Anthropic's Claude Opus 5 Matches Flagship Intelligence at Half the Price

Claude Opus 5 is out today, and Lovable is one of the first platforms to ship it. Anthropic's new flagship Opus model delivers near-Fable-5 intelligence at half the price, with the strongest coding and knowledge-work scores the company has ever published on a generally available model.

Opus 5 is the fourth model Anthropic has shipped in under two months, following Mythos 5, Fable 5, and Sonnet 5 in June — an unusually compressed release cadence even by AI industry standards.

The gap Opus 5 fills

The Opus tier had been caught in an awkward position. Sonnet 5 arrived at the end of June performing close to Opus 4.8 at a fraction of the cost, while Fable and Mythos sat above it. Enterprises on Max, Team, and Enterprise plans needed a capable middle ground that didn't carry Mythos-class pricing. Opus 5 is that answer: more capability than its predecessor at the same cost, closing most of the gap to Fable 5 without closing the gap in price.

What the benchmarks actually show

On coding and knowledge-work evaluations, Opus 5 sets new state-of-the-art marks, though it trails Mythos 5 on cybersecurity tasks. The numbers across key benchmarks:

• Frontier-Bench v0.1: Opus 5 surpasses all other models and more than doubles Opus 4.8's performance at a lower cost per task.
• CursorBench 3.2: At max effort, Opus 5 performs within 0.5% of Fable 5's peak score at half the cost per task.
• ARC-AGI 3 (novel problem-solving): Opus 5's score is three times higher than the next-best model.
• OSWorld 2.0 (computer use): Achieves state-of-the-art visual UI interaction at 70.6%, covering software navigation, file management, and desktop tool operation.
• Zapier AutomationBench: Pass rate runs roughly 1.5x the next-best model at the same cost per task. Even at its lowest effort setting, Opus 5 passes more tasks than any competing model.

On scientific work, Opus 5 shows meaningful gains over Opus 4.8 across every life sciences evaluation Anthropic ran. The largest jumps are in organic chemistry, up 10.2 percentage points on spectroscopy inference, and protein function prediction, up 7.7 percentage points.

The effort dial: a cost lever for builders

One of the most practically useful additions is a tunable effort setting. Developers can set the model to low, medium, or high effort depending on the complexity of the task, letting simpler requests run cheap while reserving full compute for hard reasoning steps. Think of it as a throttle built into the model itself.

A Fast mode is also available. Standard pricing is $5 per million input tokens and $25 per million output tokens. Fast mode runs about 2.5 times quicker than the default but costs twice as much.

What Lovable measured

Lovable co-founder Fabian Hedin is quoted directly in Anthropic's launch post: Opus 5 is up 22% over Opus 4.7 on their hardest agentic coding tasks, and it's steadier, with far less variance run to run. Lovable built a scaffolding layer on top of the Claude API covering the frontend, backend, auth, database, deployment infrastructure, visual builder interface, and workspace management.

For a platform like Lovable, consistency matters as much as peak performance. A model that scores 90% on average but swings wildly between attempts produces a worse product experience than one that scores 85% reliably. Opus 5's reduced run-to-run variance is the kind of improvement that shows up in user satisfaction before it shows up in any benchmark table.

Where it falls short

Opus 5 is not Anthropic's most capable model. Anthropic intentionally avoided training it on cyber exploitation tasks, and while the model improved on vulnerability finding — approaching Mythos 5 territory — it remains substantially behind on developing exploits. Teams doing serious offensive security research should stay on Mythos 5 with CVP access.

Self-correction and agentic judgment

Anthropic says Opus 5 is better at checking its own work and retrying when an initial approach fails. The canonical example from their launch: given an image of a machine part and no direct way to view it, Opus 5 was asked to recreate the part as a 3D FreeCAD model. Rather than stopping, the model built its own computer vision pipeline to extract geometry from the image's raw pixels, then reconstructed the part. Anthropic reports it succeeded repeatedly under conditions where competing models failed after five attempts.

Early-access customers across finance, legal, and genomics report similar behavior: the model pushes back on flawed designs, cross-checks its own results using independent methods, and stays on track through long multi-step analyses without needing to be re-prompted.

Safety and alignment

Anthropic says Opus 5 scores 2.3 on their automated behavioral audit for overall misaligned behavior, the lowest of any recent model, and shows a lower rate of deceptive behavior than its predecessors.

Two new API-level features ship alongside the model:

• Mid-conversation tool changes: Developers can swap which tools Claude can access mid-conversation without invalidating the prompt cache, a practical improvement for complex agentic workflows.
• Automatic fallbacks: Requests flagged by safety classifiers on Opus 5 or Fable 5 can now automatically route to a safer model rather than being blocked outright.

How to access it

Claude Opus 5 is rolling out across all of Anthropic's platforms as the new default on Claude Max and the strongest model available on Claude Pro. On the API, call it as claude-opus-5 at $5/$25 per million input/output tokens. Lovable users get it automatically, no configuration needed.

For teams building agentic pipelines, the effort setting is worth experimenting with right away. Routing simpler subtasks to low-effort mode and reserving high or xhigh for hard reasoning steps can cut per-session costs without sacrificing output quality where it counts. The official prompting guide covers how to get the most out of each effort level.