Ask which AI model is "the best" and you get a marketing answer. Ask which model is best at a specific task and the leaderboards hand the crown to a different model almost every time — and sometimes disagree with each other about that, too. This study collects the current per-task leaders from public benchmarks and arenas, keeps every number tied to its source, and publishes the disagreements instead of blending them away.
Method. Fifteen research queries run through Perplexity's Agent API with live web search, one per category, cross-checked across the leaderboards named below. All figures are a snapshot of November 15, 2026. The tables from this study also live, in full, on our Rankings page.
Writing
LMArena's Creative Writing arena ties Google's Gemini 4 line and Anthropic's Opus 5.5 at 1533/1532. EQ-Bench's judged board disagrees and puts GPT-6 Astra first — methodology, not noise: crowd preference and rubric judging reward different things.
LMArena Creative Writing
ELO · NOV 2026| # | Model | Score |
|---|---|---|
| 1 | Gemini 4 Argon (High) | 1533 |
| 1 | Gemini 4 Carbon (High) | 1533 |
| 3 | Claude Opus 5.5 (High) | 1532 |
| 4 | Claude Fable 5.1 (Max) | 1514 |
| 5 | Gemini 4 Hydrogen (High) | 1509 |
EQ-Bench Creative Writing v3: GPT-6 Astra 2176.9, Kimi K3.1 Pro 2172.6, Claude Fable 5.1 2155.3. Budget leader: MiMo-V2.6-Pro, 1468 arena points at $0.87 per million output tokens.
Math & science reasoning
GPQA Diamond's leaders cluster at 94–96%, and the order depends on whose run you read. LLMBenchmark's board has Gemini 3.1 Pro Preview first; the Artificial Analysis-run board has GPT-6 Astra first at 96.3%. Both are on the Rankings page. AIME figures in current sources blend two competition years into one number, so this study publishes no AIME table rather than mislabel one.
Agentic coding
The clearest result in the study. Measured SWE-bench Verified results are an Anthropic podium, and the harness effect is just as visible: the best Terminal-Bench 4.1 result (65.8%) belongs to an agent+model pair, not a model alone.
SWE-bench Verified
MEASURED · BENCHGAP / BENCHLM| # | Model | Resolved |
|---|---|---|
| 1 | Claude Opus 5 | 96.0% |
| 2 | Claude Mythos 5 | 95.5% |
| 3 | Claude Fable 5 | 95.0% |
| 4 | Ember-1 | 92.2% |
| 5 | Sakana Fugu-Ultra | 91.8% |
Code review
Review is not generation, and the rankings shuffle accordingly. Ironbee's benchmark (50 real pull requests) is led by Claude Sonnet 5.5 at low effort — F2 64.4 at $0.142 per PR — ahead of Opus 5.5 (62.1) and Grok 4.7 (61.9). Kodus's CodeReviewBench instead puts Kimi K3.1 Pro first (63.5). And a genuine gap in the field: no defensible public ranking exists for architecture judgment or code explanation. The evaluations test bug-finding, not judgment.
Research accuracy
Two skills hide under "research": recalling facts (SimpleQA, independent board, top score 74.8%) and finding hard-to-locate answers on the live web (BrowseComp, vendor-reported, clustered at 91–92%). The 17-point gap between the two boards says more about measurement than about models.
Recall vs. retrieval
SIMPLEQA · BROWSECOMP| Benchmark | Leader | Score |
|---|---|---|
| SimpleQA | Gemini 3.1 Pro Preview | 74.8% |
| SimpleQA | GPT-6 Astra | 73.1% |
| BrowseComp vendor-reported | GPT-5.6 Sol Ultra | 92.2% |
| BrowseComp vendor-reported | LongCat-2.5 | 91.9% |
| BrowseComp vendor-reported | GPT-6 Astra | 91.5% |
Everyday chat
LMArena's Text Arena (Nov 11) is led by Gemini 4 Carbon (1526) and Argon (1522), then Opus 5.5 (1514) and Fable 5.1 (1512). The entire top ten fits inside 26 points — general chat is the most contested category there is, and the least useful one for choosing a model.
Image generation
OpenAI's GPT Image line sweeps both major arenas. GPT Image 2 still ranks fifth on LMArena (1382) — and leaves the API on February 10, 2027. Its successors split by temperament: Sunburst is #1 on quality but takes ~2m15s per image at max quality; Flare is #2 at a ~30-second median. Artificial Analysis estimates both at roughly $0.21 per 1024×1024 image at max quality — a price point that breaks products metered for the old model.
Long context
On MRCR v2 (8-needle), Muse Spark 1.3 posts 98.5% overall and 98.1% in the 512K–1M band, where the next model manages 73.8%. The caveat stays attached: MRCR and RULER measure finding information in long inputs — evidence about retrieval, not a complete measure of summarization.
Overall intelligence & value
Artificial Analysis's Index is led by Claude Fable 5.2 (Max) at 65. On value — index points per dollar — the leaders are small open models: inclusionAI Ling 3.1 Flash (608.9), Z.ai GLM 5.3 Flash batch (561.3), Upstage Solar Mini 4 (275.4). And AA's cost-per-task metric puts GPT-6.1 Terra (Max) cheapest among the ten highest-scoring models at $0.96 per task. Three different "best overall" answers, all defensible — which is the point of this study.
Routing itself
The category this company exists for. Among general models, Claude Haiku 5.5 is explicitly positioned for classification and routing at $0.10/$0.50 per million tokens. Dedicated decision models are now a category of their own — and the independent numbers are humbling: on AA's Decision Index, Gemini Decision 1 scores 66, exactly the field median, at $345 per million tasks and 211ms, while TypeSafe's Jev 1.14 scores 66.2 at 92ms and $156. Google's launch claims (best-in-suite calibration, 150–250ms, $0.10/M input) are vendor-reported; the independent index tells a more even story.
The one-sentence version: there is no best model — there is a best model per task, the gap between first and fifth is usually a few points, and the only way to answer well is to route. That's the finding this company will keep testing.
Sources
- LMArena Creative Writing & Text Arena leaderboards (snapshots, Nov 11–15, 2026)
- EQ-Bench Creative Writing v3 results (updated Nov 14, 2026)
- LLMBenchmark.ai — GPQA Diamond leaderboards (Nov 15, 2026)
- Benchgap & BenchLM — SWE-bench Verified; Harbor — Terminal-Bench 4.1; Artificial Analysis — Terminal-Bench 4.0
- Ironbee, "The Honest Raw LLM Code Review Benchmark"; Kodus CodeReviewBench; SWE-PRBench
- Kaggle — SimpleQA leaderboard (Oct 20, 2026); TensorFeed & Steel — BrowseComp leaderboards
- LMArena Text-to-Image Arena; Artificial Analysis Image Arena & AA-Image-T2I v2.1 (Nov 15, 2026)
- LLMBoard & BenchLM — MRCR v2 results (Nov 13–15, 2026)
- Artificial Analysis — Intelligence Index v5 leaderboard (Nov 15, 2026); SmophyAI value ranking (Nov 14, 2026)
- Google — Gemini Decisions API announcement (Nov 11, 2026); Artificial Analysis — Decision Index; Stackness decision-model comparison (Nov 13, 2026)
- OpenAI — Images 2.6 announcement, API pricing, and API deprecations (gpt-image-2 removal, Feb 10, 2027)