The frontier is a photo finish. Everyone else is racing for fourth.
Gemini 4 Carbon (high) leads with a 1-point margin, but it has no public API. The best model you can actually call is Claude Opus 5.5 (max w/ fallback) at 58. US labs hold 9 of the top 10 slots, and China's best trails the leader by 13 points.
- Top score59Gemini 4 Carbon (high)
- Lead over #21Claude Opus 5.5 (max w/ fallback)
- Best with public API58Claude Opus 5.5 (max w/ fallback)
- Frontier-tier models7Score 50+
- US-held top-10 slots9Out of the top 10 ranks
- Field median32Across 28 ranked models
Market trends
Three models within a whisker, then a long slide
Every model in the snapshot, ranked by index score. Change the colour key to see who owns the ranking and which models you can't buy.
Index score of every model
- 1Gemini 4 Carbon (high)59no API
- 2Claude Opus 5.5 (max w/ fallback)58
- 3GPT-6 Astra (max)57
- 4GPT-6.1 Sol (max)55
- 5Claude Sonnet 5.5 (max w/ fallback)54
- 6GPT-6.1 Terra (max)51
- 7Muse Wonder Full (max)50no API
- 8Claude Fable 5.1 (max w/ fallback)49
- 9Kimi K3.1 Pro46
- 10Muse Wonder (max)43
- 11Claude Haiku 5.5 (max)36
- 12Grok 4.7 (xhigh)34
- 13Qwen3.8 Max (0902)34
- 14DeepSeek V4.1 Flash 1017 (max)32
- 15GLM-5.3 (max)32
- 16Gemini 3.8 Flash (high)31
- 17MiMo-V2.6-Pro30
- 18Ling 3.1 Flash 101929
- 19MiniMax-M3.1-Flash29
- 20Step 5 Preview29
- 21GPT-6 Luna (max)28
- 22Mistral Large 4 (high)26
- 23Muse Spark Open (xhigh)25
- 24Qwen3.8 27B (xhigh)20
- 25K2 Horizon 375B A23B18
- 26Beam 1.018
- 27Nemotron 3.5 Super VL14
- 28Inkling (xhigh)12
Market structure
Four US labs own the frontier tier
Ranks only tell you who won. Depth tells you who can keep winning: how many strong models a lab ships and how many top-10 slots it holds.
Lab depth
Lab scorecard
One row per lab, sorted by its best model. Bar shows the best score.
| Lab | Best | Average | Models | Top 10 |
|---|---|---|---|---|
| 59 | 45 | 2 | 1 | |
| Anthropic | 58 | 49.3 | 4 | 3 |
| OpenAI | 57 | 47.8 | 4 | 3 |
| Meta | 50 | 39.3 | 3 | 2 |
| Moonshot AI | 46 | 46 | 1 | 1 |
| xAI | 34 | 34 | 1 | 0 |
| Alibaba | 34 | 27 | 2 | 0 |
| DeepSeek | 32 | 32 | 1 | 0 |
| Z.ai | 32 | 32 | 1 | 0 |
| Xiaomi | 30 | 30 | 1 | 0 |
| Ant Group | 29 | 29 | 1 | 0 |
| MiniMax | 29 | 29 | 1 | 0 |
| StepFun | 29 | 29 | 1 | 0 |
| Mistral | 26 | 26 | 1 | 0 |
| MBZUAI | 18 | 18 | 1 | 0 |
| Beam | 18 | 18 | 1 | 0 |
| NVIDIA | 14 | 14 | 1 | 0 |
| Thinking Machines | 12 | 12 | 1 | 0 |
Where the field sits
Models per score band. The frontier band is thin; the crowd lives in the middle.
- Frontier50+7
- Contender40–493
- Mid-tier25–3913
- Lightweightunder 255
- The three busiest labs hold 8 of the top 10 slots.
- Only 7 models clear the frontier band; all of them are American.
Geopolitics
China's labs cluster a full tier below the US frontier
US models spread from the very top to the very bottom. Chinese models bunch tightly in the low thirties, with one breakout.
US and China compared
Every model, by country
Each dot is a model. The tick marks each group's median; the dashed line marks the leader.
US vs China, head to head
Best against best, then the typical model against the typical model.
- Best US model
- 59
- Best Chinese model
- 46
- Median US model
- 46
- Median Chinese model
- 30
- China's top-10 slots
- 1
Kimi K3.1 Pro is the only non-US model in the top 10, at rank 9.
Financial impact
The price of going small is steep
Every lab sells a cheaper tier. This shows how much capability you give up to get it: each line runs from a lab's weakest model in the snapshot to its strongest.
Flagship versus lightest model
Flagship vs lightest model, per lab
Labs with two or more models in the snapshot, longest drop first.
- OpenAI4 models28−29
- Google2 models31−28
- Meta3 models25−25
- Anthropic4 models36−22
- Alibaba2 models20−14
What this means for spend
- Steepest drop: OpenAI, 29 points from flagship to lightest model. Budget tiers are a different product, not a discount.
- Paying for the best callable model costs you 1 point against the unreleased leader. The premium for the gated model buys almost nothing.
- The top three sit within 2 points. At the frontier, negotiate on price and latency, not capability.
- China's best scores 46, beating Claude Haiku, Gemini Flash and GPT-6 Luna outright. That is the cost-arbitrage zone.
The index has no price data. Pair these gaps with per-token pricing (see the models page and the cost calculator) before committing spend.
Technological risks and opportunities
What to act on
Opportunities and risks
Opportunities
- The frontier is buyable: 5 of 7 frontier-tier models have a public API.
- Low lock-in at the top. Three vendors within 2 points makes multi-model routing cheap to adopt.
- A non-US option exists near the top: Kimi K3.1 Pro at rank 9.
- The mid-tier is crowded, so prices there should fall fastest. Half the field scores 32 or less.
Risks
- The #1 model is gated. 2 models in the snapshot have no public API, and both sit in the frontier tier.
- Concentration: three labs hold 8 of the top 10 slots. Supplier risk sits with very few companies.
- Small-model cliff: going light costs up to 29 points. Agentic and coding work will feel it first.
- Geopolitical exposure: the US–China gap at the top is 13 points, and export rules can widen or close it overnight.
Scores transcribed from the Artificial Analysis Intelligence Index v5.0 chart shared on Oct 29, 2026 (nine evaluations including Terminal-Bench 4.0, FrontierCode 2.0, HLE-Diamond and GDP.pdf). “No public API” follows the chart's checkered marking. Lab and country tags, score bands and all commentary are this report's own analysis, not Artificial Analysis's. Index data: Artificial Analysis. For the live leaderboards see every leaderboard on one page.
Frequently asked questions
Who leads the Artificial Analysis Intelligence Index v5.0?
Gemini 4 Carbon (Google) leads with 59, one point ahead of Claude Opus 5.5 at 58 and two ahead of GPT-6 Astra at 57. Gemini 4 Carbon has no public API, so the best model you can call today is Claude Opus 5.5.
How far behind is China?
China's best model, Kimi K3.1 Pro, scores 46 at rank 9, 13 points behind the leader. It is the only non-US model in the top 10. The typical Chinese model scores in the low thirties.
Are these your test results?
No. The scores are transcribed from an Artificial Analysis chart shared on 29 October 2026. The lab and country tags, score bands and commentary are this report's own analysis, and BestSuperintelligent did not re-run any evaluation.
Does the index include price?
No. The index scores capability only. Pair the score gaps with per-token pricing before deciding where to spend; our models page lists the list prices we hold.