Claude Fable 5.2 scores 65 on Artificial Analysis index but costs about 2x Opus 5.5 per task

Anthropic announced Claude Fable 5.2, its top-tier model, on November 9 (US time). Japanese outlet ITmedia AI+ then compared it with Claude Opus 5.5 and Claude Sonnet 5.5 on two white-collar tasks. The result: Fable 5.2 produced the most audience-ready output in one task but was the most expensive option in both. The article is in Japanese; this report translates its findings.

What Anthropic claims

Anthropic claims Fable 5.2 beats Opus 5.5 on benchmarks including coding and science. The company still recommends starting with Opus 5.5 for most coding and knowledge work, according to ITmedia.

Third-party scores

Artificial Analysis, an independent evaluator, rates Fable 5.2 (max) at 65 points on its Intelligence Index, first place, ahead of Opus 5.5 (max) at 59. With thinking effort lowered to "high", Fable 5.2 holds 62 points on the index, still above Opus 5.5 (max). Artificial Analysis says that, depending on settings, Fable 5.2 can deliver higher performance at lower cost than Opus 5.5.

On AA-Briefcase, which measures knowledge work such as building spreadsheets and presentations from large document sets, Fable 5.2 (max) scores 1785. That is slightly below Opus 5.5 (max) at 1800 and Sonnet 5.5 (max) at 1813.

How ITmedia tested

Reporter Tomoki Murata ran each task once per model on November 10, using the Claude Code CLI, a fresh session per task, and thinking effort set to max. OpenAI's GPT-6 Astra checked outputs against source material for accuracy. Cost was calculated from CLI-reported token counts times API prices at 1 dollar = 158 yen. ITmedia notes that results are single runs and may change with updates.

Task 1: slide deck from a white paper. Each model received the 268-page Ministry of Internal Affairs and Communications 2026 Information and Communications White Paper (Reiwa 8 edition) and was asked to build a PowerPoint deck that non-IT staff could absorb in 15 minutes. All three produced editable charts, speaker notes on every slide and citations to the paper, with no obvious transcription errors in key figures.

• Fable 5.2 delivered 13 presentation slides plus a glossary, showed a time allocation on the cover, and wrote the shortest speaker notes. Sonnet 5.5 wrote the most detailed notes, too long for 15 minutes.
• On a survey finding (Japanese firms citing efficiency gains at 61.6% versus quality improvement at 32.2%), Opus 5.5 and Sonnet 5.5 headlined Japan as stopping at efficiency. Fable 5.2 used a more measured framing.
• Fable 5.2 took about 1.6 times as long as Opus 5.5 and cost about 2 times as much as Opus 5.5 and about 3.6 times as much as Sonnet 5.5. Its output tokens were about 1.5 times those of Opus 5.5.

Task 2: pre-interview research on Anthropic. Web access was limited to official sites and major media. Token and cost figures include the helper model Claude Haiku 5.5. All three covered Anthropic's IPO roadshow and the Fable 5.2 announcement.

• Fable 5.2 gave the longest answer, with extra detail on Japan-related moves, flagged what it could not verify, and proposed Japan-specific questions.
• Sonnet 5.5 was shortest and the only model to state whether each source was read in full or only seen as a search summary.
• Fable 5.2 finished about 2 minutes faster than Opus 5.5 and cost about 1.8 times as much as Opus 5.5 and about 5 times as much as Sonnet 5.5.

What this means

In ITmedia's reading, Fable 5.2 earns its premium where the goal is hard-to-quantify quality, such as a deck structured to fit a 15-minute talk. For routine research, the reporter judged the three models close in practical usefulness and considered lower-tier models sufficient. The reporter's view is that Opus 5.5 should be the default, Sonnet 5.5 suits quick overviews, and Fable 5.2 fits cases where thoroughness and accuracy justify the cost. The reporter says these impressions are partly subjective.

The Artificial Analysis data points the same way: Fable 5.2 leads on its overall index but not on AA-Briefcase. Lowering Fable 5.2's thinking effort is an option worth testing before switching models, a step Anthropic itself suggests in its model selection guidance. API pricing for Fable 5.2 and its context window are not stated in the article.

Source: ITmedia AI+ (Japanese), published November 12, 2026.