Claude Fable 5.2 vs Opus 5.5: Which Should You Use?

This comparison combines Anthropic's launch figures from 09/11/2026, independent measurements from Artificial Analysis, and the first head-to-head tests from creators including Bart Slodyczka, AI Revolution, AI Search and Matt Wolfe. Every number below is sourced. Where a figure comes from Anthropic rather than an independent party, we say so. For the full launch story, including Mythos 5.2 and the system card, read our Claude Fable 5.2 and Mythos 5.2 review. For the older model, see our Claude Opus 5.5 review.

• Capability: Fable 5.2 leads Opus 5.5 on 12 of the 13 benchmark lines where Anthropic publishes both scores. The gaps range from small (OSWorld 2.1: 89.3% vs 87.2%) to very large (RoboDojo Sim: 64.9% vs 39.4%; Remote Labor Index: 41.3% vs 21.3%).
• The one Opus win: on AA-Briefcase v1.1, a knowledge-work benchmark, Opus 5.5 has the higher Elo (1,800 vs 1,785), even though Fable 5.2 has the higher rubric score (66.1% vs 61.2%).
• Independent ranking: Artificial Analysis places Fable 5.2 (Max) first of 74 models with an Intelligence Index score of 65. Opus 5.5 (Max) sits fifth with 59.
• Token price: Fable 5.2 is exactly twice Opus 5.5: $8 vs $4 per million input tokens, $40 vs $20 per million output tokens.
• Real cost per task: at Max effort, Artificial Analysis measured $26.54 per Intelligence Index task for Fable 5.2 against $10.04 for Opus 5.5. Fable also used more output tokens (275 million vs 218 million across the index).
• The twist: at lower effort settings, Fable 5.2 often matches Opus 5.5's best scores for a fraction of the cost. Anthropic reports parity on FrontierCode 2.0 at High effort for about half the cost per task, and parity on Humanity's Last Exam at Low effort for under a tenth.
• Speed: Opus 5.5 is slightly faster to generate (115 vs 101 output tokens per second on Artificial Analysis), but Fable 5.2 reached its first answer token sooner in the same tests (234 vs 356 seconds at Max effort).

Anthropic now runs two premium lines. Opus is the long-standing flagship tier, sitting above Sonnet and Haiku. Opus 5.5 launched on 22/09/2026 at $4 input and $20 output per million tokens, 20% below Opus 5, and became the reference model for agentic coding through the autumn. Our Opus 5.5 vs GPT-6 Astra comparison covers how it stacked up against OpenAI at launch.

Fable is the newer, more capable tier. It is the broadly available sibling of Mythos, Anthropic's restricted frontier line for vetted cyber-defence and life-science organisations. Fable 5.2 and Mythos 5.2 are the same underlying model with different safeguards. The background is in our Fable 5 vs Mythos 5 explainer and the Claude Fable 5.1 review.

Anthropic's own framing is that Fable 5.2 is "a step above Opus 5.5 in almost all the capabilities we evaluated". That is a fair summary of the table, with one important addition: Opus 5.5 is still on sale, still supported and still half the price. Anthropic is not positioning Fable as a replacement for Opus. It is positioning it as the model you reach for when Opus runs out of road.

The table below takes every row from Anthropic's Fable 5.2 announcement where both models have a published score. Anthropic notes that scores for Fable 5.2, Fable 5.1 and Opus 5.5 are shown with safeguards on and at each model's highest available effort unless stated otherwise. These are vendor-reported numbers, so read them as Anthropic's best case rather than a neutral audit.

Three patterns stand out. First, the coding gap is real but modest. On FrontierCode 2.0, which measures whether a model's code changes would be merged by open-source maintainers, Fable leads by 3.6 points. On FrontierSWE v2 the gap is 7.6 points. Those are meaningful, but nobody who is happy with Opus 5.5 in Claude Code today will find it suddenly useless.

Second, the biggest gaps are on long-horizon and physical tasks. RoboDojo Sim (+25.5 points), the Remote Labor Index (+20.0 points), BenchCAD (+18.2) and Terminal-Bench-Science (+15.9) all involve sustained, multi-step work where small errors compound. This is where Fable's extra capability turns into qualitatively different outcomes rather than marginally better ones.

Third, Opus 5.5 is not beaten everywhere. The AA-Briefcase Elo result, where Opus 5.5 scores 1,800 against Fable's 1,785, suggests that in head-to-head preference judgements on business documents, Opus output is at least as well liked. That matches what several creators reported anecdotally: for writing-heavy knowledge work, the difference is hard to see.

Artificial Analysis runs its own Intelligence Index across a fixed blend of reasoning, knowledge, coding and agentic evaluations, using each model at a stated effort level. Because it uses the same harness for every model, it is the cleanest independent comparison available two days after launch.

