Introducing Claude Opus 5

While the whole AI community is still staring at the highest benchmark scores, Anthropic has turned around and released a 'half-price valedictorian' — Claude Opus 5. It doesn't outshine the flagship Fable 5 in raw IQ, but it sends a series of signals that force us to rethink what a model should look like.

Context: The model race enters a new phase

For the past year, model companies have been competing on parameters and benchmark peaks. But there's an awkward reality: the top-scoring API models are often too expensive for complex agent scenarios. Agents run for long periods and make numerous calls, so costs leak like a faucet. Opus 5's positioning is crystal clear: half the cost of Fable 5, yet nearly as smart on coding and knowledge work, and in some agent tasks even surpasses it thanks to its autonomy. This is a 'pragmatism manifesto' for developers.

Dissection: Not just smarter, but better at getting things done

The performance numbers are strong: Opus 5 tops all models on Frontier-Bench v0.1; scores three times higher than the next best on ARC-AGI 3; and on the computer-use benchmark OSWorld 2.0, it beats Fable 5's best result at just over a third of the cost. But what truly impresses is its 'ability to act'.

One described case: given a drawing of a machine part and asked to write code to reconstruct a 3D model, but with no way to directly view the drawing. Opus 5 didn't complain about missing tools—it wrote its own computer vision pipeline to extract geometry from raw pixels and completed the reconstruction. No other model with the same setup succeeded after five attempts. Another case saw it fix a real edge-case bug in a popular open-source package manager, from root cause analysis to patch, seamlessly.

This reveals a deeper shift: models are moving from 'answer machines' to 'problem solvers'. They no longer just react, but actively seek paths and build missing tools. Opus 5 is explicitly designed to 'verify its work and iterate carefully until it succeeds'. This kind of self-correction and engineering muscle is exactly what long-running agents need.

Trend insight: The 'agent counterattack' of mid-tier models

Many assume only cutting‑edge models can power good agents. Opus 5 suggests a different paradigm: if a model is cheap enough and possesses enough tool-using and self-iterating ability, it can compensate for a slight intelligence deficit with 'more diligence'. It's like human expert teams—a slightly less experienced but more meticulous, tool-savvy engineer can be more reliable than a lone genius.

Moreover, the introduction of 'effort' settings—letting users dial between intelligence and cost/speed—is an important engineering mindset. It acknowledges real-world budget and latency constraints, allowing the same model to flexibly fit scenarios from instant chat to background heavy lifting. This may become an industry standard.

Practical value: How developers should use it

If you're building multi-step, long-running agent applications (like automated testing, code review, research analysis), Opus 5 offers a cost-effective new choice. You can use high-effort for the hardest subtasks, then dial down to save cost. For enterprises, its cost control and autonomy mean more predictable agent budgets and less human intervention.

Contrarian view: The best model isn't always the 'smartest'

In the benchmark arms race, we obsess over top scores. But Opus 5 reminds us that in real workflows, 'getting things done' matters more than being the smartest. A model that can proactively build tools and verify its work is commercially more valuable than a higher-IQ but budget-busting one. With this move, Anthropic has shifted the agent competition from 'who has the higher IQ' to 'who is better at actually doing the job'.

In summary, Opus 5 isn't just a model update—it may be a turning point in how we design models for the agent era. It's no longer about chasing the absolute strongest, but about delivering the most reliable, autonomously capable assistant within a budget. For developers building agents day in and day out, this is the real 'version answer'.