category of agentic work. It is still a useful counterweight to "Fable level on most work." The word "most" is carrying weight in that sentence, and the only way to know which side of it your work falls on is to test with your own tasks.

The other known weakness comes from Anthropic’s own system card, which says Opus 5.5 "is more likely than previous models to follow malicious instructions in text that a user pastes into their own prompt." The same document says it performed similarly to or better than Opus 5 on every prompt injection evaluation Anthropic reports, the kind where hostile text arrives from a web page or a tool. We will cover that distinction in detail separately. For now, the short version: if staff paste emails, documents or form submissions into Claude, that is the channel where this model is weaker than the last one.

Anthropic put that finding in the executive summary rather than an appendix. That is how a disclosure should work.

Retention: the real difference from Fable 5.1

The launch page states: "Like previous Opus models, Opus 5.5 is available with zero data retention." Anthropic’s 30-day retention requirement applies to a named list of covered models on its data retention page: Fable 5, Fable 5.1, Mythos 5 and Mythos 5.1. Opus 5.5 is not on it.

For anyone working with member records, student data, donor files or case notes, this is the most useful fact in the launch. Our Fable 5.1 coverage turned on that retention rule, and for a lot of organizations it was the reason Fable was not an option at any price. If Opus 5.5 does reach Fable-level results on your work, it gets there without the retention condition.

One caution. The retention page was last updated before this launch. The announcement is explicit about zero data retention, so the conclusion is sound, but confirm it against your own agreement before you rely on it for a compliance decision. AWS separately states that Bedrock offers zero data retention by default.

What happens to Opus 5 and the older models

Nothing, for now. Opus 5 is listed on the deprecations page as "Active (legacy)," with retirement not sooner than July 24, 2027. Anthropic’s docs suggest you "consider migrating," which is advice, not a deadline. Opus 5.5 itself carries a not-sooner-than date of September 22, 2027.

The dates that matter this quarter belong to older models, and this launch did not change them:

• Claude Sonnet 4.5: not sooner than September 29, 2026.
• Claude Haiku 4.5: not sooner than October 15, 2026.
• Claude Opus 4.5: not sooner than November 24, 2026.

If anything you maintain still pins one of those IDs, that is a more urgent job this week than adopting Opus 5.5.

Anthropic also says "Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks, with many of the same improvements." If your production work runs on Sonnet or Haiku for cost reasons, which for most high-volume tasks it should (see our tier selection guide), the sensible plan is to wait for those rather than move volume work up to Opus.

The ninety-minute overlap

TechCrunch reported that Opus 5.5 went out roughly ninety minutes before OpenAI announced GPT-6 Sol and Luna. Other outlets give different gaps, and neither company published a time, so treat the number as approximate.

The consequence runs both ways. OpenAI’s launch comparisons were against Opus 5 and Fable 5.1, a generation behind by the time anyone read them. Anthropic’s table compares against GPT-6 Astra and GPT-5.6 Sol, not the GPT-6 Sol that shipped the same morning.

Artificial Analysis has both. Opus 5.5 scores 58 on its index and GPT-6 Sol scores 48. GPT-6 Sol costs $2.00 input and $10.00 output, half of Opus 5.5’s price. Those two facts together are the real decision for a cost-sensitive team, and neither vendor’s table sets it out for you.

Subscriptions and usage limits

For Claude subscribers, Anthropic says it is "increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans," and it gave subscribers a one-time rate limit reset they can save for later. Anthropic has not published the new limit figures. MacRumors reports the saved reset can be used until October 22; we could not confirm that date with Anthropic.

Anthropic also says Opus 5.5 "generates output more than 30% faster than Opus 5." Independent speed figures had not been published at the time of writing.

What to do

Set effort explicitly. Do not rely on the default in either direction. If you want the cost saving, set medium and test. If a task needs high or above, budget for it, and do not assume Opus 5.5 at max costs less than Opus 5 did.

Measure tokens per request, not price per token. Run a day of representative traffic at your chosen effort level and compare it with your Opus 5 baseline before you switch production.

Check your integration before you change the model ID. Opus 5.5 rejects some request settings that Opus 5 accepted, including turning thinking off and forcing a specific tool call. Code that does either will get errors, not slightly different output.

Tell your Claude Code users. Their default model has already changed.

If retention kept you off Fable 5.1, test Opus 5.5 on that work now. That is where this release most changes what is possible for organizations handling sensitive records.

Hold volume work for Sonnet 5.5 and Haiku 5.5. They are weeks away, not months.

Frequently Asked Questions