Yuval Avidani
Author
Claude Opus 5.5 is Anthropic's new model, and according to the company it matches Fable 5.1 - its expensive flagship - on most tasks, while costing an average of 40% less to run than Opus 5. It launched on September 22, 2026, and within two days it grabbed the #1 spot on both the Artificial Analysis index and Arena's WebDev leaderboard. But the internet filled up with something no benchmark table shows: full 3D worlds, games, and Blender models, all built from a single prompt.
Let's break this into two parts: the numbers that matter to whoever's paying the bill, and the 3D hype that matters to everyone else.
What Opus 5.5 actually is, and why "5.5" is confusing
Anthropic now has two floors upstairs. Fable is the top, expensive flagship ($10 per million input tokens, $50 for output). Opus is the floor below it: very smart, cheaper, the one most developers actually run their agents on. Opus 5.5 is the first in a new family - "5.5" - and Anthropic already announced that Sonnet 5.5 and Haiku 5.5 are coming in the following weeks.
Think of it like a new executive-car model that ships with the luxury model's engine, but at the price of the previous model - and even a bit less. That's exactly the pitch here: not "smartest in the world," but "as smart as the flagship, priced like the tier below it."
The numbers: the price, and the detail everyone's missing
The official pricing is $4 per million input tokens and $20 per million output tokens - 20% less than Opus 5, which cost $5 and $25. But the real story is hiding in a different row of the table.
Cache read pricing dropped from $0.50 to $0.20 per million tokens - a 60% cut - and Anthropic writes that cache reads make up most of the cost of agentic and coding work.
Cache is the trick where the model "remembers" part of the conversation it already read, so you don't pay full price for it every single time. Think of it like a coffee shop punch card: you pay full price for the first coffee, and every refill after that comes at a steep discount. A coding agent reads the same codebase over and over, dozens of times an hour, so most of its bill is exactly those "refills." That's how a token price drop of only 20% translates, according to Anthropic, into 40% less on a typical workload's bill.
It's worth playing around with the calculator: in almost every scenario, Opus 5.5 costs less than half of Fable 5.1 - and that, more than any benchmark, is what's going to make teams switch.
The benchmarks: where it leads, and where it doesn't
Here are the numbers from the official System Card, all at maximum effort and averaged over five runs:
On SWE-bench Pro, a test that hands the model real bugs from real codebases and asks it to fix them, it scored 89.9%, versus 79.2% for Opus 5 and 81.2% for Fable 5.1. On Terminal-Bench 4.0, which tests independent work in a terminal, it jumped to 66.4% from 52.3%. But fair's fair: there are tests where OpenAI's GPT-6 Astra leads - on Terminal-Bench-Science (64.6 vs. 58.7) and AutomationBench (41.4 vs. 40.0). And Anthropic itself writes a rare sentence for a launch post: that benchmark gaps have "become a less reliable guide," and that the gap versus Fable 5.1 is "narrower than the scores suggest."
The outside measurements back this up too. Artificial Analysis, which runs every model through the same tests, ranks it #1 on its intelligence index with a score of 58. And on the Code Arena WebDev leaderboard, where real people vote on which generated website turned out better, it's #1 with 1,818 - 26 points above GPT-6 Astra and 126 above Opus 5.
And now the 3D stuff: why the internet is full of worlds
Here's the part that got everyone sharing. On the launch page there's one modest line: a tester who had several Claude models build a game from a single prompt found that Opus 5.5 scored the highest, thanks to its graphics and polish. Online, it looked a lot less modest than that.
Meng To, a well-known designer and developer, used Opus 5.5 to build a three.js scene where you sail a boat through Japanese landscapes with changing weather, a day-night cycle, and 3D characters - and wrote that he's "genuinely impressed by the details." three.js is an open-source library that draws 3D graphics directly in the browser, no installation needed. Think of it as a tiny game engine living inside every web page.
And there's more. Higgsfield AI showed off a Shinkansen bullet train built in Blender, the 3D modeling software, with 5,112 separate objects - down to the 430 seats inside. And a developer named riba2534 uploaded three complete 3D games to GitHub, each one generated from a single-sentence prompt, without a single line of code written by hand. All the models, textures, and sound were generated through code. The repo hit 800 stars within four days.
Two important corrections here, so we don't get carried away by the hype. First, this wave didn't start now: Andrej Karpathy's viral "Lord of the Rings"-style world from August was built with Opus 5, and the connectors linking Claude to Blender have existed since April. Opus 5.5 just does it better. Second, this isn't the same thing as "world models" like World Labs' Atlas or Google's Genie, which generate a world directly as video or a scene. Opus 5.5 writes code that draws the world - which means you can open it, edit it, and run it in any browser.
The fine print: what's broken and what's worth worrying about
For developers, four breaking changes: you can't turn off thinking (it returns a 400 error), forced tool use (tool_choice) returns an error, thinking blocks are now tied to the specific model and conversation, and the old computer use tool has been deprecated. The default effort level dropped to medium, so if you're used to high, you now need to set it explicitly.
And the System Card - in transparency worth crediting - admits to two weaknesses: the model is more likely than its predecessors to follow malicious instructions that a user pastes into their own prompt, and more likely to accept permission claims that can't be verified. On the flip side, it attempted to circumvent guardrails about 85% less often than Opus 5. And the model isn't available on Claude's free tier - only on Pro and above, and via the API.
Bottom line
In my view, Opus 5.5 is the clearest proof yet that the race has shifted from "who's the smartest" to "who gives you the most intelligence per dollar." Flagship-level performance, cache that's 60% cheaper, and 30% faster - that's a combination that's going to change which model a lot of us actually run in production. And the 3D wave proves something deeper: once a model writes code well enough, "build a game" turns from a months-long project into a single sentence. Two caveats, though: these benchmarks come from the company itself, and before you hand it an agent that touches permissions - go read the safety section first.
So if a full 3D world can already get built from one sentence - what world would you build first?