Claude Fable vs Opus, and when Fable is worth the cost

Should you use Claude Fable or Opus?

By Andrej Ruckij · November 10, 2026 · prices and plan rules checked on Anthropic’s own pages the same day

TL;DR: I think Fable 5.2 is worth it, but only for the right jobs. It’s Anthropic’s strongest widely available model (#1 on Artificial Analysis, score 65) and it costs twice what Opus 5.5 does per token: $8/$40 per million input/output tokens against $4/$20. Use it for work that runs for hours and has to end in something finished: a codebase audit, a migration plan, an analysis that has to end as a finished spreadsheet or deck. Run it at high effort, not max. At high it scores 62 on Artificial Analysis for $9.08 per task, which beats Opus at max (59, $10.04). For routine coding, Opus 5.5 is still cheaper per solved task, and Anthropic’s cost guide says so; for everyday prompts it’s simply the cheaper default. On a Pro plan, Fable is included until November 16 (up to 50% of your weekly limit), then it runs on usage credits.

What changed on November 9

Anthropic released two models: Claude Fable 5.2 and Claude Mythos 5.2 (Anthropic). They’re the same underlying model. Here’s the split:

Fable 5.2 has safeguards for cybersecurity and biology and is open to everyone on a paid plan or the API.
Mythos 5.2 has the same capabilities with fewer of those blocks, and you only get it after verification: the Cyber Verification Program for security teams, the Life Sciences Verification Program for researchers, or Claude Security on Enterprise. Same price as Fable (Anthropic pricing).

If you work in marketing, sales, ops or ordinary software, Mythos doesn’t matter to you. Fable is the one you can use. (The Register also reports Cisco and IBM are putting Mythos inside their security products, so security teams may get it that way (The Register).)

Against Fable 5.1, the new one is:

20% cheaper per token ($8/$40, down from $10/$50);
up to 30% faster at writing output;
much better at agentic coding on Anthropic’s benchmark (49.1% vs 32.4% on FrontierCode 2.0, both at xhigh effort);
blocked less often: about 55% fewer safeguard blocks overall.

Fable 5.1 is now a legacy model (models overview).

Fable 5.2 vs Opus 5.5, side by side

I took these from Anthropic’s model, API pricing and plan pages today.

	Claude Fable 5.2	Claude Opus 5.5
What Anthropic says it’s for	”demanding reasoning and long-horizon agentic work"	"long-running agentic coding and knowledge work”
API price (input / output, per 1M tokens)	$8 / $40	$4 / $20
Cache hit	$0.20	$0.20
Batch API	$4 / $20	$2 / $10
Fast mode	$16 / $80	$8 / $40
Default effort on the API	`high`	`medium`
Context window / max output	1M / 128K	1M / 128K
Pro plan ($20/month)	usage credits (free until Nov 16, see below)	included
Max, Ultra, Team premium seats	up to 50% of your weekly limit, then credits	included

Two things in that table matter more than the headline price.

Cache hits cost the same. Fable’s cache-read rate is 2.5% of its input price and Opus’s is 5%, so both land at $0.20 per million tokens (pricing). In a long agent loop where most of the input is cached, the gap is mostly on output.

The default effort is different. Fable defaults to high on the API, Opus to medium (effort docs). So if you switch the model name in your code and change nothing else, you’re also switching to a model that thinks harder by default. That’s a lot of the “why is my bill so high” surprise.

Why “20% cheaper” doesn’t mean cheaper

Price per token went down. Price per task can go way up, because Fable spends tokens.

Artificial Analysis runs every model through the same set of tests. At max effort, Fable 5.2 used 275 million output tokens to get through them, against a 72.8 million median for comparable models, and it cost $26.54 per task (Artificial Analysis). Opus 5.5 at max used 218 million and cost $10.04 per task (Artificial Analysis).

The leaderboard also has every effort level, and this is the table I’d actually look at before picking (leaderboard):

Model and effort	Score	Cost per task	Time per task
Fable 5.2, max	65	$26.54	46.4 min
Fable 5.2, xhigh	62	$14.83	26.4 min
Fable 5.2, high	62	$9.08	17.6 min
Fable 5.2, medium	56	$3.80	8.3 min
Opus 5.5, max	59	$10.04	25.1 min
Opus 5.5, xhigh	57	$6.05	16.2 min
Opus 5.5, high	52	$3.46	9.5 min
Opus 5.5, medium (default)	48	$2.45	7.1 min

What I take from it:

Fable at high beats Opus at max. Higher score, slightly less money, about 30% less time. This is where Fable earns its price.
Fable at max is a ceiling test. It costs almost three times Fable at high for 3 more points. Anthropic’s own docs say the same thing: “Reserve max for measuring the model’s ceiling … It runs slower and costs more than at xhigh, often for only a small gain” (effort docs).
Opus at its default is the cheap everyday option. $2.45 a task, a quarter of Fable at high.

