Claude Fable 5.2 vs Opus 5.5: Same Scores at Almost 4x the Cost?
Key Takeaways
- Anthropic released Claude Fable 5.2 on 9 November 2026 as its most capable model, at 8 dollars per million input tokens and 40 dollars per million output tokens, exactly double the 4 and 20 dollars charged for Claude Opus 5.5.
- On Anthropic's own benchmarks, Fable 5.2 does lead: 49.1 percent against 45.5 percent on FrontierCode 2.0, 69.9 percent against 62.3 percent on FrontierSWE v2, and 41.3 percent against 21.3 percent on the Remote Labor Index.
- But a hands-on comparison by Jessica Wachtel for The New Stack, published 15 November, found both models scored a perfect 150 out of 150 across three coding and agent tasks at three difficulty levels each, while Fable cost an estimated 8.98 dollars against 2.29 dollars for Opus, nearly four times as much, and took 1,579 seconds against 701 seconds.
- Anthropic itself tells most developers to start with Opus 5.5, calling it the recommended default for everyday knowledge and coding work, and reserves Fable 5.2 for the hardest, longest-running jobs.
- The gap between the two stories is not a contradiction. Anthropic's benchmarks measure maximum capability at high effort settings, while the hands-on test measured what each model does at its default settings on ordinary engineering work.
- The practical rule that emerges from both sets of evidence is simple: start with Opus 5.5, and move up to Fable 5.2 only when Opus stops producing the results you need.
The Two Models at a Glance
Anthropic's Claude line-up now has a clear hierarchy, and understanding it is the first step to choosing between these two models. Claude Opus 5.5 launched on 22 September 2026. For about seven weeks it was Anthropic's top public model, and it remains the model Anthropic recommends as the default for most work. Claude Fable 5.2 launched on 9 November 2026, replacing Fable 5.1 as the company's most capable model, and it arrived alongside Claude Mythos 5.2, which Anthropic describes as the same model with fewer safety blocks, available only to vetted security and science teams.
The price difference is the simplest fact in this comparison, and it is worth stating without qualification. Fable 5.2 costs 8 dollars per million uncached input tokens and 40 dollars per million output tokens. Opus 5.5 costs 4 dollars and 20 dollars for the same. Fable is exactly double the price on both input and output, and it is 20 percent cheaper than its predecessor Fable 5.1, which cost 10 and 50 dollars. Both models handle one million tokens of context and up to 128,000 output tokens, according to comparison coverage verified against Anthropic's models overview, effort documentation and pricing page on 13 November 2026.
There is one more price detail that matters for subscribers rather than API users. Fable 5.2 is not inside Claude Pro's normal limits, while Opus 5.5 is. Where Fable is included, on Max, Ultra and Team premium seats, it is capped at half of weekly usage. That is a usage limit, not a price, but it has the same practical effect: heavy Fable use is rationed in a way that heavy Opus use is not. A promotional offer reported at launch allowed spending up to half of the weekly limit on Fable at no extra cost until 16 November 2026 at 11:59 pm Pacific, which is today, so readers should not assume that offer is still available.
The naming deserves a brief note, because it confuses people. Fable is the top tier of the Claude family, above Opus, and the version numbers do not line up across tiers: Fable 5.2 is newer than Opus 5.5, even though 5.2 is a lower number than 5.5. The generations share the number 5, but each tier has its own version sequence. If you remember only one thing about the names, remember this: Fable is the flagship, Opus is the workhorse, and the question this article answers is whether the flagship earns its premium on the work you actually do.
What Anthropic's Own Benchmarks Say
Anthropic's position is unambiguous. The company says Fable 5.2 is a step above Opus 5.5 in almost all the capabilities it evaluated, and its launch table supports that statement on most rows. The numbers below are Anthropic's own, reported by the company and not independently checked, and the effort settings matter, so we state them where they are known.
On FrontierCode 2.0, an agentic coding benchmark, Fable 5.2 scored 49.1 percent at extra-high effort, against 45.5 percent for Opus 5.5 at maximum effort. On FrontierSWE v2, another agentic coding test, Fable scored 69.9 percent against 62.3 percent for Opus. On the Remote Labor Index, which uses real freelance jobs to measure knowledge work, Fable scored 41.3 percent against 21.3 percent for Opus, nearly double. On AutomationBench, which measures business workflows, Fable scored 51.4 percent against 42.5 percent. On AA-Briefcase v1.1, a knowledge-work test, Fable scored 66.1 percent on the rubric against 61.2 percent for Opus, although Opus still edges Fable on the same test's Elo rating, 1800 against 1785. That Elo result is worth remembering, because it is one of the few places where Anthropic's own table shows Opus ahead, and Anthropic printed it rather than hiding it.
