Claude Strikes Back With Opus 5.5
Better results, real-world builds and practical ways to get more from Claude’s latest model.
Hey smarties 👋
Claude needed a stronger answer to OpenAI. Opus 5.5 looks like one. With improvements over Opus 5 and performance approaching Fable 5.1 at a lower price, it gives readers who switched to Astra a reason to reconsider. This week, we’re exploring what changed, how it compares, and what it costs to use.
Sponsored
Webinar: AI-Native SDLC, 10x Faster with Even Higher Quality
What would your roadmap look like if your team delivered 10x more, at the same quality?
Leobit's experts will show how we run AI agents across the full lifecycle: discovery, feature development, and production.
Topics Covered:
- Traditional SDLC vs AI-native SDLC
- Vibe coding vs agentic engineering for production work
- The building blocks: context, conventions, quality gates
- 10x productivity on our own ERP
- The judgment gates humans keep, and what to measure
- Manager playbook: metrics, roles, budget, risk
Discover how you can apply it in your team next sprint.
📅 September 30, 2026
01 / The big picture
What Has Changed in Opus 5.5?
If you’ve ever asked Claude a simple question and received a dense, overly technical answer, the writing changes may be the first improvement you notice.
It sounds more natural. Opus 5.5 is designed to write more conversationally, put the main point first and explain things in everyday language. The early examples are easier to follow without repeatedly asking Claude to simplify its answer.
More thorough work, less waiting. In Anthropic’s comparison, it produced a more detailed financial analysis and a clearer presentation than Opus 5, with fewer errors. It also finished faster and cost less.
It gets further on difficult assignments. In a research test, Opus 5.5 successfully tracked down a hard-to-find company report and built an accurate analysis from it. Neither Opus 5 nor Fable 5.1 completed that task successfully.
02 / Benchmark
Benchmark Comparison
Opus 5.5 gives Astra serious competition, especially in coding. It scores 66.4% on Terminal-Bench, ahead of Astra’s 57.9% and Fable 5.1’s 55.8%. It also leads both models in the professional-work test, which covers tasks such as preparing reports, spreadsheets and presentations.
Astra keeps the edge in business workflows and scientific research in this chart. But for coding and professional work, Opus 5.5 now looks like a strong alternative worth trying.
03 / Access & pricing
Pricing and Usage
You don’t need a Max subscription to try Opus 5.5. It is available on Claude Pro for $20 a month, with Max starting at $100 a month for higher usage.
Model | Low | Medium | High | Extra high | Maximum |
|---|---|---|---|---|---|
GPT-6 Sol | $0.13 | $0.25 | $0.37 | $0.53 | $1.06 |
GPT-6 Astra | $0.82 | $1.54 | $1.73 | $2.31 | $3.26 |
Opus 5.5 | $0.55 | $1.34 | $1.82 | $3.46 | $5.98 |
Fable 5.1 | $2.37 | $2.98 | $3.91 | $5.98 | $7.63 |
Average cost per Artificial Analysis benchmark task, not subscription charges or fixed prices for your work.
Opus 5.5 costs less than Fable 5.1 at every setting shown. But turning up the effort changes the bill: $1.34 on medium becomes $5.98 on maximum. Medium is the default, so start there and increase it when the work needs it.
For API use, standard rates are $4 per million input tokens and $20 per million output tokens. Anthropic has also announced higher five-hour allowances on eligible paid plans and a usage reset subscribers can save for later.
04 / Practical Use Case
From a Rough Sketch to a Playable Game
Opus 5.5 made this game in just five minutes.
I gave it a rough sketch of a bow and some balloons, along with a few sentences explaining the idea. No detailed design brief. This was the entire prompt:
I want you to create a mini game where I have a bow and I shoot these different balloons that rise up from the bottom part of the screen and go above the screen. Give it a medieval look and also make it gloomy and dark. Make sure to add balloons that are red that I need to avoid that would lose me points.
Here’s what it built:
You can aim, shoot rising balloons and earn points, while avoiding the red ones. There’s a countdown, hit effects and a results screen when time runs out. Claude even added a moonlit castle and glowing targets to match the gloomy setting.
What impressed me was how much it filled in without being asked. One rough sketch and a short prompt were enough to turn the idea into a game I could actually play.
05 / Opus 5.5 Showcase
What People Are Creating With Claude
Example 01
A Website With 25 Miniature Worlds
Kevin Ngo used Opus 5 to create a website featuring 25 miniature rooms, each showing Claude keeping someone company. The surprising detail is how it was made: every room, person and tiny mark was drawn through code, all inside one HTML file, without imported images or libraries.
Example 02
Motion Graphics With a Distinct Visual Style
Min Choi’s post features a motion graphics experiment by mexicat, combining wireframe forms, bold typography and orange accents against a dark background. The creator took inspiration from another Opus 5.5 experiment, chose a different visual direction and refined it over several rounds. It’s a useful example of giving Claude creative direction, then developing the result instead of settling for its first attempt.
Example 03
building an interactive lens lab
Ryan Sael asked Opus 5.5 to explain camera focus by building an interactive lens lab. Turn the focus ring, and you can see the glass elements move and watch which part of the scene becomes sharp. It turns a technical explanation into something you can explore yourself. Ryan reports that the build took 1 hour and 26 minutes in a single run, costing $25.66 through the API.
06 / Comparison
From a Rough Sketch to a Playable Game
Pat Simmons gave Opus 5.5, Fable 5.1 and GPT-6 Astra the same three assignments: illustrate scenes from Moby Dick, create a product launch film and build a 3D skateboarding game. He compared the finished results alongside the time and cost of each build.
Opus 5.5 was his pick in all three tests, delivering richer illustrations, a stronger launch film and more convincing game detail. That extra quality took longer and cost more overall, but for these creative projects, he judged the results worth it.
07 / Prompting Guide
How to Get Better Results From Opus 5.5
Better results start with clearer prompts. To create something like the examples above, explain what you want, show a reference and tell Claude what the finished result should include.
Say exactly what you need. Describe the task, who it is for and the format you want, whether that is a website, presentation or short report.
Show the style you want. Attach a screenshot or sketch and explain what you like about it. Name specific colours, layouts or details instead of simply asking for a “professional design.”
Have it check the context first. Ask Claude to review the relevant files and connected sources before starting, so it does not miss information the work depends on.
Define when the work is finished. List what must be completed and checked. Ask it to finish those items, but get your approval before publishing or deleting anything.
Label anything you paste. Make it clear when an email, article or document is reference material, so Claude does not mistake instructions inside it for your request.
Start with Medium effort. Increase the setting when a difficult task needs more reasoning. Focus on clear instructions rather than repeatedly adding “think harder.”
07 / READER POLL
Last Edition’s Poll Results
Thank you to the 9 readers who voted. Here’s how you rated the last MonDive.
Amazing, very clear
66.7% · 6 votes
Good, but a few parts were tricky
33.3% · 3 votes
Not impressive, it still felt overwhelming
0% · 0 votes
A NOTE FROM A READER
Selected: Amazing, very clear
“It was good”
Our reply: Thank you! We’re glad you enjoyed the edition and appreciate you taking a moment to tell us. We’ll keep working to make each MonDive clearer and easier to follow.
How did you feel about today’s MonDive?
Know someone who may be interested?
Thanks for reading. See you next Monday.