For a feel of what long jobs cost: Decrypt reports that Matthew Berman’s three big Fable 5.2 builds came to estimates of about $11,700 to $14,710 each at pay-as-you-go prices (two of them ran 43 and 71 hours), and still came out with bugs (Decrypt). Those were stress tests, but they show what happens when nothing is capped. If you want the bigger picture of where AI money goes in a business (seats, tokens, people’s time), I wrote that up in how much AI costs a business.

The case against Fable (from Anthropic itself)

The strongest counter-evidence I found comes from the vendor’s cost guide, which it labels internal and directional (Anthropic: optimizing for cost and intelligence):

Routine coding: on a 463-problem SWE-bench Pro subset, Fable 5.2 at high solved 96.3% at $1.39 per solved task. Opus 5.5 at xhigh solved 96.5% at $0.75, “level with Fable 5.2 at high for a little over half the cost.” Opus at its default solved 93.1% for $0.22.
Deep research: on DeepResearch Bench II, Fable 5.2 cost 2.9 times as much as Opus 5.5 “for a score 1.9 points lower”, and 4.3 times as much as Sonnet 5.5 “for no measurable gain in score.”
As the boss of other agents: with Fable as the orchestrator instead of Opus on a defect-search job, it “cost 19% more and scored about the same.”

So Anthropic is telling you, in writing, that Fable isn’t automatically better value. Its advice: “Most workloads start with Claude Opus 5.5” (choosing a model).

What people who’ve tried it say

It’s been out a day, so these are early tests and anecdotes, not studies.

A tie on ordinary coding (small test). On Hacker News, one developer ran 30 coding tasks: Fable 5.2 and Opus 5.5 both solved 26 (Fable 5.1 solved 25). Several others said they can’t tell the two apart on everyday coding (Hacker News).
It eats plan limits (anecdote). One person on the $200 plan said an evening used 25% of the Fable limit, “3x faster consumption than Opus 5.5”. Another said a code review used up their 5-hour budget in half an hour, though it found 7 bugs (Hacker News).
Faster but pricier on one big job (anecdote). DataCamp relays a run where Fable finished in 1 hour 57 minutes for $81.10 and Opus took 3 hours 11 minutes for $45.97: about 40% faster, roughly 75% more expensive (DataCamp).
Better at running agents (anecdote). A couple of developers said Opus doesn’t manage agents as well, and that Fable gets further on ambitious ideas (Hacker News).
Finishing the job (vendor-reported). In launch quotes relayed by The Deep View, Uber’s Andrew Mager says it “works until the job is actually complete instead of stopping at a plausible checkpoint”, and Datadog’s Matt Lee calls it “almost like an incident commander that can send out faster models” (The Deep View). Customer quotes around a launch are picked by the vendor, so treat them as claims.

I found no Reddit threads or measured business tests on 5.2 yet. Give it a couple of weeks.

What Fable 5.2 is good for

From what Anthropic says and what people saw on day one, I’d use Fable when:

The job runs for hours and has to end in something finished. Anthropic’s list: “Agent sessions that run for hours, multistep deep research, analysis carried through to a finished document, spreadsheet, or deck” (choosing a model). Deep research is on that list, but in the deep-research test above Opus came out cheaper and slightly ahead, so I’d test that one before switching.
Opus has already failed at it. Anthropic’s step-by-step advice is to move to Fable when your tests on Opus at xhigh or max “still fall short”.
It’s coordinating other models, maybe. Fable can act as an advisor that cheaper models call on hard decisions, or as the orchestrator that hands simpler work down. This is the pattern we describe in glossary/advisor-strategy: a cheap model does the volume, an expensive one makes the hard calls. Every current executor model accepts Fable 5.2 as its advisor (advisor tool docs). The “maybe” is because Anthropic’s own orchestrator test above came out at 19% more cost for about the same score, and nobody has published a 5.2 advisor test yet. Measure it on your work.
Being wrong is expensive. A strategy pack for a client, a security review, a migration plan. Paying several times more per task is fine when a bad answer costs a week.