Two further results complicate the simple story that Fable is better everywhere. On controlling a robot, Fable scored 65 percent against 39 percent for Opus, a large gap. On expert-level examinations, Fable scored 74 percent against 68 percent. On web search for large research jobs, Fable scored 82 percent against 72 percent. On designing hardware parts, Fable scored 83 against 65. The pattern is consistent: the gap is smallest on ordinary coding, where Fable leads by 3.6 points on FrontierCode, and largest on long, open-ended work, where the Remote Labor Index gap is 20 points.
But the most important row in Anthropic's table may be the one that undermines the case for always using Fable. Anthropic reports that Fable 5.2 at high effort scored 45.3 percent on FrontierCode 2.0, against Opus 5.5's best of 45.5 percent, for roughly half the cost per task. In other words, turn Fable down from its maximum effort and it matches Opus's best score while costing less per task. And at its lowest setting, according to analysis of the launch, Fable matched the best score Opus could reach for less than a tenth of the cost. The effort dial is the hidden variable in this whole comparison. Benchmark scores are always reported at a chosen effort level, and the cost of a score depends on where that dial is set. A model that wins at maximum effort may not be the cheapest way to reach a given score, and that is exactly what the hands-on testing set out to check.
There is also independent evidence that the gap between the two models is small on some kinds of work. Crescendo Lab ran both models through the same website agent on the same 150 customer conversations, judged by two independent AI judges. Fable 5.2 fully passed 131 of 150 cases, or 87 percent, while Opus 5.5 passed 133, or 89 percent. Case by case the two were a dead heat, with a statistical test returning a p-value of 0.80, which means the difference is indistinguishable from chance. The same benchmark produced a striking result at the other end of the price list: Claude Haiku 5.5 also passed 131 of 150 cases, the same as Fable 5.2, at 0.58 dollars per 1,000 conversations, against 48.67 dollars for Fable and 21.65 dollars for Opus. That finding is about a customer-service agent, not coding, so it should not be generalised beyond its setting, but it reinforces the central theme of this comparison: on many real tasks, the cheaper model is already good enough, and the expensive one buys little or nothing extra.
The Hands-On Test: Same Scores, Nearly Four Times the Cost
The comparison that gave this article its headline was run by Jessica Wachtel, a developer marketing writer at InfluxData, and published by The New Stack on 15 November 2026. Wachtel's question was a practical one. Anthropic itself steers most people toward Opus 5.5 rather than its best model, which is an unusual thing for a company to do, and she wanted to find out whether the cheaper model really is enough.
Her method was deliberately ordinary. She gave both models three coding and agent jobs: building a cron scheduler, fixing a bug in an invoicing repository, and investigating a production incident. She measured how well each model did, how many tokens it used, what it cost, and how long it took. She ran each model once per test through Anthropic's API at default settings, with no turn cap, and she was explicit about the limits of what she did. Her tests are much smaller than Anthropic's benchmarks and cannot prove or disprove those scores. They measure what a developer gets without tuning anything, which is a different question from what a model can do at its best.
The results were unambiguous on quality and striking on cost. Both models scored perfectly, 50 out of 50, on the first version of every test. Wachtel then made each test harder and ran it twice more. Across all three versions of all three tests, both models scored 150 out of 150. There was no task, at no difficulty level, where Fable outscored Opus.
The difference was in cost and time. Across all nine runs, Fable 5.2 consumed an estimated 8.98 dollars in API costs and took 1,579 seconds, while Opus 5.5 consumed an estimated 2.29 dollars and took 701 seconds. Fable cost nearly four times as much and took more than twice as long. The gap held on every individual test. On the cron scheduler, both models scored 150 points across three versions, but Opus finished in 358 seconds for 0.99 dollars while Fable took 661 seconds and cost 3.29 dollars. On the invoicing repository fix, both models achieved perfect scores, but Opus finished in under a minute with minimal reasoning tokens, while Fable took up to four minutes and consumed far more tokens. Wachtel's conclusion was direct: the tests give her a reason to start with Opus 5.5 for work like this, because it matched Fable 5.2's score and finished faster on every version, at about a quarter of the estimated cost overall. She said she would start with Opus 5.5 and move up to Fable 5.2 only when Opus stops producing ideal results.
