Claude Opus 5 lands: frontier work at half the price
Anthropic didn't wait for its IPO roadshow to drop the big one. Opus 5 is out, and the pitch is blunt: frontier-class work at half the price of Fable 5.
The numbers are the story. On Frontier-Bench it more than doubles Opus 4.8's score while costing less. On CursorBench 3.2 it lands within half a percent of Fable 5's peak at half the money. ARC-AGI 3, three times the next-best model. OSWorld 2.0, it beats Fable 5 at a third of the cost. Pricing stays at 5 dollars in, 25 dollars out per million tokens, same as 4.8, so you're getting a generation of gains for free.
The demo everyone's passing around: handed a machine part drawing it couldn't directly view, Opus 5 wrote its own computer vision pipeline to pull the geometry out of raw pixels, then reconstructed the whole part. That's the shape of what changed. It verifies its own work and iterates instead of guessing once and moving on.
For agents there are two real additions. You can swap tools mid-conversation now (beta), and flagged requests auto-fall-back to other models instead of just failing. Small on paper, big when you're running long agent loops that can't afford a dead end.
The quiet subtext is that this is the model that resets where the daily-driver line sits. When frontier costs half of last month's frontier, the question stops being can it do the task and becomes why are you still running the expensive one.
https://www.anthropic.com/news/claude-opus-5
← Back to all articles
The numbers are the story. On Frontier-Bench it more than doubles Opus 4.8's score while costing less. On CursorBench 3.2 it lands within half a percent of Fable 5's peak at half the money. ARC-AGI 3, three times the next-best model. OSWorld 2.0, it beats Fable 5 at a third of the cost. Pricing stays at 5 dollars in, 25 dollars out per million tokens, same as 4.8, so you're getting a generation of gains for free.
The demo everyone's passing around: handed a machine part drawing it couldn't directly view, Opus 5 wrote its own computer vision pipeline to pull the geometry out of raw pixels, then reconstructed the whole part. That's the shape of what changed. It verifies its own work and iterates instead of guessing once and moving on.
For agents there are two real additions. You can swap tools mid-conversation now (beta), and flagged requests auto-fall-back to other models instead of just failing. Small on paper, big when you're running long agent loops that can't afford a dead end.
The quiet subtext is that this is the model that resets where the daily-driver line sits. When frontier costs half of last month's frontier, the question stops being can it do the task and becomes why are you still running the expensive one.
https://www.anthropic.com/news/claude-opus-5
Comments