Claude Opus 5 has entered the frontier AI race with a clear message: performance gains matter, but cost and reliability now matter just as much. Across the benchmarks cited in the source, the model often matches or beats Claude Fable 5 while coming in at a lower price on several important comparisons.
The picture is not one-sided. Opus 5 looks especially strong in analytical quality, knowledge work, and coding tasks, yet its factual reliability remains a concern because it answers more often when uncertain.
Benchmarks Put Claude Opus 5 Near the Front
Artificial Analysis ranks Claude Opus 5 as the most capable model in its Intelligence Index, where it scores 61. That index combines nine tests covering knowledge work, coding, scientific reasoning, and factual accuracy.
The score places Opus 5 just ahead of Claude Fable 5 at 60, GPT-5.6 Sol at 59, Kimi K3 at 57, and Claude Opus 4.8 at 56. Artificial Analysis worked with Anthropic to test the model before its public release.
For coding, Opus 5 also performs at the top tier. At "xhigh" with Claude Code, it shares first place on the Artificial Analysis Coding Index, which evaluates how well models handle programming tasks independently, including finding and fixing bugs.
On Terminal-Bench v2.1, Opus 5 scored 89 percent at "max." That benchmark tests agents as autonomous engineers in real terminal environments, and the result matches the previous leader GPT-5.6 Sol.
Science And Accuracy Tell A More Mixed Story
Opus 5 also posts strong results on scientific reasoning benchmarks. On Humanity's Last Exam, described in the source as a very difficult knowledge test across many academic fields, it scored 53 percent. That ties Claude Fable 5.
On CritPt, a physics benchmark from researchers at Argonne National Laboratory and UIUC, Opus 5 again matches Fable 5. However, it trails GPT-5.6 Sol, GPT-5.5 Pro, and GPT-5.6 Terra on that test.
The more serious weakness is factual accuracy. On AA-Omniscience, which tests whether a model's knowledge claims are accurate, Opus 5 improved by 7 points over Opus 4.8 but still sits behind Fable 5.
The issue is not only what the model knows. According to the source, Opus 5 responds more often when uncertain, which pushes its hallucination rate up 14 points to 50 percent. That makes the model's reliability a central concern for high-stakes applications, even when its benchmark scores are strong.
Cost Changes The Comparison With Fable 5
The price-performance story is one of the most important parts of the Opus 5 launch. The average Intelligence Index task costs $2.03 with Opus 5. That is below Claude Fable 5 with fallback at $2.75, though it remains above Opus 4.8 at $1.80 and Sonnet 5 at $1.53.
At the "high" and "xhigh" performance levels, Opus 5 beats both Opus 4.8 and Sonnet 5 while costing less. That makes the reasoning tier choice more than a technical detail. It directly affects whether users get better output for the money.
Vals.ai tested Claude Opus 5 across all five reasoning tiers using Vibe Code Bench. The model scored 76.7 percent at "low," 82 percent at "medium," and 89.8 percent at "high." Performance then dipped slightly at the two highest tiers, with "xhigh" at 88.3 percent and "max" at 88.4 percent despite higher costs.
Vals.ai found that the highest tiers tend to create more complex solutions that contain errors more often. The "high" tier produced simpler solutions that met requirements more reliably.
Terminal-Bench 2.1 shows a similar pattern. The "high" tier beats "max" because the model spends more time on each attempt at the top tier, leaving fewer attempts within the time limit. That aligns with Anthropic's guidance, which sets "high" as the default tier in both the API and Claude Code.
Token pricing remains $5 per million input tokens and $25 per million output tokens. Cache writes cost $6.25 per million tokens with a five-minute lifetime, while cache hits cost $0.50 per million tokens.
Knowledge Work Is Where Opus 5 Looks Strongest
Opus 5 stands out most clearly on AA-Briefcase, a benchmark built around office-style tasks such as writing research reports, building presentations, and analyzing spreadsheets from thousands of input files. The benchmark scores correctness, analytical quality, and presentation quality, then combines them into an Elo rating.
At max reasoning, Opus 5 reaches an Elo of 1720. That is 146 points ahead of Claude Fable 5 at 1574. Its "max," "xhigh," and "high" tiers take the top three spots, while Anthropic models including Fable 5, Sonnet 5, and Opus 4.8 hold most of the top-10 positions.
The cost per AA-Briefcase task also falls. Opus 5 at "max" costs $17.79, down from $22.30 for Fable 5. The "xhigh" variant costs $14.26, and "high" costs $10.41, which is less than half of Fable 5 while still beating it in the Elo ranking.
The model's biggest gains are in analytical quality. At "max," Opus 5 reaches an Analytical Quality Elo of 2016, almost 300 points ahead of Fable 5. Its Rubric Pass Rate is 58 percent at "max," 57.2 percent at "xhigh," and 56 percent at "high."
Presentation quality is less dominant. Opus 5 scores a Presentation Elo of 1628, while GPT-5.6 Sol at "max" reaches 1666. More reasoning also takes more time: at "max," Opus 5 needs over 36 minutes per task and averages 103 passes, compared with Opus 4.8 at 24 minutes and 55 passes.
A Tight Race, Not A Clear Breakaway
Epoch AI's results point to the same broader conclusion: the frontier model race remains close. The research institute gives Claude Opus 5 an overall Epoch Capability Index score of 159, just below Fable 5 at 161. GPT-5.6 Sol leads in that category.
On software engineering benchmarks, measured by SWE-ECI, Opus 5 ties Fable 5 at 161 and outperforms GPT-5.6 Terra and Claude Opus 4.8. GPT-5.6 Sol leads there as well.
Taken together, the benchmark results show that no single model has separated decisively from the field. Opus 5 raises the bar in several places, especially for analytical and knowledge-heavy work, but the tradeoffs are real: cost, reasoning tier, runtime, and hallucination risk all affect whether it is the right model for a given task.