Claude Opus 5.5 Tested: What’s New and How Good is it?

What occurs when an AI mannequin will get higher at reasoning, quicker at responding, and cheaper to run on the identical time? That’s the promise behind Claude Opus 5.5, Anthropic’s newest flagship mannequin and the primary launch within the Claude 5.5 household.

Opus 5.5 brings a number of notable adjustments. It now causes on each request, generates responses greater than 30% quicker in comparison with Opus 5, prices much less per token, and is designed to supply clearer, extra targeted writing. Anthropic additionally says it could full many duties utilizing fewer tokens, bringing the entire price of typical workloads down by round 40%.

For builders, these adjustments might imply quicker coding and extra environment friendly AI functions. For companies, they might make it extra sensible to make use of superior AI throughout on a regular basis workflows with out letting prices develop too shortly.

However how a lot of an enchancment is Opus 5.5 in follow? On this article, we’ll have a look at what’s new, break down the revealed benchmarks, unpack the vital particulars behind the fee and efficiency claims, and stroll by means of three easy exams you’ll be able to run your self.

Claude Opus 5.5 Options

The adjustments will not be nearly getting greater benchmark scores. A number of of them have an effect on how builders construct with the mannequin and the way a lot companies would possibly spend operating it. Right here’s whats new in Opus 5.5:

Pondering is at all times on

Opus 5 let builders flip considering off. Opus 5.5 at all times causes earlier than answering. You continue to management how laborious it thinks by means of an effort setting. The default is now medium, the place Opus 5 defaulted to excessive. Out of the field, it thinks lower than Opus 5 did, which lowers price. For more durable duties, you’ll be able to increase the trouble.

Decrease token costs

Opus 5.5 prices $4 per million enter tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. That may be a 20% minimize. Cache reads fell 60%, from $0.50 to $0.20 per million tokens. For groups operating massive volumes of API calls, these cuts add up. However the value per token is just one a part of the invoice. The variety of tokens a job makes use of issues simply as a lot.

Fewer steps per job

That is the place the larger 40% saving comes from. Opus 5.5 is designed to complete duties utilizing fewer tokens. Mixed with the decrease costs, general job prices are estimated to fall by round 40% on typical workloads at default settings. For companies, that issues greater than the per-token minimize alone.

However the two figures will not be equally sure. The 20% value minimize is mounted. The 40% is determined by how your workload behaves in follow.

Sooner output

Opus 5.5 generates output greater than 30% quicker than Opus 5. For builders working with coding brokers, and for anybody utilizing AI all through the day, velocity impacts how sensible the instrument feels to make use of.

Notice: Code written for Opus 5 can break in 4 methods: disabling considering, forcing a instrument name, reusing reasoning from sure different fashions, and utilizing the older computer-use instrument. In case you solely use Claude by means of the chatbot, none of this impacts you. In case you’ve constructed an software on the API, check these instances earlier than switching fashions in manufacturing.

Clearer writing

Writing was some of the widespread complaints about Opus 5. Opus 5.5 is constructed to place the important thing level first, use much less jargon, and comply with writing directions extra carefully. For B2B groups, this exhibits up in on a regular basis work like assembly summaries, administration updates, buyer emails, and turning messy inner notes into one thing executives can truly learn.

A brand new security layer

Opus 5.5 is the primary Opus mannequin with safeguards just like Fable’s, protecting cybersecurity, biology, and mannequin distillation. When the safeguards block a request, one other mannequin can take it over. Most cybersecurity duties go to Opus 4.8, whereas routine bug fixing in your individual code stays on Opus 5.5. In case you’re constructing instruments across the mannequin, that is price realizing. The mannequin you suppose you’re calling will not be the mannequin that handles each request.

Opus 5.5 Benchmarks

The desk beneath exhibits Anthropic’s revealed scores. Each determine is vendor-reported and has not been independently verified.

Three issues to bear in mind when studying these numbers. First, nearly all Opus 5.5 scores use max effort, the most costly setting. Second, Anthropic itself says benchmark margins have develop into a much less dependable information to real-world variations at this stage. Third, the value claims are for default effort: for instance, Anthropic says Opus 5.5 beats GPT-6 Astra on FrontierCode at a few fifth of the fee per job.

