Skip to content
AIKI

AI Benchmarks Explained: 14 Tests Behind Claude Fable 5.2

AI benchmarks explained: what each of the 14 tests behind Claude Fable 5.2 measures, what it scored, and which matter most for your business.

6 min lesing

Key takeaways

  1. An AI benchmark is a standardized test where every model gets the same tasks, so results can be compared.
  2. Anthropic showed 14 benchmarks when it launched Claude Fable 5.2 on 9 November 2026, and the article explains what each one measures.
  3. For most businesses, AutomationBench, AA-Briefcase, Remote Labor Index and OSWorld are closest to daily work.
Woman marking a stack of test papers with a red pen at an office desk in Oslo, while a glowing white sphere with a star floats above the desk next to a stopwatch and a results table

An AI benchmark is a standardized test where every model gets the same tasks, so the results can be compared. The number in the table is usually the share of tasks the model solved, and higher is better. When Anthropic launched Claude Fable 5.2 on 9 November 2026, it showed 14 such tests. Here is what each one measures, what Fable 5.2 scored, and which ones say the most for a Norwegian business.

What is an AI benchmark?

A benchmark consists of three things: a fixed set of tasks, a rule for what counts as correct, and a score. The score comes in three forms:

  • Share solved: the percentage of tasks the model got fully right. Most of the tests use this.
  • Partial credit: the model earns points for each sub-goal it reaches, even if the full task stays unfinished.
  • Elo: a ranking built on pairwise comparisons, the way chess players are ranked. Two deliverables are judged against each other, and the winner moves up.

How to read a benchmark table

  • The rows are the tests, the columns are the models. Higher numbers are better in every row.
  • The thinking level (effort) is listed under the number. It runs from Low through Medium, High and Xhigh to Max, and says how long the model was allowed to think. A higher level gives better results and a higher cost per task. In Anthropic's table, the highest level applies unless stated otherwise.
  • A dash means not measured. No number exists, so the cell tells you nothing about the model's quality.
  • The vendor chooses the setup. The numbers on a launch page are the vendor's own. Independent evaluators such as Artificial Analysis run all the models with the same setup, which makes their results safer to compare. The Qwen4-Max article puts Alibaba's launch figures next to Artificial Analysis's measurements of the same model.
  • Some rows concern a different model. Fable 5.2 and Mythos 5.2 are built on the same base model, but Fable has strong safeguards against misuse in cybersecurity and biology. The cybersecurity and protein design rows are therefore measured on Mythos 5.2, which only approved organizations may use.

All 14 tests in one table

TestWhat it measuresFable 5.2Best of the others (excluding Mythos)
FrontierCode 2.0Whether code changes would be accepted in open-source projects49.1%45.5% (Opus 5.5)
FrontierSWE v234 very long programming assignments69.9%65.5% (GPT-6 Astra)
AutomationBenchBusiness workflows across business systems51.4%53.9% (Gemini 4 Carbon)
AA-Briefcase v1.1Long analysis projects with hundreds of source files per task66.1% and 1,785 Elo61.8% (Gemini 4 Carbon) and 1,800 Elo (Opus 5.5)
Remote Labor IndexReal, paid freelance assignments41.3%21.3% (Opus 5.5)
OSWorld 2.1Long tasks on an ordinary computer89.3%87.2% (Opus 5.5)
WANDRLarge assignments involving data collection82.1%72.3% (Opus 5.5)
ChartographyReading charts and figures73.4%72.5% (Gemini 4 Carbon)
Terminal-Bench-Science 0.1Research tasks in a terminal window74.6%68.1% (GPT-6 Astra)
Humanity's Last Exam2,500 expert questions across more than a hundred subjects74.2%67.7% (Opus 5.5)
De novo binder designProtein design in 24 hours85.1% (Mythos 5.2)82.6% (Opus 5.5)
CyScenarioBenchMulti-step attacks in simulated networks76.3% (Mythos 5.2)Not comparable
CVE RediscoveryKnown security vulnerabilities the model finds on its own89.6% (Mythos 5.2)81.3% (Opus 5.5)
RoboDojo SimControlling a simulated two-armed robot64.9%39.4% (Opus 5.5)

All figures are Anthropic's own, from the launch page. Opus 5.5 was measured without safety safeguards on the cybersecurity and protein design rows. How OpenAI's successor to GPT-6 Astra scores on several of these tests is covered in GPT-6.2 Astra and Sol.

Which tests measure coding?

FrontierCode 2.0

FrontierCode 2.0 measures whether the code changes the model produces would be accepted by the people who maintain open-source projects. That is a stricter standard than passing tests: the code must also be good enough for an experienced developer to merge it. Fable 5.2 scored 49.1 percent at the second-highest thinking level (Xhigh).

FrontierSWE v2

FrontierSWE v2 is 34 very long programming assignments from Proximal Labs, covering systems development, performance optimization, scientific computing and AI research. One of the assignments is to build a circuit simulator from scratch. The model can work for many hours on each task, and each task scores 0 to 100 percent.

Which tests measure workflow and office work?

AutomationBench

AutomationBench is Zapier's test of whether an AI agent can carry out a complete business workflow across 47 simulated business systems in sales, marketing, operations, customer service, finance and HR. The tasks are based on real workflows from Zapier's platform. The model gets a single instruction and must find the right functions in systems such as CRM, inbox and calendar, follow a set of rules and steer clear of misleading data. Only the end result counts, and a task is solved only when all the final checks pass. Here Gemini 4 Carbon is ahead of Fable 5.2.

