AI Model Leaderboard — Compare LLMs by Price, Quality & Speed
Gradient's opinionated leaderboard of frontier AI models. Weekly-refreshed quality, pricing and speed data, with our own composite Gradient Index to rank them for real-world use.
This Week's Signal
- Mistral Large 4 output price halved to $2.10, a 49.8% cut that undercuts every top-five model on cost.
- GPT-6 Astra intelligence rose 4.2 points to 56.9, second only to Claude Opus 5.5 at 58.4, but costs $50 output.
- GPT-6.1 Sol gained 2.7 points to 54.5, matching Sonnet 5.5's intelligence at the same $10 output price.
- Anthropic still holds the top two Gradient slots; Gemini 3.8 Flash remains the cheap outlier at $3.75 with a 30.5 intelligence score.
- Data is flagged stale, so treat these deltas as directional until the next refresh.
Gradient's Picks Editor choices, refreshed every Sunday
Claude Sonnet 5.5
Within 2 points of Opus 5.5 at half the price ($10 vs $20/M output), with 45% faster output.
GLM-5.3-Flash
41.8 on the Intelligence Index for $0.50/M output — about 73% of the top score at 1/40th the price. Open weights.
MiMo-V2.6-Pro
The highest-scoring open-weight model (46.3), at $0.87/M output.
Claude Sonnet 5.5
Anthropic- Quality (AA)
- 54.2
- Output $/M
- $10
- Input $/M
- $2.00
- Speed
- 136 tok/s
- Context
- 1M
- Released
- Sep 2026
Within 2 points of Opus 5.5 at half the price, and faster.
Claude Opus 5.5
Anthropic- Quality (AA)
- 58.4
- Output $/M
- $20
- Input $/M
- $4.00
- Speed
- 98 tok/s
- Context
- 1M
- Released
- Sep 2026
Highest Intelligence Index on Artificial Analysis. At Max effort it can think for minutes before answering.
GPT-6.1 Sol
OpenAI- Quality (AA)
- 54.5
- Output $/M
- $10
- Input $/M
- $2.00
- Speed
- 80 tok/s
- Context
- 1M
- Released
- Sep 2026
OpenAI's mainstream flagship. Same price as Sonnet 5.5 and Gemini 4 Argon, slower output.
GPT-6 Astra
OpenAI- Quality (AA)
- 56.9
- Output $/M
- $50
- Input $/M
- $10
- Speed
- 81 tok/s
- Context
- 1M
- Released
- Sep 2026
OpenAI's premium tier. Same price as Fable 5.1, a few points lower on both indexes.
Gemini 3.8 Flash
Google- Quality (AA)
- 30.5
- Output $/M
- $3.75
- Input $/M
- $0.75
- Speed
- 124 tok/s
- Context
- 1M
- Released
- Sep 2026
Highest score in the Fast tier, pricier than the rest of it.
Claude Haiku 5.5 (Medium)
Anthropic- Quality (AA)
- 24.6
- Output $/M
- $0.50
- Input $/M
- $0.10
- Speed
- 160 tok/s
- Context
- 1M
- Released
- Oct 2026
Listed at Medium effort. At Max it scores 43.4 but takes ~4 min to start answering.
GLM-5.3-Flash
Z AI- Quality (AA)
- 27.8
- Output $/M
- $0.50
- Input $/M
- $0.15
- Speed
- 69 tok/s
- Context
- 1M
- Released
- Aug 2026
About 73% of Opus 5.5's score for 1/40th of the output price.
MiMo-V2.6-Flash
Xiaomi- Quality (AA)
- 26.7
- Output $/M
- $0.28
- Input $/M
- $0.14
- Speed
- 42 tok/s
- Context
- 1M
- Released
- Sep 2026
Lowest output price in this list for a near-40 score.
Muse Wonder
Meta- Quality (AA)
- 43.3
- Output $/M
- $20
- Input $/M
- $4.00
- Speed
- 150 tok/s
- Context
- 1M
- Released
- Oct 2026
Meta's newest flagship (Oct 12), and the fastest frontier model here.
Step 5 Preview
StepFun- Quality (AA)
- 28.5
- Output $/M
- $2.70
- Input $/M
- $1.00
- Speed
- 131 tok/s
- Context
- 1M
- Released
- Sep 2026
Still a preview. Close to GLM-5.3 at about 60% of its price.
MiMo-V2.6-Pro
Xiaomi- Quality (AA)
- 29.9
- Output $/M
- $0.87
- Input $/M
- $0.43
- Speed
- 42 tok/s
- Context
- 1M
- Released
- Sep 2026
Highest-scoring open-weight model, under $1/M output. Slow output.
Mistral Large 4
Mistral- Quality (AA)
- 26.0
- Output $/M
- $2.10
- Input $/M
- $0.70
- Speed
- 169 tok/s
- Context
- 524K
- Released
- Oct 2026
Mistral's new flagship (Oct 6). The only European lab in this list.
Gemini 4 Argon
Google- Quality (AA)
- 50.4
- Output $/M
- $20
- Input $/M
- $4.00
- Speed
- 80 tok/s
- Context
- 1M
- Released
- Sep 2026
Google's new flagship. Starts answering in ~23 s, far quicker than its frontier peers.
Muse Spark 1.3
Meta- Quality (AA)
- 36.4
- Output $/M
- $4.25
- Input $/M
- $1.25
- Speed
- —
- Context
- 1M
- Released
- Sep 2026
Frontier-level score at under half Sonnet 5.5's price.
Grok 4.7
SpaceXAI- Quality (AA)
- 34.3
- Output $/M
- $6.00
- Input $/M
- $2.00
- Speed
- 80 tok/s
- Context
- 500K
- Released
- Sep 2026
Cheapest frontier option at $6/M output, with a quick time to first answer.
