XNews

Anthropic Releases Claude Opus 5 with Competitive Pricing and Benchmarks

Anthropic Releases Claude Opus 5 with Competitive Pricing and Benchmarks

Anthropic released Claude Opus 5, matching Opus 4.8's $5/$25 per million token pricing and half of Fable 5, with near-frontier intelligence. It doubled Opus 4.8 on Frontier-Bench, nearly matched Fable 5 at half cost on CursorBench, and surpassed Fable 5 on OSWorld 2.0 at one-third cost. Now default on Claude Max and API.

Key insights

Pricing Parity with Opus 4.8, Half of Fable 5

Claude Opus 5 launches at $5 per million input tokens and $25 per million output tokens, exactly the same as Opus 4.8 and half the price of the flagship Fable 5. Anthropic positions Opus 5 as delivering near-frontier intelligence at significant cost savings, making advanced capabilities more accessible to developers who previously faced high API bills.

Benchmark Scores Show Dramatic Gains Over Opus 4.8

Opus 5 achieved more than double Opus 4.8’s score on Frontier-Bench v0.1 while lowering per-task cost. On CursorBench 3.2 at maximum effort, it came within 0.5% of Fable 5’s peak at half the cost. It scored three times the second-place model on ARC-AGI 3, and on OSWorld 2.0 it beat Fable 5’s best result at just over one-third the cost. Organic chemistry and protein tasks improved 10.2 and 7.7 percentage points respectively over Opus 4.8.

Real-World Task Examples Demonstrate Autonomous Problem-Solving

Opus 5 independently built a computer vision pipeline to reconstruct a 3D mechanical part from raw pixels in FreeCAD. It identified and fixed a root-cause bug in an open-source package manager that a community patch had missed. In a trading firm use case, it created its own test framework to validate market data parsing when live data was unavailable.

Improved Safety and Alignment Metrics

Anthropic reports Opus 5 as its most aligned model, with a composite misalignment score of 2.3 on automated behavioral audits, lower than Opus 4.8, Sonnet 5, and Fable 5. It also showed the lowest rate of deceptive behaviors. Biosecurity filters were relaxed compared to Fable 5, allowing previously blocked scientific requests to be handled by Opus 5 instead of an older model.

Cybersecurity: Strong Vulnerability Discovery, Limited Exploit Capability

Although not trained on offensive cybersecurity tasks, Opus 5’s general capability lifted its vulnerability discovery to near Mythos 5 levels. However, it remains far behind in generating working exploits. The safety classifier blocks binary-level scanning, penetration testing, and exploit generation, with 85% fewer interventions than Fable 5. Blocked requests fall back to Opus 4.8 in Claude products.

Agent-Friendly Beta Features: Persistent Cache and Auto-Degradation

Two beta API features target agent developers. First, changing tools mid-conversation no longer invalidates prompt caching, preventing cost spikes. Second, an automatic degradation option reroutes requests blocked by safety classifiers to an alternative model instead of failing outright, improving reliability for production agents.

User Relief: Finally a Usable Model After Opus 4.8 Disappointment

Early user reactions describe Opus 5 as the usable model Anthropic needed to ship. One commenter noted that continuing with Opus 4.8 would have made the Claude Max subscription unjustifiable due to poor performance. The new model is seen as restoring value to the premium tier, with multiple benchmark results exceeding those of the more expensive Fable 5.

Possible Benchmark Chart Inconsistency Spotted

A user noticed a red-highlighted comparison in Anthropic’s published scores where 53.4 appeared to be marked as greater than 53.5, raising questions about the accuracy or presentation of certain benchmark figures. The post did not indicate whether this was a confirmed error or a visualization oversight.

Source posts