AA-Briefcase v1.1

AA-Briefcase is Artificial Analysis's test of knowledge work over several weeks. Each project has many connected tasks and hundreds of source files per task: emails, Slack threads, meeting minutes, documents and large data exports. The percentage is the share of checkpoints in the scoring sheet that are met. The Elo figure combines the checkpoints with an assessment of how good the analysis and presentation are. On Elo, Opus 5.5 is slightly ahead of Fable 5.2.

Remote Labor Index

Remote Labor Index is real, paid freelance assignments from many industries, collected by the Center for AI Safety and Scale AI. Each assignment has a brief, files and a deliverable made by a professional. The score is the share of assignments where the deliverable reaches a level that would have been accepted as commissioned work.

Which tests measure computer use, search and images?

OSWorld 2.1

OSWorld 2.1 from XLANG Lab measures whether the model can solve long tasks on a real computer, controlled with screenshots, mouse and keyboard. Anthropic reports the figure for a subset of the tasks ("offline set") with partial credit ("partial"), so it cannot be compared directly with full OSWorld results.

WANDR

WANDR is Perplexity's test of agents on large assignments involving data collection. The figure in the table uses a more lenient scoring rule ("relaxed score"), and there Fable 5.2 is almost ten percentage points ahead of Opus 5.5.

Chartography

Chartography measures how accurately the model reads charts and figures, without auxiliary tools. The test says something about how reliably a model can pull numbers out of reports and presentations. The gap between the models is small here: Gemini 4 Carbon is less than one percentage point behind. Google reports 73.5 percent for the newer Gemini 4 Neon.

Which tests measure research and reasoning?

Terminal-Bench-Science 0.1

Terminal-Bench-Science 0.1 is 70 research tasks from life sciences, physics, earth sciences, mathematics and engineering, led by researchers at Stanford. The model works in a terminal window, and a task is solved when the result passes task-specific tests that domain experts have reviewed. Mythos 5.2 scored 75.3 percent, close to Fable 5.2.

Humanity's Last Exam

Humanity's Last Exam is 2,500 questions across more than a hundred subjects, written by subject-matter experts and collected by the Center for AI Safety and Scale AI. The questions were chosen because the best models could not answer them when they were collected. Fable 5.2 scored 74.2 percent with tools, meaning with access to aids such as search along the way. Figures without tools are lower and cannot be compared directly.

De novo binder design (24 hours)

De novo binder design means designing new proteins that bind to a specific target protein. Anthropic reports the result for Mythos 5.2 with a 24-hour time limit, and gives no further description on the launch page.

Which tests measure cybersecurity and robotics?

CyScenarioBench

CyScenarioBench from the security company Irregular measures whether a model can plan and carry out complete multi-step attacks in simulated networks, where each task builds on the previous one. Each lab runs its own tasks, so the figures cannot be compared across vendors.

CVE Rediscovery

CVE Rediscovery is an internal test at Anthropic. A CVE is a publicly recorded security vulnerability. The model is given a software project as it looked before the vulnerability was fixed, and the test measures how many known vulnerabilities it finds on its own.

RoboDojo Sim

RoboDojo Sim measures how well the model controls a simulated two-armed robot from camera images. Anthropic ran the test itself. Fable 5.2 scored 64.9 percent, more than twice Fable 5.1's 27.6 percent.

Which tests say the most for a Norwegian business?

For most businesses, four of the tests are closest to everyday work:

  • AutomationBench for automating workflows between business systems.
  • AA-Briefcase and Remote Labor Index for analysis, reports and deliverables.
  • OSWorld for agents that need to use ordinary programs on screen.

If you build your own software, FrontierCode and FrontierSWE matter most. The tests on research, cybersecurity, protein design and robotics show where the technology is heading, but say less about an ordinary office day. None of the 14 tests is in Norwegian.

Remote Labor Index shows how fast things are moving. When the test launched in October 2025, the best agent completed 2.5 percent of the assignments. In Anthropic's table, Fable 5.1 scored 17.9 percent and Opus 5.5 scored 21.3 percent. Fable 5.2 scored 41.3 percent at a high thinking level, measured in Anthropic's own setup.

Freelance assignments delivered well enough to be acceptedFreelance assignments delivered well enough to be accepted: 25 av 1 000 før, 413 etterFreelance assignments delivered wellenough to be acceptedRemote Labor Index, per 1,000 assignments. Differentsetups and possibly a different test versionBest agent, October 2025: 2.5 percentFable 5.2, November 2026: 41.3 percent

The best yardstick is still your own tasks. Pick 20 real tasks where you already know what a good answer looks like, run each model three times, and let the person who owns the task today assess the answers. A test like that tells you more about which model fits than any table, because it also tests Norwegian language and your own systems. If you want help with the setup, see AI consultant or AI development. The comparison of ChatGPT and Claude shows the differences in daily use.

Sources

Del:LinkedInXFacebook

New articles by email

We use your email address to send you a confirmation link. Only after you confirm do we add you to the newsletter list at Resend, which sends on our behalf. You can unsubscribe in every email. Read the privacy policy.

Start with a pilot

  • Price after scoping

    excl. VAT, fixed price

  • One task

    One defined automation in the systems you already use

  • 4 weeks

    until the solution is live

Agreed goal before start. If we miss it, we fix it for free.

Afterwards: AI Partner at Price after scoping (normally 6 months' commitment), and the pilot amount is deducted from the first month. Or you stop.

In new projects, personal data is only processed on servers in the EU. Read about data and security