Claude Opus 5 Beats Fable 5 at Half the Price with Superior Coding

Anthropic released Claude Opus 5 on a Friday, keeping the same pricing as Opus 4.8 ($5 per million input tokens, $25 per million output tokens). Despite identical costs, Opus 5 reaches the performance level of Fable 5 while costing roughly half of Fable 5’s price.

On the Frontier‑Bench v0.1, Opus 5 achieved 43.3% while Fable 5 scored 33.7% and Opus 4.8 only 21.1%, more than doubling the previous best. On CursorBench 3.2, Opus 5 running at max effort fell just 0.5 percentage points behind Fable 5’s peak, yet used about 40% fewer tokens. In the Artificial Analysis Coding Agent Index, Opus 5 paired with Claude Code tied for first place.

A developer prompted the model with a single sentence—“Build a Wolfenstein 3D style FPS game in three.js”—and Opus 5 instantly generated a playable 3D game, something earlier models could not accomplish.

Beyond coding, Opus 5 performed strongly on knowledge‑work and reasoning benchmarks. In the Artificial Analysis Intelligence Index it scored 61 points, narrowly beating Fable 5 (60) and GPT‑5.6 Sol (59). On the AA‑Briefcase knowledge‑work benchmark it earned 1720 Elo, outpacing Fable 5 by 146 points, and on GDPval‑AA v2 it exceeded Fable 5 by over 100 points.

A notable result comes from ARC‑AGI‑3, which measures novel problem‑solving ability: Opus 5 obtained 30.2%, more than three times the second‑place GPT‑5.6 Sol (7.8%) and far above Opus 4.8’s 1.5%.

On OSWorld 2.0 (computer‑use benchmark) Opus 5 topped the chart with 70.6% while costing only one‑third of Fable 5’s best result. In AutomationBench (end‑to‑end business‑process benchmark) its pass rate was about 1.5× that of the next best model.

However, Opus 5 is not universally dominant. On DeepSWE v1.1, GPT‑5.6 Sol led with 72.7% versus Opus 5’s 68.8%. In legal and health domains Fable 5 retains a slight edge. Opus 5’s AA‑Omniscience knowledge‑accuracy improved by 7 points over Opus 4.8, but hallucination rate rose by 14 points to 50%, indicating a trade‑off where the model prefers answering over refusing when uncertain.

Speed-wise, Opus 5 outputs 56.3 tokens/second, below the class median of 73.6. First‑token latency averages 64.52 seconds, primarily due to thinking time. Enabling Fast mode multiplies speed by 2.5× at the cost of doubling the price.

Opus 5 introduces five effort levels (low, medium, high, xhigh, max). Moving from low to max expands the GDPval‑AA v2 score by 407 Elo points while token consumption varies by roughly eightfold, allowing users to balance cost and performance per task.

The average cost per Intelligence Index task is $2.03, lower than Fable 5’s $2.75. At higher effort levels Opus 5 can surpass Opus 4.8 and Sonnet 5 at lower cost.

In safety evaluations, Anthropic claims Opus 5 is its most aligned model. In automated behavior audits it received an inappropriate‑behavior score of 2.3, better than Opus 4.8, Sonnet 5, and Fable 5. Its safety classifier is about 85% more permissive than Fable 5’s, allowing source‑code vulnerability detection while blocking binary scanning and penetration testing. Biological‑security safeguards remain similar to Opus 4.8 and avoid the over‑blocking seen in Fable 5.

The context window stays at 1 million tokens. Opus 5 is now the default model for Claude Max and the strongest model for Claude Pro across all platforms. The API endpoint is provided under the name claude-opus-5, and two beta features have been added: tool‑change within conversations and automatic API fallback.

In short, K3’s engineering has produced a model that delivers near‑Fable 5 capabilities at a lower price; for users focused on coding or agent‑style tasks with limited budgets, Opus 5 currently appears to be the most cost‑effective choice.