Every task gets its own table

Overall

Intelligence Index — top 10

Artificial Analysis's composite across reasoning, coding, knowledge and agentic tasks. The single number people quote — read it as a map, then check the task tables.

Cheapest in the top 10 by AA's cost-per-task: GPT-6.1 Terra at $0.96 per task, then GPT-6.1 Sol at $1.27.

Reasoning

GPQA Diamond — two boards, two orders

Graduate-level science questions, the standard reasoning discriminator. The leaders cluster at 94–96% and swap places depending on who ran the eval — which is itself the finding.

AIME: current sources blend AIME 2025 and 2026 in single figures, so no AIME table is published here rather than mislabel a year.

Coding

Agentic coding & code review

The Claude Opus 5 family holds SWE-bench Verified outright. Terminal work is a different sport — there, the agent harness around the model decides as much as the model does.

Review ≠ generation: Kodus's CodeReviewBench puts Kimi K3.1 Pro first (F2 63.5) instead. Sonnet 5.5's Ironbee win costs $0.142 per PR reviewed. No credible public benchmark yet covers architecture judgment or explanation quality.

Writing

Creative writing

Human-preference arenas, two methodologies, two different winners — LMArena's crowd vote ties Gemini 4 and Opus 5.5 at the top; EQ-Bench's judged board puts GPT-6 Astra first.

Budget pick: MiMo-V2.6-Pro — 1468 arena score at $0.87 per million output tokens, the strongest writer under $1/M.

Research

Factual recall vs. finding things

SimpleQA measures whether a model knows; BrowseComp measures whether it can find. Different skills, different leaders — and BrowseComp's top scores are vendor-reported, which we flag wherever they appear.

Everyday chat

LMArena Text Arena — top 10

The most contested table on this page: ten models within 26 Elo of each other. General chat has no runaway winner.

Images

Text-to-image

One lab sweeps both arenas: OpenAI's GPT Image line takes the top five on LMArena. The split that matters is speed — Flare's median generation is ~30s; Sunburst at max quality runs ~2m15s. GPT Image 2, still #5, leaves the API on February 10, 2027.

Artificial Analysis estimates Sunburst and Flare at ~$0.21 per 1024×1024 image at max quality.

Long context

MRCR v2 — finding needles at scale

Retrieval across very long inputs. Muse Spark 1.3 doesn't just lead — in the 512K–1M band it scores 98.1% where the next model manages 73.8%. Retrieval isn't the whole of long-document work (summarization quality isn't directly measured here), but a gap this size is a result, not noise.

Value

Intelligence per dollar

Index points per dollar of blended API price. The leaders are small, fast, mostly open models — none of them is the smartest model here, and all of them are the right answer for high-volume work.

Routing

The decision models

A new model category: systems whose only job is picking what happens next — which model, which tool, allow or deny. Google's launch claims for Gemini Decision 1 (150–250ms, $0.10/M input, best-in-suite calibration) are vendor-reported. The independent Decision Index below is the check on them.

The general-model alternative: Claude Haiku 5.5 is explicitly positioned for classification and routing at $0.10/$0.50 per million tokens — a prompted small model instead of a dedicated decider. Which approach wins per workload is an open research question, and it's ours.

How this page is built: 15 research queries through Perplexity's Agent API with live search, cross-checked across the sources named in each table. Full method and narrative in Publication 001. Leaderboards move weekly; treat every table as a dated snapshot, not a standing verdict.