It’s a poor fit for chat, quick drafts, ad copy variations, email replies, summaries and routine bug fixes. Opus, Sonnet or Haiku do those for a fraction of the money.

What it costs on your plan

Plan rules are from Anthropic’s help centre (Fable on your plan, Fable 5.2 launch offers):

Free: no Fable.
Pro and Team standard seats: normally Fable runs only on usage credits. For launch week you get up to 50% of your weekly limit on Fable plus one Fable limit reset, until November 16, 11:59 pm PT. After that it’s back to credits.
Max, Ultra, Team premium seats: Fable is included up to 50% of your weekly limit, then usage credits or another model. For the launch you also get one to three Fable resets, depending on how long you’ve subscribed (under 6 months: 1, 6 to 12 months: 2, over 12 months: 3), valid until March 8, 2027.
API and usage-based Enterprise: standard API rates, the $8/$40 above.

One trap: using a reset doesn’t switch your default model. You still have to pick Fable in the model picker. And in Claude Code, fast mode always bills to usage credits, even if you have plan usage left (Claude Code fast mode).

Which to pick

Your situation	Pick
Everyday questions, drafts, summaries, ad copy	Opus 5.5 at default, or Sonnet 5.5
High-volume extraction, routing, classification	Haiku 5.5
Routine coding and bug fixes	Opus 5.5 (Anthropic’s cost guide says it’s cheaper per solved task)
A long agent job that must finish: audit, research pack, migration	Fable 5.2 at `high`
Opus at `xhigh` keeps getting it wrong	Fable 5.2 at `high`
A system where cheap models do the volume	Opus or Sonnet as executor; try Fable 5.2 as advisor, but measure it (no published 5.2 advisor test yet)
Security research or biology work that Fable blocks	Apply for Mythos 5.2 verification
You’re on Pro and it’s after November 16	Opus, unless the job is worth paying credits for

How to keep the bill under control

Set effort to high and leave it there. Not max. If you’re on the API, set it explicitly; in Claude Code, use /effort.
Don’t set max_tokens too low. It sounds like a cap, but Anthropic’s tests show the opposite: at 16,384 tokens, 85% of Fable’s attempts were cut off and each solved task cost $27.96; at 64,000 it solved 77.4% for $24.42 (cost guide). Anthropic recommends a large max_tokens at high and above.
Put a dollar cap on long runs. Claude Code has --max-budget-usd and /usage; usage credits can have a monthly spend limit, and Team and Enterprise admins can set limits per member (Claude Code costs).
Watch the subagents. Anthropic warns that “subagents can account for a large part of what a session costs”, and at xhigh and max Fable starts subagents to review its own work. One more reason to stay at high.
Don’t switch models mid-conversation casually. Moving from Fable 5.2 to another model loses its reasoning for those turns (what’s new in Fable 5.2).
Price it on your own work. Anthropic’s line is the best advice here: “Price every candidate in cost per completed task on your own traffic.”

What I’d do this week

If you’re on a paid plan, use the free week. Before November 16:

Pick one job you’d never hand to Opus because it’s too long or too messy: a full audit of a codebase, a migration plan, an analysis that has to end as a finished spreadsheet or deck.
Run it on Fable 5.2 at high, and note how much of your weekly limit it used in Settings > Usage.
Run the same job on Opus 5.5 at xhigh. Compare the result, not the benchmark.
Write down one rule for your team from what you saw, like “Opus by default, Fable only for jobs that run more than an hour”.

If you run Claude through the API, the equivalent is 20 real tasks on both models, priced per finished task. If you haven’t connected Claude to your own data yet, connecting it to your Meta and Google Ads accounts is a good job to test both on, and we compared the plans in ChatGPT Business vs Claude for business.

Gaps in what I could check

All the scores beyond Artificial Analysis are from Anthropic. Its cost guide is frank about where Fable loses, but it’s still the vendor’s data.
I found no published test with Fable 5.2 as an advisor. The only published advisor result in Anthropic’s docs uses Fable 5.1.
The early user reports are from one day of Hacker News threads, not controlled tests.

Sources

Cover image: Anthropic, from the Fable 5.2 launch post.

https://t.me/aidaily_business