Sonnet 5.5 vs Opus 5.5: Half the Price and Double the Fun ???

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Sonnet 5.5 vs Opus 5.5: Half the Price and Double the Fun ??? · Channel: Prompt Engineer 48 · 05:41 · Watch on YouTube

Topics: AI evaluation and benchmarks, Claude, Markets and economy, AI models, AI safety

Summary

The video reviews Claude Sonnet 5.5, comparing Anthropic’s reported benchmark results, pricing, and capabilities with Sonnet 5 and Opus 5.5. Sonnet 5.5 performs close to Opus on several tests and exceeds it on one cited coding benchmark, while costing half as much per token as Opus. The presenter also checks the model’s summary of Anthropic’s announcement and reports that the figures match the page. The video notes caveats around benchmark claims, frontier coding performance, effort settings, and migration, and recommends trying Sonnet for routine work while reserving Opus for tasks requiring careful judgment.

Key points

The presenter verifies in the Claude app that the selected model is Sonnet 5.5.
Anthropic describes Sonnet 5.5 as faster and cheaper than Sonnet 5 for most work.
Anthropic reports Sonnet 5.5 scoring 70.6% on Terminal Bench 4.0, compared with 66.4% for Opus 5.5 at its highest effort setting.
Sonnet 5.5 is close to Opus 5.5 on several cited benchmarks, but Opus remains ahead on Frontier code and slightly ahead on some other results.
Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens; Opus 5.5 costs $4 and $20 respectively.
Effort settings affect speed, token use, and how thoroughly the model reasons and checks its work.
Customer reports cited by Anthropic and other companies describe faster task completion or lower token use, but the presenter notes these are not their own tests.
Anthropic says Sonnet 5.5’s safety performance matches or improves on Sonnet 5 on most tested measures.
The presenter advises checking the migration guide when using Sonnet with thinking off, due to a new between-tools setting.
The recommendation is to try Sonnet 5.5 for everyday coding, documents, and agent work, while using Opus 5.5 for harder tasks requiring careful judgment.

Chapters

Start	Chapter	Summary
00:00	Introducing Sonnet 5.5	The presenter verifies the model in the Claude app and outlines Anthropic’s positioning of Sonnet 5.5 within the Claude model family.
01:02	Benchmarks and comparisons	The video compares reported Sonnet 5.5 benchmark scores with Sonnet 5 and Opus 5.5, noting that results vary by test.
02:07	Pricing, effort settings, and customer reports	The presenter explains token pricing and effort settings, then shares performance and token-use claims from companies.
03:39	Testing the model and reviewing safety	The presenter asks Sonnet 5.5 to summarize Anthropic’s announcement, checks its figures against the page, and describes reported safety results.
04:40	Caveats and who should use it	The video covers benchmark and migration caveats, recommends Sonnet 5.5 for routine work, and says Opus remains suited to tasks needing careful judgment.

Quotes

faster and cheaper than the last Sonnet. — 00:31

So, on this one test, the cheaper model is ahead. — 01:02

Sonnet is half the price, but it needs fewer tokens per task. — 02:07

these benchmarks are Anthropic’s own numbers, so treat them that way. — 04:40

If your work is hard and needs careful judgment, Opus 5.5 is still the one — 05:11

Concepts and entities

Claude Sonnet 5.5, Opus 5.5, Claude Haiku 5.5, Anthropic, Claude app, Terminal-Bench 4.0, GB Val AA, Frontier Code

Full transcript (original language)

00:00 All right. In this video, we will look at Claude Sonnet 5.5, which Anthropic just released, and I already have it open here in the Claude app. Now, mind you, most of the numbers you will see come from Anthropic’s own page, so I will tell you when it’s their claim and when it’s something I checked on my screen. Let’s go ahead and start. So, first thing, I open the Claude app, I pick Sonnet 5.5 down here at the bottom, and I ask it, “Which model are you?” And you can see it says Claude Sonnet 5.5, and the model ID is Claude Sonnet 55. Cool. That is the right model, so we can

00:31 go ahead. So, what is it? It’s the second model in the Claude 5.5 family. Anthropic calls it a clear upgrade over Sonnet 5. It runs 30% or more faster, and it costs up to 30% less for most work. So, the pitch is really simple, faster and cheaper than the last Sonnet. Now, where does it sit? Opus 5.5 is still the one for complex work that needs careful judgment. Sonnet 5.5 is strongest at everyday, well-scoped tasks like fixing bugs and creating polished documents, slides, and spreadsheets. And Claude Haiku 5.5 is coming in the next

