Anthropic blindfolded its new model and it built its own eyes
Claude Opus 5 wrote its own vision system to pass a test no rival could solve, and it runs at half the price of Anthropic's flagship.
Anthropic gave its newest model a drawing of a machine part and asked it to rebuild the thing in 3D.
Then it took away any way of looking at the drawing.
Claude Opus 5 wrote its own computer vision pipeline, pulled the geometry out of the raw pixels, and built the part.
Anthropic says it did it repeatedly, and that no competing model solved the task in five attempts.
That is the pitch for the model released on Friday, and it comes with a price tag that undercuts Anthropic's own flagship.
Opus 5 costs $5 per million input tokens and $25 per million output — identical to its predecessor Opus 4.8, and half what the company charges for Fable 5, the model that sat above it until now.
On Frontier-Bench, an agentic coding evaluation, Opus 5 scored 43.3 per cent against Fable 5's 33.7 and OpenAI's GPT-5.6 Sol on 34.4.
On CursorBench it landed within half a percentage point of Fable 5's best result at half the cost per task.
On OSWorld 2.0, a computer-use benchmark, it beat Fable 5's peak score at just over a third of the cost.
On ARC-AGI 3, which tests novel problem-solving, it scored roughly three times the next-best model.
The exception is cybersecurity, where Anthropic says it stays behind its Mythos-class models.
For anyone already running Opus 4.8, the jump is bigger than the version number suggests.
Anthropic says Opus 5 more than doubles 4.8's Frontier-Bench score at a lower cost per task, and beats it on every life sciences evaluation the company runs — by 10.2 percentage points on organic chemistry work like reading molecular structure from spectroscopy data.
Box measured an 8 per cent overall lift over 4.8, rising to 11 per cent on data analysis and 17 per cent on due diligence.
Legal AI firm Harvey said it matched 4.8's quality while generating 26 per cent fewer tokens.
The other examples Anthropic offers follow the same pattern as the blindfolded drawing.
Handed a bug in an open-source package manager, Opus 5 found the root cause that the community's own patch had missed.
An engineer at a trading firm used it to build a market data feed in a single session, and when there was no live feed to test against, it built its own test harness.
Anthropic also claims it as its most aligned model yet, scoring 2.3 on an internal misaligned-behaviour audit — the lowest of any recent Claude.
Opus 5 is now the default on Claude Max and the strongest model available on Claude Pro.
Anthropic has just made its own flagship a harder sell.