Long Live the New King: Can Your Team Afford to Keep Claude Fable 5.2?

Anthropic released Claude Fable 5.2 on November 9, and the model scored 65 in Max reasoning mode on the November 10 Artificial Analysis Intelligence Index, ahead of Gemini 4 Carbon High at 60, Anthropic's own Opus 5.5 Max at 59, GPT-6 Astra Max at 57 and the previous Fable 5.1 Max at 49. Anthropic priced Fable 5.2 at $8 per million input tokens and $40 per million output tokens, 20% below Fable 5.1's $10 and $50, but Artificial Analysis estimated Fable 5.2 Max at $26.54 per Intelligence Index task versus $10.04 for Opus 5.5 Max. Anthropic reported Fable 5.2 at XHigh scored 49.1% on the agentic coding test FrontierCode 2.0, up from 32.4% for Fable 5.1, while Vals AI's science index placed GPT-6 Astra ahead of Fable 5.2 in November 9 results.

Anthropic has released Claude Fable 5.2, and a new king has taken the throne. In its Max reasoning mode, the model leads the Artificial Analysis Intelligence Index https://artificialanalysis.ai/models/releases/claude-fable-5-2 , ahead of competitors from Google and OpenAI as well as Anthropic’s own Opus 5.5. But there is a price for that crown. Fable 5.2 Max costs substantially more per benchmark task than its closest rivals, raising a question for every team considering it: can you afford to keep the king working? On November 9, Anthropic introduced Fable 5.2 for demanding technical work. In the November 10 Artificial Analysis snapshot, the model scored 65 with Max reasoning. Its nearest non-Fable competitor in the selected chart, Gemini 4 Carbon High, scored 60 . Opus 5.5 Max scored 59 , GPT-6 Astra Max 57 , and the previous Fable 5.1 Max 49 . The difference is not just one coding exercise. The independent index combines nine evaluations across coding, science, knowledge work, and general reasoning. Its score gives us a broader comparison of the tested configurations, though it cannot predict the winner of every individual task. One of the evaluations included in the index is FrontierCode 2.0, an agentic coding test. Anthropic reports that Fable 5.2 at XHigh scored 49.1% , up from 32.4% for Fable 5.1 at XHigh. The evaluation looks at whether changes to open-source projects would be accepted by maintainers. Anthropic reports 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra, with settings as described in its release announcement https://www.anthropic.com/claude-fable-and-mythos-5-2 . These are the company’s reported results on a test that also forms part of the index, not an entirely separate category of evidence. A separate evaluator also puts Fable at the top of the Vals Index https://www.vals.ai/benchmarks/vals index , which combines agentic tasks in finance, coding, legal work, and tax. In its November 9 results, Fable scored 71.31% , followed by Sonnet 5.5 at 70.85% and Gemini 4 Carbon at 70.45% . The top scores are close in this index. The two indexes use different mixes of work but overlap on Terminal-Bench 4.0, so they are not entirely separate sets of tests. There are other ways to measure a model. In Vals AI’s science index https://www.vals.ai/benchmarks/science index , GPT-6 Astra placed ahead of Fable 5.2 in the November 9 results. That does not erase Fable’s lead on Artificial Analysis. It tells us what the crown means: Fable came first in one broad set of tests, while a different set can produce a different winner. When Anthropic announced Fable 5.2, its official API price came with a pleasant surprise. The new model costs $8 per million input tokens and $40 per million output tokens . Fable 5.1 cost $10 and $50 , so the new version’s standard input and output rates are 20% lower. Anthropic’s pricing details https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2 tell an encouraging story: a stronger model at a lower price per token. How does that compare with the other contenders? These are API rates, not subscription prices. Cache and processing options have their own rates. For Astra, OpenAI charges more for the entire request when the prompt exceeds 272,000 input tokens. Google’s Carbon prices are announced rates for a limited rollout; they are not a promise of general access. So, based on the price per token, Fable 5.2 is cheaper than both its predecessor and Astra, while costing twice as much as Opus. But the price of a token is only one part of the bill. The other is how many tokens the model uses to finish the work. This is where the Artificial Analysis cost chart makes the story more interesting. It estimates the cost of running the evaluated tasks using the tokens consumed and their prices. In