Third-party review: Opus 5 slightly beats Fable 5 on the intelligence index, cuts costs by 26%, but the hallucination rate jumps sharply

Independent evaluation firm Artificial Analysis released the score sheet for Claude Opus 5. In the Intelligence Index across 9 tests, Opus 5 scored 61, narrowly beating Fable 5’s 60 by 1 point. GPT-5.6 Sol scored 59, and Kimi K3 scored 57, closely following behind. Opus 5’s average cost per task is $2.03, 26% lower than Fable 5, and it also topped the rankings in two knowledge-focused evaluation categories.
(Background: Anthropic released Claude Opus 5! Performance nearly matches Fable 5’s price but is cut in half, making it the latest default model)
(Additional context: Anthropic no longer shows summaries of Claude’s thought process to prevent Fable 5 from being distilled)

Table of contents

Toggle

• First place by 1 point
• Paying more Tokens for performance
• Topping the leaderboard but stumbling on factual knowledge

Key takeaways

• Opus 5 scored 61 on the Intelligence Index across 9 tests, narrowly beating Fable 5’s 60 by 1 point
• Average cost per task is $2.03, 26% lower than Fable 5’s $2.75, alongside Claude Code tied for first in the Coding Agent index
• The shortcoming is factual knowledge: the AA-Omniscience hallucination rate jumped to 50%, which is 14 percentage points higher than the previous Opus 4.8

On July 24, Anthropic released Claude Opus 5, and the third-party score sheets followed. Independent evaluation firm Artificial Analysis ran Opus 5 through its Intelligence Index across 9 tests, ultimately giving it 61 points—exactly 1 point higher than its own strongest model, Fable 5—while pushing GPT-5.6 Sol’s 59 and Kimi K3’s 57 behind it. First place is locked in, but this top win doesn’t come with a huge margin.

First place by 1 point

The top four are clustered within 4 points, and Opus 5’s lead is so small it’s nearly within the margin of error. What truly separates the models is price. Opus 5’s average cost per task is $2.03, 26% lower than Fable 5’s $2.75—meaning you can buy the same level of intelligence for a cheaper price.

Detailed performance is what gives Opus 5 the confidence to take first place. In the simulated real-knowledge work evaluations GDPval-AA v2 and AA-Briefcase, it scored first in both. Paired with Claude Code, it also tied for first in the Programming Agent (Coding Agent) index, and in the Terminal-Bench v2.1 test for terminal operation capability, it scored 89%, roughly matching GPT-5.6 Sol. Cheap and capable—that’s exactly the selling point Anthropic highlighted on release day.

Paying more Tokens for performance

This time, Opus 5 offers 5 reasoning strength tiers, ramping from low all the way to max. The cost of this design is direct: moving from the lowest tier to the highest tier leads to about an 8x difference in the number of output tokens, and the GDPval-AA v2 score widens as well by 407 Elo.

In other words, how well the same model performs is largely in the hands of the user. If you’re willing to burn more Tokens, you get stronger performance; if you want to keep costs down, you can also proactively dial down the reasoning tier. By handing the “how much more intelligent do you want, and how much does it cost” knob back to users, a single model can serve both budget-conscious and cost-unconstrained needs at the same time.

Topping the leaderboard but stumbling on factual knowledge

Opus 5’s factual knowledge is still behind Fable 5. In AA-Omniscience, which tests knowledge breadth, its hallucination rate rises to 50%, which is 14 percentage points higher than the previous Opus 4.8. For a model that tops a composite intelligence benchmark, there’s a 50% chance it will present what it doesn’t know as if it were real—and that contrast isn’t easy to ignore.

On the same leaderboard, Opus 5 is both the first-place model in composite intelligence and a high-risk zone for hallucinations. Winning the scores and being trustworthy don’t happen at the same time.

Another detail not covered over by marketing language is that Opus 5’s value at the low reasoning tiers is still slightly worse than the GPT-5.6 series. In other words, only when you raise the reasoning strength does its advantage truly materialize—and that requires more Tokens. Third-party reporting brings that release-day line—“top-tier intelligence, commoner-friendly price”—back down to reality.

Common questions

What rank did Claude Opus 5 get in the Artificial Analysis evaluation?

Opus 5 topped the board with 61 points on the Intelligence Index across 9 tests. It led Fable 5’s 60 points, GPT-5.6 Sol’s 59 points, and Kimi K3’s 57 points, but the gap among the top four is only 1 to 4 points—meaning a narrow lead.

What are the clear weaknesses of Claude Opus 5?

Its factual knowledge is still behind Fable 5. In the AA-Omniscience test, its hallucination rate rises to 50%, which is 14 percentage points higher than the previous Opus 4.8, and its CP value at low reasoning tiers is also slightly worse than the GPT-5.6 series.