· ai · agentic · tooling · —
Opus 5.5 vs Fable 5.2: The Cheaper Model Is Not Always the Cheaper Job
Opus 5.5 costs half as much per token as Fable 5.2, and Anthropic says to start with Opus. But on hard tasks, Fable at lower effort matches Opus for less money. Here's how I pick between them.
Anthropic now sells two top models. Claude Opus 5.5 came out on 22 September. Claude Fable 5.2 came out on 9 November. Opus costs half as much per token. Fable scores higher on almost every test Anthropic ran.
So the easy answer is to use Opus unless money doesn't matter. I think that's close, but it's wrong in one way. The price per token is not the price per job. Here's the comparison I wanted before choosing, with Anthropic's own numbers and what I've seen myself.
Side by side, the sticker says Opus
| What | Opus 5.5 | Fable 5.2 |
| Input, per million tokens | $4 | $8 |
| Output, per million tokens | $20 | $40 |
| Cache reads, per million tokens | $0.20 | $0.20 |
| Fast mode, per million tokens | $8 input, $40 output | $16 input, $80 output |
| Context window | 1M tokens | 1M tokens |
| Max output | 128K tokens | 128K tokens |
| Default effort on the API | medium | high |
| Data retention | Available with zero data retention | 30 days by default |
Fable costs twice as much, in standard mode and in fast mode. The context window and the max output are the same, so size is not a reason to pick either one.
Two rows are easy to miss. Cache reads cost the same on both. Anthropic's Opus 5.5 post says cache reads "make up the majority of agentic and coding work costs". So for long agent sessions, the gap is likely smaller than two to one. That's my reading, not Anthropic's claim.
The other row is the default effort. If you call the API and don't set it, Fable runs at high and Opus runs at medium. That makes the gap bigger again, unless you set effort yourself.
The scores say Fable
These are Anthropic's numbers, from the Fable 5.2 launch post. Both models ran with safeguards on, at their highest effort unless the row says something else.
| Test | Fable 5.2 | Opus 5.5 |
| FrontierCode 2.0, agentic coding | 49.1% (Xhigh effort) | 45.5% |
| FrontierSWE v2, agentic coding | 69.9% | 62.3% |
| AutomationBench, business workflows | 51.4% | 42.5% |
| Terminal-Bench-Science 0.1, science research | 74.6% | 58.7% |
| WANDR (relaxed score), agentic search | 82.1% | 72.3% |
| OSWorld 2.1 (offline set, partial), computer use | 89.3% | 87.2% |
| Remote Labor Index, knowledge work | 41.3% (High effort) | 21.3% |
| Humanity's Last Exam (with tools) | 74.2% | 67.7% |
| AA-Briefcase v1.1, knowledge work | 66.1% rubric, 1785 Elo | 61.2% rubric, 1800 Elo |
Fable leads in every row but the last one, and that one is a split. On AA-Briefcase, Fable has the higher rubric score and Opus has the higher Elo.
Look at the size of the gaps, because they aren't equal. On computer use, Fable is two points ahead. On the Remote Labor Index, it almost doubles Opus. The gap grows as the work gets longer and less defined.
The clearest example is the Fermat run. Anthropic gave both models the same setup of about 1,000 agents and asked them to formalize the proof of Fermat's Last Theorem. Fable finished in 30 hours. Opus completed 26% in 18 hours "and barely made any more progress after 30". It was one run per model, so I wouldn't lean on the exact numbers. But the shape matches the table.
All of this is a vendor testing its own products. I treat it as a claim until other people repeat it.
The price per job is a different number
Both models have an effort setting, from low to max. Higher effort means more thinking and more tokens. Anthropic's launch post compares the two models at different effort levels, and this is where the sticker price stops being the whole story.
- On FrontierCode 2.0, Fable at High effort "scores about the same as Opus 5.5's best score (45.3% vs. 45.5%) for roughly half the cost per task".
- On WANDR, Fable at Medium effort "matches Opus 5.5's top score for about 40% of the cost per task".
- On Humanity's Last Exam, Fable at Low effort "matches Opus 5.5's top score for less than one-tenth of the cost per task".
So on these tests, Fable turned down matches the best score Opus reached, and it costs less to get there.
Be careful with what is being compared, though. It's Fable at a lower setting against Opus at its best score, which on FrontierCode means Max effort. Most of my Opus work doesn't run at max. For daily work, this proves less than it seems. For hard work, where you'd push Opus to max anyway, it proves more. On those jobs, the model with the higher price per token can be the cheaper way to finish.
The price per token is not the price per job.
Where Opus 5.5 still holds
The price. Half the cost per token is a real advantage on every task where both models get it right. For most daily coding, that's most tasks.
Anthropic's own advice. The Fable page says it plainly: "For most everyday coding and knowledge work, we recommend starting with Claude Opus 5.5." The launch post calls Opus 5.5 "our strongest daily driver". When the company that sells the expensive model tells you to start with the cheaper one, I'd listen.
Fewer conditions on your data. The Opus 5.5 post says "Opus 5.5 is available with zero data retention." Fable needs 30-day retention by default, and only eligible customers can keep zero data retention.
One thing that is not a difference: blocked security work. Both models have safeguards, and in the Claude apps both pass most security tasks to Opus 4.8. The detail on blocks, fallbacks and retention is in my post on the Fable 5.2 release.
What I run, and when I switch
I mostly run Opus 5.5. I use it for client frontend work, my own projects, and the skills and agent tooling I build. For most of that, Opus is enough.
Fable 5.2 has one habit I like a lot. It keeps going until the job is done. It also has one I don't. It eats my usage limits fast.
That second point is the part the cost charts don't show. The charts are in API dollars. On a subscription, the cost is your weekly limit. This week's launch promotion includes Fable "for up to 50% of weekly usage limits" on Pro plans and standard Team seats, so on those plans it is rationed. A model can be cheaper per task on paper and still use up your week.
So this is my rule. Opus 5.5 is the default. I move up to Fable 5.2 only for hard, open-ended tasks, where a model that stops early costs me more than the extra usage.
Quick answers
Is Fable 5.2 worth double the price?
For long, hard tasks, Anthropic's numbers say yes, and at lower effort it can cost less per job than Opus at max. For daily coding, I don't think so, and Anthropic doesn't say so either.
Can I use both together?
Yes. Anthropic says Fable 5.2 can work as an Advisor, where Haiku, Sonnet or Opus models ask it for help on harder problems. It can also work as an Orchestrator, where it hands simpler tasks to smaller models.
Do they have the same context window?
Yes. Both have a 1M token context window and a 128K token max output.
Which one should I start with?
Opus 5.5. Move up when the work is hard and open-ended, and move back down when it isn't.
The numbers come from Anthropic's Fable 5.2 launch post, its Fable page, the Opus 5.5 launch post, and the models overview in the docs.