its November 10 snapshot, Fable 5.2 Max cost an average of $26.54 per Intelligence Index task . Opus 5.5 Max came to $10.04 and GPT-6 Astra Max to $6.24 . Fable Max was more than twice as expensive as Opus and more than four times as expensive as Astra in this comparison. That $26.54 is not the price of asking Claude a question. Artificial Analysis calculates a weighted average across its benchmark tasks, using the tokens consumed and their input, output, and caching prices. A real team’s cost depends on the work it sends, the reasoning setting, context, caching, and how often the model is used. Its methodology explains the calculation. https://artificialanalysis.ai/methodology The volume of work behind those numbers is substantial. Artificial Analysis reports 275 million output tokens for Fable 5.2 Max across the Intelligence Index evaluations, compared with 54.3 million for GPT-6 Astra Max. Those are totals across the evaluations, not measurements for one task or one conversation. They are also not the weighted per-task figures used to calculate the cost above, so the totals alone cannot tell us exactly how much each model’s output contributed to that price difference. The underlying point is that a lower rate per token does not automatically mean a lower bill: the number of tokens used matters too. For the index’s cost calculation, the mix of input, output and cached tokens all contributes. Fable’s results https://artificialanalysis.ai/models/releases/claude-fable-5-2 and Astra’s results https://artificialanalysis.ai/models/gpt-6-astra show the measured totals. A more capable model may justify a higher bill on difficult work. But the numbers raise a practical question: does a team need the most expensive setting for every job, or can it get what it needs with less reasoning effort? Fable 5.2 lets developers choose how much effort the model can spend thinking. Max is the setting that placed first in the index. High asks the same model to work at a lower level of effort. More effort may help with difficult work, but it can also consume more time and tokens. In the same Artificial Analysis snapshot, Fable 5.2 High scored 62 on the index, compared with Max’s 65. Its weighted average cost was $9.08 per task , about a third of Max’s $26.54. Artificial Analysis labels this tested configuration “High, Default Fallback” https://artificialanalysis.ai/models/claude-fable-5-2-high ; the figures describe that configuration. Look back at Opus 5.5 Max: 59 points for $10.04 per task . In this index, Fable High scored three points higher while costing a little less per benchmark task. Those are aggregate results, not proof that Fable High will outperform Opus on every coding problem or cost less in every workflow. They make High a particularly interesting setting to try. That’s particularly interesting because Fable’s uncached input and output rates are twice those of Opus. In this benchmark mix, Fable High’s measured cost per task is still lower. The token mix and caching help determine the total, so the same relationship may not hold for a different task. Anthropic offers another useful comparison in its FrontierCode results: Fable 5.2 High scored 45.3% , close to Opus 5.5 Max at 45.5% , at roughly half the cost per task in that evaluation. FrontierCode is one component of the broader Intelligence Index, so these are overlapping comparisons rather than fully independent evidence. They suggest that the reasoning setting deserves as much attention as the model name. In this index, Max’s three-point lead over High comes with an additional $17.46 per benchmark task . That is a large premium for a team to consider, even when the model itself is the same. The numbers do not tell us what those three points mean for a particular problem. They do suggest a reason to look at High first, and to ask whether a difficult task warrants paying more for Max. My first impression is that Fable 5.2 deserves its place at the top. Its results are exciting, especially the lead it has established over the other models in the Artificial Analysis index. But at the price Artificial Analysis measured, I expect to use Max almost never. High is the version I want to try: it scores above Opus 5.5 Max on that index, while its measured cost per task is a little lower. I want to see whether that tradeoff holds for the work I actually give it. Long Live the New King: Can Your Team Afford to Keep Claude Fable 5.2? https://pub.towardsai.net/long-live-the-new-king-can-your-team-afford-to-keep-claude-fable-5-2-2f669161865b was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.