A six-point lead on the Intelligence Index is substantial at the top of the leaderboard, where new frontier releases often move the ceiling by two or three points. AI Revolution's launch video leaned on this ranking for its headline that Fable 5.2 "ranks first worldwide", and that claim checks out against the live leaderboard.

The time figures need careful reading. Fable 5.2 at Max effort spends far longer per task (45.2 minutes against 25.0) because it thinks and acts for longer, yet it starts producing its answer sooner. In practice, Max-effort Fable is the slower model to finish a hard job, not the slower model to respond. For interactive chat at default settings, neither number reflects what you will feel.

List prices are the simple part. Converted at roughly 79p to the dollar:

One detail jumps out: Fable 5.2 on the Batch API costs exactly the same per token as Opus 5.5 at standard rates. If your workload can wait for asynchronous processing, such as overnight document analysis, test-suite generation or bulk classification, you can get the stronger model for the price of the weaker one.

The harder part is that token price is not the bill. A model that reasons for longer or takes more steps can cost more per finished task even at the same rate. AI Revolution's summary, that "lower token prices hide potentially higher bills", refers to Fable 5.2 being 20% cheaper per token than Fable 5.1 while potentially using more tokens. The same logic applies against Opus: at Max effort, Artificial Analysis measured Fable 5.2 at $26.54 per task against $10.04 for Opus 5.5, a ratio of about 2.6 to 1, not 2 to 1.

For a team running 1,000 hard agentic tasks a month at Max effort, that difference is roughly £21,000 ($26,540) against £7,900 ($10,040). That is the wrong way to run Fable, as the next section shows, but it is how many teams will run it if they simply swap the model string and leave every setting at maximum.

Both models expose an effort setting (Low, Medium, High, Xhigh and Max) that controls how much reasoning and how many tool calls the model spends. Medium is the default in the Claude apps and Claude Code. Anthropic's launch page publishes accuracy-against-cost curves for several benchmarks, and they change the comparison completely:

• FrontierCode 2.0: Fable 5.2 at High effort scores 45.3%, essentially level with Opus 5.5's best score of 45.5%, "for roughly half the cost per task". At Xhigh, Fable reaches 49.1%, the highest of any released model Anthropic tested.
• WANDR (agentic search): at Medium effort, Fable 5.2 matches Opus 5.5's top score for about 40% of the cost per task.
• Humanity's Last Exam (with tools): at Low effort, Fable 5.2 matches Opus 5.5's top score for less than one-tenth of the cost per task.
• Chartography (no tools): at Medium effort, Fable 5.2 scores 66.4%, ahead of Opus 5.5 at Max effort (64.4%), for less than one-tenth of the cost per task.
• BenchCAD 2.1: at Max effort Fable 5.2 scores 83.5 against 65.3 for Opus 5.5, a gap that no effort setting on Opus closes.

The practical reading is that the fair comparison is not Fable Max against Opus Max. It is Fable at a lower effort against Opus at a higher one. On Anthropic's curves, a lower-effort Fable frequently reaches the same score as a maximum-effort Opus for less money, because it gets there in fewer, better-chosen steps. That reverses the headline price story: per finished task at a fixed quality bar, Fable 5.2 can be the cheaper model.

Two caveats apply. These curves are Anthropic's, on Anthropic's chosen benchmarks, and the company has an obvious interest in showing its premium model as cost-effective. And cost per task depends heavily on your prompts, tools and context sizes. A coding agent that reads a 200,000-token repository on every turn will see input costs dominate in a way that benchmark harnesses do not capture. If you want a primer on getting more from the effort controls, our guide to prompting Claude Opus 5 still applies to both models.

Benchmarks tell you about averages. The first creator tests show what the difference feels like on individual builds, with the obvious caveat that a handful of prompts is not a controlled study.

Bart Slodyczka ran the most structured comparison: identical prompts to both models across four builds, graded blind. The rounds were an open-world city game, five motion-graphics videos, an "un-slop" round (a website, a pitch deck and a blog post judged on how generic they felt) and a Blender truck model, followed by a refusals segment and a final tally of score and cost. The format matters more than any single verdict: blind grading removes the bias of knowing which model is the expensive one, and tracking cost alongside quality is exactly the trade-off most teams need to judge.

AI Search pushed Fable 5.2 across a wide range of creative and technical demos, from animation and full song production to an Unreal game, physics reasoning, deep research and a segment on specifications and "nerfing", before ending with benchmarks and usage limits. Matt Wolfe published a launch-day reaction calling the model "quite insane". In both cases, these are capability showcases for Fable rather than side-by-side tests against Opus.

A consistent thread across launch coverage is that Fable's advantage shows up most clearly on tasks that run long: multi-file builds, 3D scenes with many parts, research that needs several rounds of searching. On short, well-specified tasks, creators found Opus 5.5 output hard to tell apart. That mirrors the benchmark pattern above, where the largest gaps sit on long-horizon tasks.

