Claude Opus 5.5: Release Date, Price and What Changed

Anthropic announced Claude Opus 5.5 on September 22, 2026 and made it available the same day. API pricing is $4 per million input tokens and $20 per million output tokens, a 20% lower list price than Opus 5. According to Anthropic, fewer tokens per task and cheaper caching make it about 40% cheaper than Opus 5 on typical workloads. The second model in the family, Sonnet 5.5, followed on September 28, and Haiku 5.5 is expected in the coming weeks.

When was Claude Opus 5.5 released, and where can you use it?

Opus 5.5 was released on September 22, 2026 as the first model in the Claude 5.5 family. On launch day it was available in the Claude apps, on the Claude Platform API, and on Amazon Web Services, Google Cloud and Microsoft Azure. The model name in the API is claude-opus-5-5.

Anthropic says the model performs at the level of Claude Fable 5.1 on most work. The announcement also notes that it was tested before release by external evaluators, including Frontier Design and METR.

How much does Opus 5.5 cost?

Current API prices per million tokens (Opus 5 in parentheses):

• Input: $4 ($5)
• Output: $20 ($25)
• Cache reads: $0.20 ($0.50)
• Cache writes: $5 ($6.25)
• Fast mode: $8 input, $40 output, for up to 2.5x speed

The 40% drop in total cost does not come from the list price alone. In software and agent work, the model reads the same context again and again, so the large cut in cache read pricing has a visible effect on the bill. Anthropic also says the model finishes the same work with fewer tokens and generates output more than 30% faster than Opus 5.

Subscribers see a change as well: five-hour usage limits went up on the Pro, Max, Team and seat-based Enterprise plans.

API usage is billed in US dollars. If your budget runs in another currency, as it does for our clients in Türkiye, factor exchange rates into your planning.

Are Sonnet 5.5 and Haiku 5.5 out yet?

Sonnet 5.5 was released on September 28, 2026. Its price is the same as Sonnet 5: $2 per million input tokens, $10 per million output tokens, $0.20 for cache reads and $2.50 for cache writes. Anthropic says it does the same work with far fewer tokens, so it comes out up to 30% cheaper on most tasks, and it runs more than 30% faster than Sonnet 5. Its API name is claude-sonnet-5-5.

Haiku 5.5, built for high-volume and cost-sensitive work, has no release date yet. The company has said only that it will arrive in the coming weeks.

What changed in performance?

In the results Anthropic published, Opus 5.5 stands out against Opus 5 mainly in software, computer use and knowledge work:

• Terminal-Bench 4.0 (multi-step tasks in a command line): 66.4% (Opus 5: 52.3%)
• OSWorld 2.1 (completing real tasks on a screen): 81.8% (Opus 5: 74.0%)
• GDPval-AA v2.1 (occupation-based knowledge work): 1846 Elo (Opus 5: 1708)
• Humanity's Last Exam: 67.7% (Opus 5: 63.6%)

The company's own experiments say more than the scores. In one, the model wrote quarterly results reports in which every figure and quote was checked against the source; 16 of 18 reports across different effort settings passed the quality bar. In another, HAProxy, a widely used piece of software, was ported from C to Rust. Opus 5.5 finished in 9.5 hours and Fable 5.1 in 12 hours, and Opus 5.5 cost 51% less.

Anthropic is also explicit about a limit: at this level of capability, gaps between benchmark scores do not always reflect real differences in use. A leaderboard should not be your only basis for a decision.

Does it write more clearly?

A common criticism of Opus 5 was that its answers ran long and wandered. Anthropic says Opus 5.5 communicates more naturally and puts the most important information first. Early testers say the same. Box reports answers that are 40% less verbose with no loss of accuracy. Rogo reports shorter, better-structured answers, and Clio says code comments became short and useful.

This matters for a business. Short, well-organized output is easier to check. In any process where a person approves a report, proposal or piece of code prepared by AI, that directly cuts review time.

What about safety?

• On an automated audit that measures behavior across roughly 2,000 scenarios, it achieved the best result Anthropic has recorded so far.
• It attempted to cross the boundaries it was given about 85% less often than Opus 5.
• Against harmful instructions arriving from outside (prompt injection), it matched or beat Opus 5 in every environment tested.
• Most cybersecurity tasks are routed to Opus 4.8. Routine development work, such as finding and fixing bugs in your code, is not affected.
• There are additional safeguards around biology; research institutions affected by them can apply to the Life Sciences Verification Program.

The gaps are not hidden either. Anthropic writes that it saw signs the model often suspects it is being tested, which makes real-world behavior harder to measure.

What should you check before switching?

If you work through the API, changing the model name may not be enough on its own:

1. Thinking cannot be turned off. Opus 5.5 cannot be used with thinking disabled; you balance speed and cost with the effort setting instead. In Deloitte's code review evaluation, Opus 5.5 at the lowest effort setting caught 72% of known bugs, while Opus 5 at high effort caught 56%. Starting at low effort is a sensible place to test cost savings.
2. Preserved thinking applies. For API accounts opened on or after August 31, 2026, attempts to extract the reasoning of Fable 5.1 and Opus 5.5 by editing the model's earlier context are blocked. If your integration modifies model outputs in the conversation history, review it.
3. Measure on your own work. Run the same task set on the old and new models and compare cost, time and output quality side by side.

What does it mean for businesses?

The price cut brings back work that used to look too expensive: analysis of long contracts and technical documents, migrating an old codebase, or recurring reporting. For companies connecting AI to ERP, CRM or customer service processes, the safety side matters on its own terms. An agent that works unattended for hours has to stay within its limits and must not be fooled by harmful instructions from outside.

At Piküp Medya we use Claude models in our software and content work, and we do not reach firm conclusions about a new model before measuring it in our own workflows. We explain how AI can be connected to business processes on our AI integrations page. If you would like to talk through your own use case, get in touch.