What Benchmark Categories Does RankedAGI Use to Rank AI Models?

Updated

RankedAGI groups its rankings into four capability areas: Code, Agentic, Reason, and Text Arena. Each area contains several individual benchmarks, and models are ranked both within each benchmark (with a numeric rank like #1, #2) and across the category. The site also tracks metadata for every model — size, version, release date, context length, knowledge cutoff, license, and input/output cost per million tokens — so a ranking can be read alongside the conditions that produced it.

Code benchmarks

These measure a model's ability to write, edit, and reason about software.

Benchmark What it roughly measures
SWE Bench Verified Resolving real GitHub issues in a verified subset
SWE Bench Pro A harder, professional-grade variant of SWE Bench
SWE Bench Multilingual The same task type across multiple programming languages
Terminal Bench 2.1 / 2.0 Completing tasks in a terminal environment
Code Arena Head-to-head coding comparisons
DeepSWE Deep software-engineering tasks
Svelte Bench Front-end work in the Svelte framework
Cyber Gym Security-oriented coding tasks

In the sample data, Claude Fable 5 leads SWE Bench Pro at 94.2% (#1), while Claude Opus 5 leads SWE Bench Multilingual at 94.2% (#1) — showing that the top spot can differ by benchmark even within one category.

Agentic benchmarks

Agentic scores reflect how well a model operates as an autonomous agent — browsing, using tools, and completing multi-step tasks.

  • RankedAGI Browse Comp — web browsing and information retrieval
  • OSWorld Verified — controlling a desktop operating system
  • τ2-Bench Telecom — task completion in a telecom domain scenario
  • Vending Bench 2 — a simulated transactional environment
  • GDPval AA and GDPval AA v2 — economic-value-oriented task evaluation

Reasoning benchmarks

These test abstract problem-solving rather than task execution.

  • HLE and HLE w/ Tools — Humanity's Last Exam, with and without tool access
  • GPQA Diamond — graduate-level science questions
  • AIME 2026 — competition mathematics
  • ARC AGI 2.0 — abstract pattern reasoning
  • MMMU, MMMU Pro, and MMMU Pro w/ Tools — multimodal reasoning
  • NYT Connections — word-association puzzle solving

The "w/ Tools" variants matter: a model's HLE or MMMU Pro score can change substantially depending on whether it may call external tools, so compare like with like.

Text Arena

Text Arena is a separate category for general text-generation quality, typically derived from head-to-head human or model comparisons rather than a fixed answer key.

How to read a ranking

  1. Pick the category that matches your use case (code, agentic work, reasoning, or general text).
  2. Within that category, choose the benchmark closest to your actual task — SWE Bench Verified for bug-fixing, Terminal Bench for shell work, GPQA Diamond for science Q&A.
  3. Check the model's metadata row (context length, cutoff, cost, license) before treating a high score as usable for you.
  4. Note that the site labels some entries as Simulated or Preview data, and the interface includes a "Show Simulated Data" toggle — treat simulated figures as illustrative rather than measured.

Because ranks are per-benchmark, a model can be #1 in one column and outside the top ten in another. The category view is a convenience grouping, not a single composite score.