Why did Fable cost so much more for the same score? The answer lies in the default settings. Anthropic's API defaults are not equal across the two models: Fable 5.2 runs at high effort by default, while Opus 5.5 runs at medium. Higher effort means more reasoning tokens, longer chains of thought, and more time spent before answering. On tasks where the extra reasoning does not change the answer, it is pure overhead. Fable was, in effect, thinking harder about problems that Opus could already solve, and charging for the privilege. That is not a flaw in Fable. It is a consequence of running a more powerful model at a higher default effort on work that does not need it. The lesson is about defaults, not about capability.
When Fable 5.2 Still Wins
None of the above means Fable 5.2 is a waste of money. It means its premium buys something specific, and buyers should know what that something is before paying for it.
The clearest case for Fable is long, open-ended work. The Remote Labor Index gap, 41.3 percent against 21.3 percent, is the largest in Anthropic's table, and it measures real freelance jobs rather than short benchmark tasks. On work that runs for hours, involves ambiguous requirements, and requires sustained planning across many steps, Fable's advantage is substantial. If your work looks like that, Opus may not be enough, and the hands-on test, which used tasks that could be completed in minutes, would not have detected the difference. A cron scheduler and an invoicing bug fix are bounded problems with clear success criteria. A multi-week research project or a large refactoring across a whole codebase is not, and that is where Fable earns its keep.
The second case is work at the edge of what models can do at all. Fable's lead on robot control, 65 percent against 39 percent, and on hardware design, 83 against 65, suggests that on unusual or highly technical tasks the flagship pulls away. If you are working in a domain where Opus regularly fails, Fable is the obvious next step, and the cost of failure may dwarf the cost of tokens. A model that succeeds where a cheaper one fails is cheap at any price, provided the task matters enough.
The third case is cost per task at matched quality, which cuts in Fable's favour in a way that the headline prices hide. Anthropic's finding that Fable at high effort matches Opus's best FrontierCode score for roughly half the cost per task shows that the effort dial can make Fable the cheaper option for a given quality bar, if you tune it. The list prices say Fable costs double. The cost per completed task says something more nuanced, and it depends on how you configure the model. Teams that tune effort settings carefully may find Fable cheaper than Opus for the same result, which is the opposite of what the list prices suggest.
Finally, there is the question of headroom. Opus 5.5 at maximum effort is already at its ceiling. Fable 5.2 at extra-high effort has more room, and Anthropic's table shows it using that room on the hardest tests. If your workload is getting harder over time, as agentic workloads tend to, buying headroom now may be cheaper than discovering the ceiling later.
The Cost Maths, Worked Honestly
It helps to translate the list prices into concrete numbers, because percentages hide the scale of real spending. The arithmetic below uses Anthropic's published list prices only. It is not a measurement, and your actual costs will depend on your mix of input and output tokens, your use of caching, and your effort settings.
Take a team that runs one million input tokens and one million output tokens through the API. On Opus 5.5, that costs 4 dollars for input and 20 dollars for output, a total of 24 dollars. On Fable 5.2, the same volume costs 8 dollars for input and 40 dollars for output, a total of 48 dollars. That is the doubling the list prices imply, and it is the right mental model for budgeting at equal volumes.
But equal volumes are not what happens in practice, and Wachtel's test shows why. Fable used more tokens per task than Opus, because its higher default effort generates more reasoning tokens. That is how a model that costs double per token ends up costing nearly four times as much per completed task. The multiplier is the product of two things: the price per token and the tokens per task. If you run Fable at its defaults on ordinary work, you pay both multipliers. If you turn the effort down, you may pay neither.
Caching changes the picture further. Both models charge less for cached input tokens than for uncached ones, and the figures quoted in this article are for uncached input. Teams with repetitive prompts, large shared contexts or long-running agents will see different effective prices, and the gap between the models may narrow or widen depending on cache hit rates. We have not calculated cached pricing here because the discount structures differ and the result would depend on assumptions we cannot verify for your workload. The honest statement is that list prices are a starting point, and your own usage pattern decides the real gap.
For subscribers, the maths is different again. If Fable is capped at half of weekly usage on your plan, then the question is not what Fable costs per token but whether your heaviest work fits inside the cap. A team that hits the cap mid-week and falls back to Opus anyway is paying for a flagship it cannot fully use. Check your plan's limits before choosing a model on price alone.
How to Decide: A Practical Framework
The evidence in this article points to a decision rule that is simple to state and easy to apply. Start with Opus 5.5. That is Anthropic's own advice, and both the company's benchmarks and the independent testing support it for most work. Opus is the recommended default for everyday knowledge and coding work, it costs half as much per token, it is inside Claude Pro's normal limits, and on ordinary engineering tasks it matched the flagship's scores in hands-on testing while finishing faster.