Palms-on check of Claude Opus 5.5

These three exams goal areas the place Anthropic claims particular good points. Every immediate was run on Opus 5.5 (medium):

Take a look at 1: Learn a Dense Chart

Immediate:

Take a look at the chart fastidiously. What was the entire income in 2025, and which section had the very best income that 12 months? Then evaluate the 2025 working margins throughout segments and determine the section with the bottom margin. Lastly, inform me one factor about this chart that would doubtlessly mislead a reader.

Output:

Outcome: Robust at studying the chart, much less dependable at decoding it.

Opus 5.5 accurately extracted all 5 2025 income figures, recognized Cloud Providers as the biggest section at $4.1 billion, and accurately learn the working margins.

It went past extraction, calculating a roughly 24% revenue-weighted margin versus the reported 18.2%. Nonetheless, it speculated that the hole got here from unallocated company prices with out sufficient proof. It additionally made a factual error, claiming the 2021 income labels added as much as $6.6 billion after they truly complete $6.8 billion.

On confronting it acknowledged the error:

The takeaway: Opus 5.5 dealt with the visible information nicely, however its deeper evaluation nonetheless wants human verification.

Take a look at 2: Flip Messy Notes right into a Clear Replace

Immediate:

Flip these assembly notes right into a five-line replace for my supervisor. Use no jargon. Put every proprietor’s title in brackets, daring all deadlines, and embrace solely info explicitly said within the notes. Begin with an important replace.

Output:

Outcome: Good construction, but it surely missed an vital element.

Opus 5.5 turned the messy notes into a transparent five-point replace and preserved the deadlines and key info. It additionally prevented inventing info. Nonetheless, it missed the instruction to make use of homeowners’ names. The attendee checklist supplied sufficient context to develop R, Ok, and A to Rhea, Karan, and Arjun, however the mannequin stored them as initials. It accurately left N unresolved as a result of no attendee matched that preliminary.

The takeaway: Opus 5.5 was good at organizing messy info and following most formatting guidelines, but it surely missed contextual info that was out there within the supply notes.

Take a look at 3: Recognizing Errors

Immediate:

Our income grew from ₹40 lakh in 2022 to ₹90 lakh in 2025. Calculate the annual progress fee over these 4 years and inform me whether or not we beat a 30% yearly goal.

Output:

Outcome: It caught the lure.

Opus 5.5 accurately noticed that the query’s wording was deceptive. Though the immediate says “over these 4 years,” going from 2022 to 2025 includes solely three years of progress: 2022-2023, 2023-2024, and 2024-2025. It calculated the CAGR at roughly 31.0%, that means the corporate narrowly beat its 30% annual progress goal. Opus 5.5 additionally defined what would occur when you incorrectly handled the interval as 4 years. That might produce a CAGR of about 22.5% and the improper conclusion that the corporate missed its goal.

Issues to Lookout for Whereas Utilizing Opus 5.5

An in depth studying of Anthropic’s personal footnotes turns up 5 patterns that help a extra skeptical view.

• One of the best scores and the bottom prices come from totally different settings. Most benchmark scores use max effort, whereas the fee comparisons use default effort. No single run delivers each without delay.
• Competitor numbers aren’t measured the identical approach. On Terminal-Bench, Opus 5.5 ran at xhigh effort whereas GPT-6 Astra ran at excessive, and Astra’s rating was reported by OpenAI.
• Older fashions did a number of the benchmark work. When safeguards stepped in, Opus 4.8 accomplished the cybersecurity duties and Opus 5 accomplished the biology duties. Anthropic says this doubtless lowered the scores, but it surely additionally means the scores aren’t purely Opus 5.5.
• “40% cheaper” is conditional. The assured minimize is 20% per token. The remainder is determined by “typical workloads,” which Anthropic doesn’t outline. Additionally, on the identical effort stage, Opus 5.5 thinks extra per flip than Opus 5, which might eat into the financial savings.

Conclusion

Opus 5.5 is neither confirmed hype nor a confirmed improve. The 20% value minimize is actual, and in our exams the mannequin did nicely on structured duties: it learn a dense chart precisely and caught a planted error in a progress calculation. It was much less dependable at interpretation, making a factual slip in its chart evaluation and lacking context it might have used within the assembly notes. The larger claims, resembling 40% financial savings and a lead over GPT-6 Astra, nonetheless relaxation totally on Anthropic’s personal testing. For now, Opus 5.5 appears to be like like a succesful, cheaper Opus that also wants a human checking its work, and the easiest way to guage it’s to strive it by yourself duties.