Data: 19 September 2026

Independent · every number dated & sourced

The top 10 AI models
in the world, right now

Ranked by blind human preference — millions of head-to-head votes, not marketing. In five minutes you'll know exactly which model wins at what you need.

402 models tracked256,733 votes behind the top 102 independent sources0 paid placements, ever

Updated 19 September 2026 · what changed

Pick your job

One winner for every job

Tap a card to meet the current champion — its price, its strengths and its honest limits. No paywalls, no sign-up.

The ranking

The top 10, by real human votes

Ranked by each model's best-performing variant on LMArena's overall board (style control on). The bar under each name shows the score's real 95% confidence interval — a top-3 gap this small is a statistical tie. Variants that differ only by effort setting share one entry; separate model versions rank separately.

The full field

Every model we track

14 honest cards — what each model is best at, what it is not the pick for, and what it costs. Filter by what you need.

Claude Fable 5Anthropic #1 Overall#1 Coding The strongest all-rounder in the world right now, by human preference. Arena Score 1505.7$10/$50 Honest read Claude Opus 4.6Anthropic The proven workhorse: near-top quality with the deepest vote history in the top 3. Arena Score 1504.6$5/$25 Honest read Claude Opus 4.7Anthropic The newest Opus iteration — statistically a tie with Opus 4.6. Arena Score 1501.7$5/$25 Honest read Muse Spark 1.2Meta Frontier-adjacent quality at roughly a tenth of Fable pricing. Arena Score 1499.6$1.25/$4.25 Honest read Claude Fable 5.1Anthropic Best reasoning The highest measured reasoning score we track. Arena Score 1498.5$10/$50 Honest read Muse Spark 1.3Meta Meta's newest Spark generation — strong, fast, still fresh. Arena Score 1493.1$1.25/$4.25 Honest read Gemini 3.8 FlashGoogle The fast daily driver of the top 10. Arena Score 1493$0.75/$3.75 Honest read Claude Opus 5Anthropic Anthropic's newest base line — top-10 quality on a huge vote base. Arena Score 1492.9$5/$25 Honest read Muse Spark 1.1Meta The stability pick: proven scores, cheap, still in the top 10. Arena Score 1492.6$1.25/$4.25 Honest read Gemini 3.7 FlashGoogle One honest step behind its sibling 3.8 — at the same price. Arena Score 1489.8$0.75/$3.75 Honest read GPT-Image 2.5 (Sunburst)OpenAI Best image The best image generator in the world, by blind human preference. Arena Score 1420.7 · image Honest read Gemini Omni 1.1 FlashGoogle Best video The best video generator in the world, by blind human preference. Arena Score 1514.6 · video$1.5/$9 Honest read Kimi K3Moonshot Best open The strongest open-weights model in the world. Arena Score 1484.8$3/$15 Honest read GLM 5.3 FlashZ AI Best value The value champion: the most intelligence per dollar we measured. Intelligence Index 41.9$0.15/$0.50 Honest read

Straight answers

Questions people actually ask

Which is the best AI model right now?

On the retrieval date of this page (19 September 2026), Claude Fable 5 leads LMArena's overall board with an Arena Score of 1506 from 30,000+ blind human votes. But the top four are separated by fewer than 5 points — a statistical near-tie. Always check the date on any leaderboard you read.

What is the best AI model for images and video?

GPT-Image 2.5 Sunburst tops the text-to-image board (Arena Score 1421) and Gemini Omni 1.1 Flash tops text-to-video (1515). Both boards are young — thousands of votes so far, not millions — so expect the leaders to move.

How much does the best AI model cost?

Prices vary hugely. Claude Fable 5 costs $10 per million input tokens and $50 per million output. Claude Opus 4.6 and 4.7 cost half that at statistically near-identical quality, and Meta's Muse Spark models cost a fraction again. For pure value, GLM 5.3 Flash returns about 176 intelligence-index points per $1M tokens (blended 3:1 price).

What is the best open-weight AI model?

Kimi K3 from Moonshot is the highest-ranked open-weights model on the overall board (Arena Score 1485, Kimi K3 License). You can download and run it yourself; GLM 5.3 Flash (MIT license) is the best-value open pick.

How often does this ranking change?

The boards move daily. We refresh the data weekly, stamp every page with the retrieval date, and log every change in the changelog — if it isn't logged, it didn't happen.

Can companies pay to rank higher here?

No. There is no paid placement and no affiliate links — not now, not ever. Ranks come only from the public measurements described in the methodology.

Why trust this list?

No paid placement.
Not now, not ever.

Read the methodology