Anthropic drops Claude Opus 5 with near-flagship power at half the cost

Anthropic just dropped Claude Opus 5, bringing near-flagship reasoning to everyday agentic tasks at half the price of Claude Fable 5.

Matching the $5 per million input and $25 per million output token price tag of Opus 4.8, Opus 5 immediately takes over as the default engine on Claude Max and the top performer on Claude Pro. It is also available right now on the Claude API and Amazon Bedrock, complete with zero data retention options for enterprise deployments.

Building its own tools under pressure

Where Opus 5 truly flexes is in autonomous execution and self-correction.

During pre-deployment tests, engineers handed the model a 3D part reconstruction challenge on Frontier-Bench, but intentionally denied it direct image-viewing access. Instead of failing, Opus 5 wrote a custom computer vision pipeline from scratch, parsed raw pixels to extract the geometry, and built the FreeCAD script. Competitors failed all five attempts under the same constraints.

Real-world coders are seeing similar grit. In one trading firm deployment, Opus 5 wired up a live exchange data feed in a single session—something earlier models could not finish even with detailed step-by-step human prompts. Finding no live feed to test against, the model independently coded a mock testing harness to verify its output.

On the benchmark board, Opus 5 put up a staggering 30%-plus score on ARC-AGI 3, tripling the score of the next-best model at abstract problem-solving. It also bested Fable 5 on OSWorld 2.0 computer use workflows at roughly one-third of the execution cost.

The fine print behind the charts

Independent testers have already spotted where the official marketing glosses over compromises.

Reviewers pointed out that Anthropic compressed price comparisons using log scales while graying out top rivals. On Deep Sweep, a critical measure for agentic software work, OpenAI's GPT-5.6 Soul actually retains the crown. Independent tracking from Artificial Analysis shows Opus 5 edging into first place overall by a hairline, but at nearly double the operating cost of GPT-5.6 Soul.

Other independent boards tell a mixed story. On Abacus AI's Live Bench, Opus 5 slipped to third behind Fable 5 and GPT-5.6 Soul. On the Valse index, it barely edged out the open-source Kimi K3. Even that dazzling ARC-AGI 3 score showed cracks: while it mastered classic logic puzzle formats, it lagged behind Opus 4.8 on novel games with completely unfamiliar mechanics.

Then there is Anthropic's trademark safety net. While cyber classifiers trigger 85% less often than on Fable 5, the model still aggressively blocks binary scanning, exploit crafting, and complex biology pipelines, falling back to Opus 4.8 when spooked.

Why it matters

For builders and enterprise teams, Opus 5 changes the math on long-running agents. Features like mid-turn tool swapping without busting prompt caches make it vastly more practical to keep complex coding bots running for hours without running up massive context bills.

We think Anthropic made a savvy tactical pivot here. Instead of chasing dangerous frontier capabilities where Mythos 5 operates, it repackaged Fable-tier agentic stamina into a predictable tier-two pricing model. It forces competitors to prove that their expensive flagships are truly worth the premium.

Still, it is not a clean sweep. With OpenAI maintaining an edge in pure agentic coding speed and open-weight models closing the reasoning delta, Opus 5 feels less like an untouchable breakthrough and more like a workhorse for teams ready to delegate messy, iterative engineering problems.