GPT-6 Luna (High)
OpenAI- Quality (AA)
- 21.4
- Output $/M
- $0.50
- Input $/M
- $0.10
- Speed
- 114 tok/s
- Context
- 1M
- Released
- Sep 2026
Same price as Haiku 5.5, slightly lower score, slower output.
GLM-5.3
Z AI- Quality (AA)
- 31.7
- Output $/M
- $4.40
- Input $/M
- $1.40
- Speed
- 78 tok/s
- Context
- 1M
- Released
- Aug 2026
Strong open-weight all-rounder with a 74.8 Coding Index.
Claude Fable 5.1
Anthropic- Quality (AA)
- 49.0
- Output $/M
- $50
- Input $/M
- $10
- Speed
- 73 tok/s
- Context
- 1M
- Released
- Sep 2026
Highest Coding Index on Artificial Analysis (81.6). Premium pricing.
Qwen3.8 Max
Alibaba- Quality (AA)
- 33.6
- Output $/M
- $6.00
- Input $/M
- $2.00
- Speed
- 38 tok/s
- Context
- 983K
- Released
- Sep 2026
Alibaba's flagship. Strong 76.2 Coding Index, slow output.
Ling 3.1 Flash
InclusionAI- Quality (AA)
- —
- Output $/M
- $0.90
- Input $/M
- $0.30
- Speed
- 217 tok/s
- Context
- 1M
- Released
- Oct 2026
Fastest open-weight model here, with a 41.1 score.
Muse Spark Open
Meta- Quality (AA)
- 24.7
- Output $/M
- $4.25
- Input $/M
- $1.25
- Speed
- 155 tok/s
- Context
- 1M
- Released
- Oct 2026
Meta's open-weight release (Oct 15). Same price as Spark 1.3, ~9 points lower.
Gemini 3.5 Flash-Lite
Google- Quality (AA)
- 11.7
- Output $/M
- $2.50
- Input $/M
- $0.30
- Speed
- 409 tok/s
- Context
- 1M
- Released
- Jul 2026
For high-volume work where latency matters more than score.
MiniMax-M3
MiniMax- Quality (AA)
- 18.7
- Output $/M
- $1.20
- Input $/M
- $0.30
- Speed
- 117 tok/s
- Context
- 1M
- Released
- Jun 2026
Fast for its price, now outscored by cheaper open models.
Qwen3.8-Flash-Next
Alibaba- Quality (AA)
- —
- Output $/M
- $0.47
- Input $/M
- $0.15
- Speed
- 53 tok/s
- Context
- 256K
- Released
- Aug 2026
Priced like GLM-5.3-Flash, slightly lower score, 256k context.
DeepSeek V4 Pro
DeepSeek- Quality (AA)
- 23.1
- Output $/M
- $3.96
- Input $/M
- $1.32
- Speed
- 81 tok/s
- Context
- 1M
- Released
- Aug 2026
DeepSeek's open-weight flagship, now behind GLM-5.3 and MiMo on the index.
Mercury 2.5
Inception- Quality (AA)
- —
- Output $/M
- $0.75
- Input $/M
- $0.20
- Speed
- —
- Context
- 260K
- Released
- Sep 2026
Fastest model here by far. Low score: keep it for simple tasks.
DeepSeek V4.1 Flash
DeepSeek- Quality (AA)
- —
- Output $/M
- $1.20
- Input $/M
- $0.30
- Speed
- —
- Context
- 1M
- Released
- Sep 2026
Open weights, 180 tokens/s, answers in ~12 s.
Kimi K2.7 Code
Kimi- Quality (AA)
- —
- Output $/M
- $4.00
- Input $/M
- $0.95
- Speed
- 48 tok/s
- Context
- 256K
- Released
- Jun 2026
Purpose-built for coding (60.8 Coding Index).
GPT-5.3 Codex
OpenAI- Quality (AA)
- —
- Output $/M
- $14
- Input $/M
- $1.75
- Speed
- 100 tok/s
- Context
- 400K
- Released
- Feb 2026
Purpose-built for agentic coding. Still OpenAI's latest Codex on AA.
Kimi K3
Kimi- Quality (AA)
- —
- Output $/M
- $15
- Input $/M
- $3.00
- Speed
- —
- Context
- 1M
- Released
- Jul 2026
Open-weight frontier contender, but priced like a closed model.
How we built this leaderboard
Where the data comes from
All quality scores, benchmarks and latency figures are pulled weekly from Artificial Analysis, which runs standardized evaluations across providers on their own infrastructure. Prices come from official API pricing pages. We cache the last good pull and mark the page as stale if the scrape fails — we never silently serve unknown data.
The Gradient Index, explained
A composite score from 0 to 100, computed as 50% quality (percentile rank on Artificial Analysis's intelligence index), 35% value (percentile rank of inverse output price), and 15% speed (percentile rank of median output tokens per second). It is deliberately opinionated: quality dominates, but a model that is twice as smart and a hundred times more expensive does not win automatically.
What we exclude
Deprecated model versions. Research previews without a public API. Self-hosted-only models that no inference provider will run for you. Models we cannot independently verify a price for. Our goal is a list a technical founder could act on this afternoon.
Why we still show Arena-style scores with caveats
Intelligence indices aggregate several benchmarks — GPQA, AIME, HLE, IFBench, HumanEval, LiveCodeBench, MATH, MMLU-Pro — which partially mitigates the known limits of any single score. We display them as one signal, not as gospel. If a number surprises you, the Signal archive is where we argue about why.
— The Gradient editors