The two models differ in a way that rarely appears in comparison tables: what happens when you ask about something sensitive. Fable 5.2 runs behind classifiers that block most cybersecurity work and a set of biology topics. Anthropic says blocked cyber queries fall back to Opus 4.8 and blocked biology queries to Opus 5 in its first-party products and for opted-in API users, and that this behaviour can vary across cloud providers.

This has two consequences for a Fable-versus-Opus decision:

• Security teams may not get Fable at all for their core work. A request that trips the cyber classifier is answered by an older model, so the benchmark gap between Fable and Opus 5.5 is irrelevant for that request. Organisations that need frontier cyber capability are pointed towards Mythos 5.2 through Anthropic's Cyber Verification Program.
• Refusals have improved but not vanished. Anthropic says Fable 5.2 triggers about 55% fewer blocks overall than Fable 5.1, with around 60% fewer in cybersecurity and 40% fewer in biology. Bart Slodyczka's video includes a dedicated refusals segment for exactly this reason: false positives on benign prompts are a real productivity cost.

On the alignment side, the joint system card reports that Anthropic deliberately limited Fable 5.2's skill at accelerator-kernel optimisation and LLM architecture research to about Opus 5.5 level. For those two narrow areas, the stronger model is intentionally no better than the cheaper one. We covered the reasoning behind that cap, and the METR findings that prompted it, in METR on Anthropic's AI R&D acceleration.

Most teams should not pick one model. They should route. A simple three-step policy captures most of the value:

1. Default to Opus 5.5 (or Sonnet 5.5) at Medium effort for interactive work, short tasks and anything high-volume. This keeps the baseline bill predictable.
2. Escalate to Fable 5.2 at Medium effort when a task is long-running, multi-file, or has already failed once on Opus. Anthropic's curves suggest a Medium-effort Fable often matches or beats a Max-effort Opus, so escalating the model while lowering effort can cost less than simply cranking Opus to Max.
3. Reserve Fable at High or Xhigh for the genuinely hard cases: migrations, gnarly concurrency bugs, research agents, CAD and robotics work. Avoid Max as a default; on Artificial Analysis it more than doubles time per task.

Track two numbers while you do this: cost per accepted task (not per call) and the rate at which a human has to redo the work. Those are the metrics that decide whether the premium model pays for itself. If you want to see how other teams structure their stacks, our Claude hub collects every Anthropic model review and guide in one place.

If you are weighing OpenAI and Google as well, read our GPT-6 Sol vs Claude Opus 5.5 comparison and Gemini 4 Carbon review. Gemini 4 Carbon beats both Claude models on AutomationBench (53.9%), and GPT-6 Astra sits between them on FrontierSWE v2 (65.5%). The rumoured next OpenAI release is covered in GPT-6.2 Sol: rumour checked.

• Vendor benchmarks: every row in the head-to-head table and every cost-curve claim comes from Anthropic's launch page. Anthropic chose the benchmarks and the settings.
• RoboDojo Sim was run internally, according to Anthropic's own footnote, so the largest single gap in the table has no external replication yet.
• Artificial Analysis measures one blend of tasks at Max effort. It is independent, but its cost-per-task figure reflects its harness, not your agent loop.
• Creator tests are small samples. Blind grading helps, but four builds cannot establish a reliable win rate.
• Prices can change. Anthropic has cut prices on each recent generation. Check the live pricing page before budgeting.
• Exchange rates: sterling figures use an approximate rate of 79p to the dollar and will drift.

Claude Fable 5.2 is the better model, and the independent leaderboard agrees: first of 74 on Artificial Analysis against fifth for Opus 5.5. But "better" and "better value" are different questions. At maximum effort, Fable costs about 2.6 times more per task in independent testing. At sensible effort settings, Anthropic's own curves show it matching Opus 5.5's best scores for half the cost or less.

So the recommendation is not "switch" or "stay". It is: stop running anything at Max by default, route hard and long tasks to Fable 5.2 at Medium or High, keep Opus 5.5 for the bulk of everyday work, and measure cost per accepted task rather than price per token. Done that way, most teams will pay roughly the same as they do today and get noticeably better results on the work that matters most.

• Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (09/11/2026): benchmark table, effort and cost curves, safeguards.
• Anthropic: Introducing Claude Opus 5.5: Opus 5.5 launch details and imagery.
• Artificial Analysis: Claude Fable 5.2 and Claude Opus 5.5: Intelligence Index, speed, tokens and cost per task (accessed 11/11/2026).
• Bart Slodyczka on YouTube, AI Revolution on YouTube, AI Search on YouTube and Matt Wolfe on YouTube: creator tests and reaction.
• Images: Anthropic launch page artwork for Fable 5.2 and Opus 5.5, credited to Anthropic and compressed to AVIF.