Claude Fable 5.2 Review: First on the Index, Last on Cost

Claude Fable 5.2 is first on Artificial Analysis Intelligence Index v5.0.1 with 65, and it is also the most expensive model on that board to run: $26.54 of provider spend to complete one index pass, against $5.48 for the model directly behind it. A review of it is therefore not a question of whether it is good — it is the best-scoring widely released model available, and the Claude Fable 5.2 API page lists $8.00 per million input and $40.00 per million output. The question is what you are paying for and whether your workload can use it.

What it is

Anthropic's successor to Fable 5.1, built for demanding reasoning and long-horizon agentic work. A 1M-token context window where the maximum is the default, up to 128K output tokens per response, text, image and document input with text out, native tool use and structured JSON. Thinking is always on and adaptive, with an effort dial from low to max, and the raw chain of thought is never returned — you get a summary or an empty thinking field.

That last detail is the one to design around. You cannot audit the reasoning; you can only audit the answer and the summary. For regulated workflows that matters more than the index score, and it is also why the effort level is worth setting deliberately rather than leaving at the default.

The score

65 on v5.0.1, rank 1 of 78. The nine evaluations are AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 Hard Subset and Terminal-Bench-Science 0.1 — an agentic and hard-science mix rather than a coding-leaderboard mix. The five-point gap over Gemini 4 Carbon is the whole of the lead at the top of the board.

The version guard is not a formality. Scores from v4.3 and v4.3.2 belong to a different evaluation set and cannot be tabulated against v5.0.1. If you see Fable 5.2 ranked next to a v4.3 score, the comparison is invalid no matter how respectable the source.

What the top of the board costs in time

267 seconds to first token, rank 76 of 77 models with latency data — second slowest on the board. 46.8 minutes of wall clock per task across the nine evaluations, 1.6 times what Carbon takes. Output speed once it starts writing is 97 tokens per second, which is 35 percent faster than Carbon's 72 and irrelevant to the total, because the clock is spent thinking.

Against live traffic through OrcaRouter the picture is different again: p50 8.09 seconds to first token, p95 10.00 seconds, 101 tokens per second, across 35.6 million tokens served in seven days. AA measures a harness at maximum effort; the gateway measures the median real request. Both numbers are honest and neither alone tells you what a user will experience.

What it costs in money

$26.54 to run the index — rank 77 of 77, the most expensive model on the board, and 4.8 times Gemini 4 Carbon's $5.48. At $8.00 in and $40.00 out the blended rate is roughly five times Carbon's. AA also reports 275 million output tokens for its run against a verbosity figure of 0.97 percent; the token count is the figure that matters, and it is higher than Carbon's 205 million despite the much lower percentage, so the two statistics cannot be read as a pair.

Caching changes the arithmetic for production but not for the benchmark. Cache reads bill at $0.200 per million and writes at $10.00, so a 40,000-token prefix costs $0.40 to write once and $0.008 per subsequent read, against $0.32 for one uncached pass. A prompt you send twenty times a minute is cheap; a one-shot document is not. The LLM API pricing comparison puts that cache arithmetic next to the rest of the catalogue, where the gap between this model and the cheap tier is roughly five to one on output.

Reliability

Across the same seven-day window the error rate was 10.8 percent of requests returning non-2xx across all upstream providers — roughly one in nine. That is far better than Gemini 4 Carbon's 39.7 percent over the same period and still not zero. Any production path needs a retry policy and, if the call is user-facing, a fallback model; the Gemini 4 Carbon API is a sensible cheap tier for the calls that fail twice and can tolerate a weaker answer.

Where it earns the money

Three patterns justify it. Long-horizon agents where a single wrong step invalidates twenty minutes of work, and the cost of a bad answer dwarfs $40 per million tokens. Deep code work on a repository the model can hold in a 1M window. And hard analytical tasks where the alternative was a three-model ensemble reconciled by hand. In each case you are buying the five index points and, more importantly, the reduction in variance.

Two patterns do not. High-volume classification or extraction at low effort is a place where a cheaper model with a structured output does the job, and interactive products where the user waits on the first token are a place where 267 seconds at max effort is disqualifying regardless of quality. Set effort per route and the model becomes usable in both; leave it global and it will be either wasteful or unusable.

For the direct comparisons, Claude Fable 5.2 vs Claude Opus 5.5 pairs it with the cheaper Anthropic option, and Claude Fable 5.2 vs Claude Fable 5.1 covers the generation it replaces — which is 20 percent more expensive on both input and output, making the upgrade a price cut as well as a score gain.

Verdict

First on the index, last on cost, second slowest to first token, and the most capable widely released model you can call today. It is worth the money for work where a wrong answer is expensive and the call can be asynchronous, and it is the wrong default for everything else. The operational discipline the model demands — effort levels per route, a retry budget, a fallback for the one-in-nine failures — is the price of admission, and it is not in the price table.