Claude Opus 5.5 Review: Why It's My New Claude Code Default

Claude Opus 5.5 built me a playable platformer from one prompt in 20 minutes, and an audio-reactive music visualizer from one prompt in about two hours. I didn't touch the code on either one. With the models I had before, builds like that would have cost me around $100 in Claude credits and burned through my whole usage window.

After a day of testing, Opus 5.5 is my default model in Claude Code. This is my hands-on review: what Anthropic actually changed, the four benchmarks I pay attention to, the effort setting I'd use (it isn't max), the two builds, and where the model fits in my setup.

What Anthropic changed#

The headline claim is simple. Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.

The per-token price dropped 20%, to $4 per million input tokens and $20 per million output tokens. That's nice. The bigger cut is cache reads, down 60% to $0.20 per million, from $0.50 on Opus 5.

That second number matters more than it sounds. A coding agent rereads the same files and the same conversation over and over, so cached tokens are most of what you pay for when you run one. Anthropic says as much in its own post: cache reads are the majority of agentic and coding costs. A 60% cut there is a bigger deal for your real bill than the headline price.

If you're on a Pro or Max plan instead of the API, Anthropic also raised the five-hour usage limits. Cheaper work plus a bigger window means the same plan goes a lot further.

The change I like most has nothing to do with code. Anthropic says the writing is clearer and puts the most important information up front. That was my biggest complaint about Opus 5. On day one, the difference was already noticeable. If you use Claude for everyday business work like emails, summaries and plans, this is the upgrade you'll feel first.

The only benchmarks I pay attention to#

Anthropic's launch table puts Opus 5.5 in first place on most rows. But benchmarks don't tell you how a model performs day to day, so I use one filter: I only care about tests that grade finished work. Would a reviewer merge this code? Did the spreadsheet actually come out right? Quiz-style questions matter a lot less to me.

Opus 5.5 leads most rows, while GPT-6 Astra takes AutomationBench and Terminal-Bench-Science. Four of these benchmarks pass my rule:

• Terminal-Bench 4.0 is real work inside a terminal. Opus 5.5 scores 66.4%, against 57.9% for OpenAI's GPT-6 Astra and 55.8% for Fable 5.1.
• GDPval-AA grades office work: documents, spreadsheets and slides. Opus 5.5 has the top score of every model in the table, 1846 against 1735 for Fable 5.1.
• FrontierCode asks whether you'd actually merge the code. This one hides the most useful trick in the whole launch, and I'll get to it in the next section.
• AutomationBench is Zapier's test of business workflows. It's the one Opus 5.5 doesn't win: 40% against 41.4% for GPT-6 Astra.

One more thing I appreciated. Anthropic says the real gap between Opus 5.5 and Fable 5.1 is narrower than these scores suggest. A lab playing down its own lead is rare, and it matches what I've seen so far. Opus 5.5 isn't miles ahead of Fable. It's roughly Fable-level work for a lot less money, and that's the whole story of this model.

If you want every chart with cost per task at every effort level, including GPT-6 Sol and GPT-6 Luna, I put it all in my Claude Opus 5.5 vs GPT-6 Sol and Luna benchmark breakdown. This post is about what happened when I actually used the model.

Use medium effort, not max#

Opus 5.5 has effort settings from low all the way up to max. Most people crank it straight to max. I wouldn't.

On FrontierCode, Opus 5.5 on medium scored 54.6% at about $0.80 per task. On max it scored 54.4% at $6.19. That's the same result for 7.7 times the price. Extra high actually did worse than medium, at 51.4%.

It gets better. Medium also beat Fable 5.1 running on max, which scored 50.3% and cost $12.83 per task. That's a higher score for about a sixteenth of the price.

Anthropic's own Terminal-Bench chart shows the same thing at the top end. Max costs more per attempt than extra high and scores lower.

My rule: start on medium, move to high when a task is big, and skip max unless you've tested that it earns its price on your own work.

What other people are building#

This is the most hyped launch I've seen in a while, and the demos explain why. People are one-shotting full games, like a Roblox game someone built from scratch. Someone on X generated 100 web pages from one prompt, each with its own design. Others are redesigning their landing pages, and one person built a Minecraft-style voxel world.