01:02 few weeks, so that’s the third one to watch. Okay, here’s a number that jumped out at me. Terminal Bench 4.0, which is an agentic coding test, Sonnet 5 scored 10.3%, Sonnet 5.5 scores 70.6%, and Opus 5.5 at its highest effort setting scores 66.4. So, on this one test, the cheaper model is ahead. That is Anthropic’s number, not mine. There are a few more. It’s the first Sonnet model to beat Pokémon Red working only from screenshots. And on GB Val AA, which is

01:34 real-world work across a variety of occupations, it scores 1844, and Opus 5.5 scores 1846. So, it’s two points below Opus, and Sonnet 5 was 1449. And the coding ones, Cursor Bench 4.0, 55.5% for Sonnet 5.5, 57.8 for Opus 5.5. So, within about two points. And OSWorld, that’s computer used 80.1% against 81.8 for Opus. Now, on Frontier code, Opus is still ahead, 54.4 against 52.1 for

02:07 Sonnet at X high effort and 46.2 at max. So, it’s not a clean sweep. Now, let’s talk cost because for me that is the main thing. Sonnet 5.5 is priced the same as Sonnet 5, $2 per million input tokens and $10 per million output tokens. Opus 5.5 is 4 and 20, so Sonnet is half the price, but it needs fewer tokens per task. And Anthropic says at low or medium effort it beats Sonnet 5’s best score on several benchmarks for about a tenth of the cost per task. There’s also an effort setting if you

02:37 look here at the bottom right of the app, it says Sonnet 5.5 and medium, that is the default in the Claude apps. On the Claude platform, the default is high. Lower settings answer faster and use fewer tokens, which suits routine work. Higher settings reason for longer and check the work more thoroughly. Now, what do the companies say? Base 44 ran 118 real app builds, and Sonnet 5.5 scored level with Opus 5 in 3.6 iterations per build, where Opus 5 took 7.7. Luvable says roughly half the shell runs to finish a task, and Anthropic

03:08 says it batches tool calls together more than Sonnet 5, so fewer steps. For everyday work, these are what customers report. Balenciaga tested 2,441 finance tasks, and it used about 121,000 tokens per answer, where Sonnet 5 used 497,000. Box says 2.4 times faster. Zendesk says tickets were processed 20% faster, and Slack says about 14% fewer output tokens. Again, these are their words, not my tests. Okay, now let me actually use it. What I do is I ask, “What are

03:39 the new capabilities that you have now?” and I paste the link to the Anthropic page. Then I open the same page in the browser, so I can compare. So, we have the blog post here, and the model is fetching it here in the app. This is the one part I have sped up, and you can see the tag on screen. And this is what came back. It gives me the capabilities, the benchmark table, then pricing, availability, and safety. The model ID is right there. Zero data retention is offered and the default effort is medium in the apps and high on the platform. I checked the table against the page and the numbers match. One thing, the

04:10 frontier code row shows the max effort score 46.2 and the page also has 52.1 at X high. So, it’s not the whole picture. Really simple and it took one prompt. Now, safety Anthropic tested it in roughly 1,850 scenarios and says it matches or improves on Sonnet 5 on most measures of alignment, misuse, and honesty. It is the least likely of any of their models to probe the limits of its container. The cyber safeguards are similar to Opus 5.5. So, higher risk security tasks will visibly fall back to

04:40 Sonnet 5. And it is the first Sonnet with classifiers that stop reasoning extraction. So, what are the catches? First, these benchmarks are Anthropic’s own numbers, so treat them that way. Second, Opus 5.5 is still ahead on frontier code and slightly ahead overall. Third, at max effort the frontier code score goes down because it ran its code review skill more and made extra edits. And fourth, if you run Sonnet with thinking off, you need to switch to the new between tools setting. So, check the migration guide before you swap it in. So, who is it for? If you’re

05:11 paying for Opus on everyday coding, documents, and agent work, this is the one to try first because it is half the price per token and nearly level on the test they show. If your work is hard and needs careful judgment, Opus 5.5 is still the one and it is available on the Claude platform and on AWS, Google Cloud, and Azure. So, in this video we have seen what Sonnet 5.5 is, the numbers Anthropic gives, what it costs, the catches, and a quick test in the app. If you’re interested, you can go ahead and try it on your own work and

05:41 let me know how it goes. In the next video I will put it on a real coding task and look at how many tokens it uses. I will see you in another one. Thank you.