New data: Claude Opus 5.5 scores 58 on the Artificial Analysis intelligence index. The previous leaders, Fable 5.1 and GPT-6 Astra, were tied at 53.
And it costs less than half of Fable 5.1 per token. I just watched a quick breakdown from Matt Wolfe of Future Tools, who recorded it in his Palo Alto hotel room before heading to Meta Connect. On the same day, OpenAI shipped GPT-6 Sol and Luna. The creator’s take was clear: one launch was a real leap and the other was a small step.
Here’s what matters for anyone who uses these models every day.
What the numbers actually say
Anthropic says Opus 5.5 performs at the Fable 5.1 level on most tasks and costs 40% less to run than Opus 5. The author shared the benchmark results, noting that he cares more about what people actually build than about scoreboards.
- Agentic coding: 66.4%, top of the chart
- Frontier Code: 54.4%, ahead of Fable 5.1
- Cursor Bench: 57.8%, also first
- GDPval (knowledge work): 1846, the best score listed
It doesn’t win everything. GPT-6 Astra still leads on Automation Bench and Terminal Bench. Still, Opus 5.5 beats Fable 5.1 and Opus 5 on almost every test, and it beats Astra on most of them.
The pricing is the real story
This is where it gets interesting. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Opus 5 was $5 and $25.
Now compare that with Fable 5.1, the model Opus 5.5 beats on most tests. Fable costs $10 input and $50 output. That’s a 60% price cut for a smarter model.
There’s a catch, though. Opus 5.5 is chatty. The expert pointed out that it uses about 119,000 tokens per task, compared with 78,000 for Fable 5.1. So the cost per task only drops from about $7.63 to about $5.98 in max mode. It’s still cheaper, just not as dramatically as the token prices suggest.
I liked the creator’s framing here: cost per task is the number that counts. Token counts on their own don’t tell you what you’ll actually pay.
The demos are wild
This is the part that got me. The original poster walked through what early-access users built, and it’s a big jump from a year ago.
- An animated short video built entirely in JavaScript, no After Effects
- A children’s-drawing-style cartoon where Opus drew every frame in code
- A working mini Game Boy emulator
- A claymation-style video rendered in Blender, where the model checked its own frames before exporting
- A Snake game with sand physics and a fading trail
- A Dark Souls-style game, a cockpit flight simulator, and a Mario Maker clone from Alex at Forward Future
The author said that if someone had sent him that Dark Souls clip a year ago, he’d have assumed it was footage from a real game.
And then OpenAI dropped GPT-6 Sol
About two hours later, OpenAI released GPT-6 Sol and GPT-6 Luna. Both are faster and much cheaper:
- GPT-6 Sol: $2 input / $10 output, half the price of 5.6 Sol
- GPT-6 Luna: $0.10 input / $0.50 output, down from $0.20 / $1.20
But Sol scores around 48 on Artificial Analysis. That’s ten points behind Opus 5.5 and still below Astra. The creator also noticed that OpenAI’s announcement carefully skipped some of the usual benchmark comparisons. On the plus side, Sol is cheap. It costs about $1.60 per task and uses only about 31,000 tokens.
One twist: on the author’s own personal leaderboard, GPT-6 Sol actually came out on top and Opus 5.5 landed eighth. He was upfront that this is one person’s taste, and that his overall read still favors Opus.
3 practical ways to use this
- 🎬 Code-based animation and motion graphics. If you make explainer videos or social clips, try asking Opus 5.5 for JavaScript or Blender scripts instead of opening a timeline editor. Start small: a 10-second loop with one character and one scene.
- 🎮 Fast game and app prototypes. The Snake, Game Boy, and flight sim demos show it can handle interactive logic plus polish. Use it to build clickable prototypes before you commit dev time. A prompt to try: “Build a browser-based [game type] in a single HTML file with smooth animations and a restart button.”
- 💸 Split your workloads by cost. Send hard reasoning and agentic coding to Opus 5.5. Send high-volume, simpler API jobs like classification, tagging, and short summaries to GPT-6 Sol or Luna. That split could cut your bill a lot without hurting quality where it matters.
Tips and pitfalls
Pros
- Top intelligence score available right now
- Big price drop compared with Fable 5.1
- Strong at creative coding, animation, and games
- Available on paid plans and the API today
Cons
- High token usage cuts into the savings
- Astra still wins some automation and terminal tasks
- Early demos come from hand-picked, early-access users
Practical tips:
- Track cost per task, not price per token. A cheaper token means nothing if the model writes twice as much.
- Test on your own workflows. The creator’s personal leaderboard ranked things very differently from the public benchmarks.
- Don’t write off Sol. If you’re a developer who needs a cheaper GPT model, it’s a solid upgrade. It’s just not a frontier jump.
- Keep an eye on Fable. If Opus 5.5 already beats Fable 5.1, the next Fable release should push things even further.
GPT-6 Sol and Luna are live in ChatGPT and Codex for paid plans. Free and Go users get Luna in the desktop app.
It’s worth watching the full video to see the demos yourself. The Dark Souls clone alone is worth the click.