Launch-day demos are fun, but they're also cherry-picked. So I ran my own.

My test: two prompts, zero edits#

I gave Opus 5.5 two prompts: a side-scrolling runner game and a music visualizer. They're the same two prompts I gave OpenAI's GPT-6 Sol and GPT-6 Luna. Each build got its own fresh chat and the same high effort level, and I didn't touch the code at all.

Sundown Courier, a one-shot platformer#

The game is called Sundown Courier. You play a comet fox carrying an ember to a beacon before the sun goes down. You run with the arrow keys, jump gaps and monsters, pick up speed on the downhill slopes, smash through crystal obstacles and get sent back to your last checkpoint when you die.

It's harder than it looks, and apparently I'm really bad at it. But the physics and game mechanics feel good, it's easy to pick up, and it's genuinely fun to play.

My favorite detail: it wrote its own test. The build includes a scripted bot that plays through the whole level, and the bot finished it in 41.3 seconds.

Anthropic's post mentions a tester who had several Claude models build a game from one prompt, and Opus 5.5 scored highest on graphics and polish. That's exactly what I saw.

Afterglow, a music visualizer#

The second prompt produced Afterglow. You load your own audio file and pick a style from the side panel: Halo, Ridgelines, Nebula, Particles or Scope. Every style reacts to the music, and each one has its own settings to play with.

This isn't one effect on a black screen. It's a small app with five modes and real controls, and all of it came from one prompt.

How did GPT-6 Sol and Luna do with the same two prompts? It wasn't even close. The full side-by-side is my next video.

Where Opus 5.5 fits in my setup#

Opus 5.5 is now my default in Claude Code. I used to switch to Fable for the hardest problems. At these prices, I don't see a reason for most of my work to go through Fable anymore. If you've never changed the model before, here's how to change the model in Claude Code.

One caveat. A great-looking first screen doesn't prove the code underneath is good. I still need to do a lot of testing to confirm Opus 5.5 writes clean code I'd actually merge into a real project. The first tests are convincing, but that's not the same as months of daily use.

What's missing right now is a small, cheap Claude. My favorite setup is one smart model that plans and a cheap, fast model that does the grunt work. OpenAI just shipped exactly that: GPT-6 Luna costs $0.10 per million input tokens. Anthropic's cheapest current model is still Haiku 4.5 at $1, ten times more. Routing like this is where the real savings are. It's how I got 19 OpenClaw agents running for $2 a month.

The good news is that Anthropic says Sonnet 5.5 and Haiku 5.5 are coming in the next few weeks. Until then, I hand the cheap sub-agent work to Sonnet 5. And since Anthropic calls this the first model in the 5.5 family, my guess is that a Fable 5.5 isn't far behind. That's my own speculation, not anything Anthropic has confirmed.

Where I'm putting it to work#

The real test for a coding model isn't a one-prompt game. It's the products I already run, where a bad merge breaks something a paying customer uses. I build and maintain all three of these with AI coding agents, and they're where Opus 5.5 has to prove itself over the next few weeks.

Rambleproof#

Rambleproof is my dictation app for Mac. Hold a shortcut, talk, and clean text lands wherever your cursor is. It works local-first, and it's how I write a lot of my long prompts for coding agents. Talking through what I want is a lot faster than typing it.

Sevenposts#

Sevenposts turns one product photo into a week of branded social posts, with captions and hashtags, in about three minutes. I built it for cafes, restaurants and small brands that need to post every day and don't have a designer.

AI Headshot Generator#

Upload 3 to 10 selfies, pick a style, and get 40+ studio-quality headshots in minutes, with no photographer and no studio. There's a free tier, so you can see your results before you upgrade.

Should you switch?#

If you code or do business work with Claude, yes. Switch today, use medium or high effort, and skip max.

Next, I'm putting Opus 5.5 head-to-head with GPT-6 Sol and Luna on these same two prompts, and I'll show how I put the winning game online on my own domain. If you want a head start on that part, I already walked through how to build and deploy a website with Claude Code on a real domain.

This review isn't sponsored by anyone.

Watch the full review on YouTube: Claude Opus 5.5 Review: Why It's My New Claude Code Default