Move up to Fable 5.2 when one of three conditions holds. First, when Opus fails or produces results that are not good enough, and the task matters enough to justify the premium. Wachtel's rule, move up only when Opus stops producing ideal results, is the right one. Second, when the work is long, open-ended or at the edge of current capability, where Anthropic's benchmarks show Fable's largest leads, particularly the Remote Labor Index and the robot-control and hardware-design results. Third, when you have tuned the effort settings and found that Fable reaches your quality bar at a lower cost per task than Opus at its maximum, which Anthropic's own FrontierCode analysis suggests is possible.
There are also two things not to do. Do not choose Fable because it is the newest or the most capable model in the abstract. Capability you do not use is money spent for nothing, and the hands-on test is a reminder that maximum capability and practical value are different things. And do not choose Opus for everything without testing Fable on your hardest work. The benchmark gaps on long tasks are large enough that a team doing genuinely hard work may be leaving quality on the table by never trying the flagship.
One more consideration applies to teams rather than individuals. If you are standardising on a single model for a whole team, Opus is the safer default, because its costs are predictable, its limits are generous, and its performance on ordinary work is proven. Reserve Fable for the specific people or projects whose work justifies it, and review that reservation regularly. Model pricing and capability change every few months, and a decision that is right in November may be wrong by February.
For readers who want the background on Fable 5.2 itself, our earlier explainer covers what the model is, what changed from Fable 5.1, and what Mythos 5.2 is for. This article deliberately does not repeat that material. It is a comparison and a cost analysis, and it assumes you already know, or can read elsewhere, what Fable 5.2 is. If you are weighing the wider question of what AI models actually cost in practice, our analysis of cheaper AI models from OpenAI and Anthropic puts the Fable and Opus prices in the context of the broader market, where list prices have been falling even as flagship capability rises.
What We Could Not Verify
- The original New Stack article itself. The figures quoted here come from secondary coverage of Wachtel's tests, including detailed summaries that reproduce her costs, timings and scores. We were able to verify the numbers across multiple independent summaries that agree with each other, but we did not read the original article text at the time of writing, so quotations from it should be treated as reported rather than directly verified.
- Whether the per-task cost breakdowns beyond the cron scheduler, the totals of 8.98 dollars and 2.29 dollars, and the timings of 1,579 and 701 seconds, are exact to the cent. They are the figures reported in the coverage of the tests, and they are estimates based on API usage, not metered bills.
- The current status of the Fable promotional offer, reported as running until 16 November 2026 at 11:59 pm Pacific. That date is today, and the offer may already have ended by the time you read this.
- The precise cached-input pricing for both models and the effective cost gap for a cached workload. Cached prices differ from the uncached list prices quoted here, and the gap depends on cache hit rates we cannot assume.
- Whether the Crescendo Lab agent benchmark, which found Fable and Opus statistically tied on 150 customer conversations, generalises to coding or other kinds of work. It measured a customer-service agent, and its results should not be read as a coding comparison.
- The exact effort-setting behaviour of the API defaults beyond the reported fact that Fable defaults to high effort and Opus to medium. How effort maps to token consumption on any given task varies, and we have not measured it ourselves.
Sources
- The New Stack, via Zot News, Claude Opus 5.5 matches Fable 5.2's scores at a quarter of the cost in hands-on tests, 15 November 2026, https://zot.news/article/claude-opus-5-5-matches-fable-5-2s-scores-quarter-cost-hands-9fs1t
- GazeWire, Claude Fable 5.2 vs Opus 5.5 Cost and Performance Analysis, 15 November 2026, https://gazewire.com/technology-innovation/ai-and-automation/claude-fable-5-2-and-opus-5-5-model-comparison
- The AI Career Lab, Claude Opus 5.5 vs Fable 5.2: Which Should You Use?, November 2026, https://theaicareerlab.com/blog/claude-opus-5-5-vs-fable-5-1-for-professionals
- Vanda Team, Claude Fable 5.2 vs Opus 5.5, GPT-6 Astra and Mythos 5.2, November 2026, https://vandatateam.com/blog/claude-fable-5-2-vs-opus-5-5-mythos-5-2
- Crescendo Lab, Claude Fable 5.2 joins the top of our agent benchmark, November 2026, https://caacgo.cresclab.com/benchmark/fable-5-2-november-2026
- Claude Fable 5.2 and Mythos 5.2 Explained
- OpenAI and Anthropic cheaper AI models: real 2026 costs