Anthropic says start on Opus, not Fable. The reason is cost per finished task

On 9 November 2026, Addy Osmani published a practical guide on Anthropic's Claude developer blog: which Claude model to build with, what they cost, and how to use them together. The current generally available lineup is Claude Haiku 5.5, Claude Sonnet 5.5, Claude Opus 5.5, and Claude Fable 5.2. The short answer in the post is to start most work on Opus 5.5 at its default effort level, medium. That is also the default model in Claude Code. Fable 5.2, the larger generally available model, costs twice as much per input and output token. The reason for the default is not that Fable is weak. It is that the bill that matters is the cost of a finished task, and a higher price per token does not always mean a higher cost, or a better result, on that task.

What the guide actually says

The post is a decision guide, not a model launch. It gives a starting model and a starting effort level for common jobs. Everyday engineering and debugging start on Opus 5.5 at medium, then high. Classification, routing, tagging, and high-volume extraction start on Haiku 5.5. Well-scoped coding, chat, repeated agent tasks, and everyday documents start on Sonnet 5.5, which the post says handles scoped work at half the Opus token price. Major refactors, architecture planning, leading many subagents, and deep scientific or quantitative work start on Fable 5.2, at medium and then high. If you are not sure, the advice is still Opus 5.5 at medium.

Effort is treated as a finer dial than a model swap. The post warns that past high, one more effort level can cost as much as moving up a model, so the two should be compared on your own work. Higher effort helps when a model needs to think longer, call more tools, or check itself more, and it also adds waiting time. Raising effort will not give a model knowledge it does not have.

The price table is Claude API list pricing. All four models have a 1 million token context window and a 128K maximum output. Cache reads are $0.01 or $0.05 per million tokens for Haiku, depending on prompt length, $0.10 for Sonnet, and $0.20 for both Opus and Fable. On a long, cache-heavy run, the post says, the gap between Fable and Opus therefore comes down to uncached input and output, where Fable still costs twice as much per token. Haiku is the only model whose token price jumps when a prompt goes over 100K tokens. Even then, the higher Haiku tier is a quarter of Sonnet's per-token price, but the post says to compare cost per finished task when every call needs a long prompt.

When a bigger model is the wrong spend

The examples are specific, and they are Anthropic's own runs. On Terminal-Bench 3.0, the previous Fable, 5.1, at max scored 58.0%, while Opus 5.5 at high scored 58.9% with half the tokens. On FrontierCode 2.0, Fable 5.2 at high scored about level with Opus 5.5 at max, 45.3% versus 45.5%, for roughly half the cost per task. The post's own lesson is that before you raise Opus 5.5 past high on a hard task, you should test Fable 5.2, and that everyday work should still start on Opus at medium.

Effort cuts the same way. On SWE-bench Pro, Opus 5.5 at medium scored about 2.5 points lower than high, for about 70% of the cost. At low it scored about 8 points lower, for about a third of the cost. At xhigh it scored about 1.4 points higher, for 2.5 times the cost of high. A cheaper pattern in the same post is to run everything at low and re-run only the failures at high. On the same SWE-bench Pro subset, that passed about 97% of tasks for about $0.17 each, against 95.3% for $0.29 when every task ran at high.

The post is also cautious about two fashionable setups. An advisor tool, still in beta, pairs a cheaper executor model with a stronger advisor. Of the four pairings in Anthropic's own table, none clearly beat the executor alone. One pairing, Opus 5.5 at high with a Fable 5.1 advisor, scored 90.1% at $2.92 per attempt, 1.7 points over Opus 5.5 alone at high, for about 2.1 times the money, a gain the post itself puts at the edge of run-to-run noise. A different pattern did save money: on an internal defect-search benchmark over about 22 million tokens of source code, an Opus 5.5 lead over Sonnet 5.5 workers, up to 25 at a time, cost about a third less than Opus 5.5 alone at its default effort, and scored about the same.

The fair case, and what is missing

The fair reading is that a lab telling customers not to default to its most expensive generally available model is useful, and that the supporting numbers are concrete enough to test. The post itself says published benches only get you a starting model, and that you should measure candidates on your own tasks.

The limits are the usual ones for a first-party guide. The scores are Anthropic's runs, often marked with "about." A move down a model tier drops the earlier model's hidden reasoning, because a model does not read the thinking blocks of a model above it. A first turn on a new model also pays the cache-write price on the whole conversation, unless you compact or summarize first. And the haiku alias is not the same everywhere: on the Anthropic API it now points to Haiku 5.5, which needs Claude Code v2.1.293 or later, but on Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS, it still points to Haiku 4.5.

Who should care, and what to watch

Teams paying API bills, and anyone who has been routing every hard ticket to the largest model by default, should care. So should people comparing Fable 5.2's headline capability with what it costs to leave it on all day.

In the next six months, the check is whether builders actually start on Opus at medium, and whether cost-per-task figures like the FrontierCode and SWE-bench Pro results still hold when other groups rerun them. The other number already in the post is cache cost on long runs: about $0.03 for a 150K-token cache read on Opus, against about $0.004 at 20K.

Close

The guide's load-bearing claim is simple enough to be wrong: most work should start on Opus 5.5 at medium, and a move to Fable, or a jump in effort, should be priced as cost per finished task. The 45.3% versus 45.5% FrontierCode result, and the $0.17 versus $0.29 re-run result, are the figures against which that advice can later be judged.