Anthropic Ships Claude Opus 5, Claims Near-Fable Performance at Half the Price

Anthropic released Claude Opus 5 on July 24, 2026, in an unusual Friday launch, marketing it as delivering performance close to its pricier Claude Fable 5 model at roughly half the cost. The company's own messaging is measured, saying Opus 5 "comes close" to Fable 5 rather than claiming outright superiority — despite several third-party benchmarks showing Opus 5 ahead on specific tasks.

Benchmark results

Epoch AI's Epoch Capabilities Index (ECI) puts Claude Opus 5 at 159, versus 161 for Claude Fable 5 — a narrow gap. On the software-engineering-specific SWE-ECI, the two models tie at 161, according to Epoch AI. Critics on social media, including a user posting as @scaling01, argued the ECI result undersells Opus 5, noting the score is only one point above the prior Opus 4.8 despite what they describe as substantially better real-world performance.

Artificial Analysis reported a more favorable picture for Opus 5, stating the model leads its AA-Briefcase agentic knowledge-work benchmark, outperforming Claude Fable 5 by nearly 150 Elo points while reducing cost per task by 20%. Artificial Analysis also lists Opus 5 as the new leader on its Intelligence Index. These are Artificial Analysis's own benchmark claims and have not been independently replicated in this report.

A separate anomaly was flagged by a user posting as @jerhadf: on the FrontierCode benchmark, Opus 5 scored better at medium inference effort than at high effort — the opposite of the pattern seen on other evaluations. Neither Anthropic nor Epoch AI has published an explanation for this result.

Availability and early reactions

Nous Research added Opus 5 to its Nous Portal with a 20% discount applied across all hosted models, according to a post from a user identified as @witcheer — a distribution detail, not a capability claim. Arena said it had posted first impressions of Opus 5 and that real-world leaderboard rankings based on community usage would follow.

Anecdotal reports from developers were largely positive on agentic tool use. Developer Mikhail Parakhin said Opus 5 beat Fable 5 in his own testing "for math and everything, really," particularly when using best-of-n sampling. Another user, posting as @abacaj, described Opus 5 successfully navigating a browser to cancel a ChatGPT Pro subscription unprompted, calling the tool-use performance notable. These remain isolated demonstrations rather than systematic evaluations.

No absolute pricing figures for either Opus 5 or Fable 5 have been disclosed; Anthropic's positioning is relative ("half the price" of Fable 5), and context window size was not specified in available materials.

What this means

The substance of this release is a cost story more than a raw-capability story: Opus 5 appears to trade a small amount of benchmark headroom against Fable 5 for a roughly 20% reduction in cost per task, based on Artificial Analysis figures. The one-point ECI gain over Opus 4.8 versus larger perceived qualitative gains — a gap flagged repeatedly by independent commentators — is the most consequential open question. It suggests current static benchmarks may be poorly calibrated to capture improvements in agentic, tool-using workloads like browser automation, which is where most of the positive anecdotal reaction concentrated. Until Arena and other community leaderboards publish real-world rankings, buyers evaluating Opus 5 against Fable 5 or GPT-5.6 Sol should treat the Epoch and Artificial Analysis numbers as directionally useful but not final.