Anthropic shipped Claude Opus 5 on July 24, and the number that jumps out is the
one that hits your API bill: it gets close to Fable 5's frontier intelligence at
half the price.
If you build with LLMs, that combination (near top-tier quality, mid-tier cost) is
the whole game. Here is what actually changed and why it matters.
TL;DR: Opus 5 reaches near Fable 5 quality at half the price, tops Anthropic's own coding and agent benchmarks, and is noticeably better at checking its own work. It is live now on the API as
claude-opus-5at $5 / $25 per million tokens.
Higher effort buys more intelligence; lower effort conserves tokens for faster, cheaper runs. Chart: Anthropic.
What's new
Anthropic reports Opus 5 as state of the art on its coding and knowledge work
benchmarks like Frontier-Bench and GDPval-AA, while still sitting behind their
Mythos 5 model on cybersecurity. On their own evals:
| Benchmark | What Anthropic reports for Opus 5 |
|---|---|
| Frontier-Bench v0.1 (software engineering) | More than doubles Opus 4.8's score, at lower cost per task |
| CursorBench 3.2 (coding) | Within 0.5% of Fable 5's peak at max effort, at half the cost per task |
| ARC-AGI 3 (novel problem solving) | Roughly 3x the next best model |
| OSWorld 2.0 (computer use) | Beats every model at a given cost; matches Fable 5's best at about a third of the cost |
Coding and agent benchmarks (Frontier-Bench, CursorBench, AA Coding Agent Index). Chart: Anthropic.
Pricing stays at $5 per million input tokens and $25 per million output, the same
as Opus 4.8. It is now the default model on Claude Max and the strongest option on
Claude Pro. On the API it is claude-opus-5.
Beyond code
The same pattern holds beyond coding. On reasoning and computer use tasks,
Anthropic shows Opus 5 leading at a given cost too.
Knowledge work and reasoning: ARC-AGI 3, GDPval-AA, OSWorld 2.0, HLE, AutomationBench, DeepSearchQA. Chart: Anthropic.
The part that caught my attention
Benchmarks are benchmarks. The examples of the model's judgment are more
interesting, because they are about agents that finish work instead of stalling:
🔹 Given a drawing of a machine part with no way to actually see the image, Opus 5 wrote its own computer vision pipeline to pull the geometry from raw pixels, then rebuilt the part in FreeCAD. Anthropic says no competing model solved it in five attempts.
🔹 Handed a real bug in a popular open source package manager, it found the root cause and fixed an edge case the community's own patch had missed. A competing model patched only the symptom and declared it done.
🔹 An engineer built a market data feed for a new exchange in one session. With no live feed to test against, Opus 5 wrote its own test harness to validate the parsing.
That is the shift: less "generate code and hope", more "verify, iterate, and check
its own work before handing it back".
Why it matters if you build with this
Cost per useful task is dropping fast. Near-frontier quality at half the price
changes what you can afford to run in a loop, which is exactly what long-running
agents need. If you were priced out of putting a strong model in an autonomous
pipeline, that math just moved.
One honest caveat: these are Anthropic's own reported numbers on their own harness.
Treat them as a strong signal, not gospel, until independent results land. And if
your work touches offensive security, note that it stays intentionally behind
Mythos 5 there.
On OSS-Fuzz, Opus 5 finds vulnerabilities nearly as well as Mythos 5 (left) but is far behind on writing exploits (right), by design. Chart: Anthropic.
Getting started
Opus 5 is live today on all platforms and on the API as claude-opus-5, at
$5 / $25 per million tokens. There is also a Fast mode at roughly 2.5x speed, plus
two beta additions: mid conversation tool changes without invalidating the prompt
cache, and automatic model fallbacks on the API.
All charts above are from Anthropic's official announcement, shown here for
commentary and reporting. Source: Introducing Claude Opus 5
If you have tried it: is the "it checks its own work" behavior holding up in your
real projects, or still hit and miss? Curious what people are seeing.
Top comments (0)