Four flagships shipped in 72 hours. Here's the board.
Claude Fable 5.1, GPT-6 Astra, Muse Spark 1.3, Gemini 3.8 Flash — 7 points apart on capability, 100× apart on price. Sort by what your workload actually needs. Vote for your pick; the crowd column is this browser's ballot box.
QUALITY KING
CLAUDE FABLE 5.1 · 100
SPEED DEMON
CELERIS-1 ULTRA · 285 T/S
CHEAPEST SEAT
DEEPSEEK V4 FLASH · $0.28/M
| # / MOVE | MODEL | CROWD | ||||||
|---|---|---|---|---|---|---|---|---|
1▲+1 | Claude Fable 5.1 Anthropic · Sep 2026 · 1M · Hardest agent work | 100 | 1535 | 82.4 | 58 | $50 | 2 | 0 |
2— | GPT-6 Astra OpenAI · Sep 2026 · 1.05M · Math + computer use | 99 | 1528 | 79.1 | 60 | $50 | 2 | 0 |
3▲+1 | Claude Opus 5 Anthropic · Jul 2026 · 1M · Agentic coding value | 99 | 1524 | 78.5 | 105 | $25 | 4 | 0 |
4▼-1 | GPT-5.6 Sol OpenAI · Jul 2026 · 1.05M · Frontier reasoning | 98 | 1516 | 76.5 | 98 | $30 | 3.3 | 0 |
5NEW | Gemini 3.8 FlashNEW Google · Sep 2026 · 1M · Fast frontier reasoning | 97 | 1512 | 78.2 | 175 | $3.25 | 29.8 | 0 |
6NEW | GLM-5.3-FlashNEWOPEN Z.ai · Sep 2026 · 1M · Open-weight speed leader | 96 | 1508 | 76.8 | 195 | $2.85 | 33.7 | 0 |
7▲+2 | Muse Spark 1.3 Meta · Sep 2026 · 1M · Coding value king | 96 | 1492 | 76 | 184 | $4.25 | 22.6 | 0 |
8▼-2 | Kimi K3OPEN Moonshot AI · Jul 2026 · 1M · Open coding frontier | 95 | 1502 | 74.5 | 58 | $15 | 6.3 | 0 |
9▼-1 | Gemini 3.1 Pro Google · Apr 2026 · 1M · Science + long ctx | 95 | 1501 | 70.2 | 135 | $12 | 7.9 | 0 |
10NEW | Celeris-1 UltraNEW GroqCloud · Sep 2026 · 256K · Sub-second agent loops | 94 | 1475 | 73.1 | 285 | $5.4 | 17.4 | 0 |
11— | Claude Sonnet 5 Anthropic · Jun 2026 · 1M · Daily driver | 93 | 1482 | 72.5 | 102 | $15 | 6.2 | 0 |
12▼-2 | GPT-5.6 Terra OpenAI · Jul 2026 · 1M · Balanced tier | 93 | 1490 | 70.8 | 122 | $12 | 7.8 | 0 |
13NEW | Mercury 2NEW Inference Labs · Sep 2026 · 512K · Ultra-low latency | 92 | 1468 | 71 | 245 | $2.4 | 38.3 | 0 |
14▲+3 | DeepSeek V4 FlashOPEN DeepSeek · Sep 2026 · 1M · Price floor disruptor | 92 | 1478 | 74.2 | 210 | $0.28 | 328.6 | 0 |
15▲+1 | Qwen 3.8 Max Alibaba · Sep 2026 · 1M · Agent workflows | 89 | 1445 | 69.5 | 105 | $4.8 | 18.5 | 0 |
16▼-1 | Grok 4.5 SpaceXAI · Jul 2026 · 500K · Engineering focus | 89 | 1452 | 71.8 | 115 | $6 | 14.8 | 0 |
17— | GPT-5.6 Luna OpenAI · Jul 2026 · 1M · Cheap frontier | 84 | 1438 | 63.4 | 192 | $1.2 | 70 | 0 |
18▼-1 | Llama 4 MaverickOPEN Meta · Apr 2025 · 128K · Open ecosystem | 82 | 1285 | 60.5 | 140 | $0.6 | 136.7 | 0 |
19— | Gemma 4 27BOPEN Google · Apr 2026 · 128K · Self-host | 78 | 1365 | 62 | 148 | — | — | 0 |
20▼-2 | Claude 3.5 Haiku Anthropic · Oct 2024 · 200K · Legacy fast | 76 | 1240 | 56.2 | 178 | $4 | 19 | 0 |
METHOD
QUALITY = COMPOSITE 0–100 · ARENA = LMARENA ELO · CODING = AGENT INDEX · VALUE = QUALITY ÷ $/M OUT · MOVE = PERIOD-OVER-PERIOD RANK DELTA VS PRIOR CYCLE. CROWD = YOUR LOCAL BALLOT, STORED IN THIS BROWSER ONLY.
SOURCES · 2026-09-22
- ARTIFICIAL ANALYSIS INTELLIGENCE, CODING & SPEED INDICES (LATE SEP 2026)
- LMARENA PUBLIC ELO SNAPSHOTS VIA INDEPENDENT BENCHMARK SWEEPS
- VENDOR LAUNCH BENCHMARKS: ANTHROPIC, OPENAI, GOOGLE DEEPMIND, Z.AI, DEEPSEEK, GROQCLOUD
- PROVIDER PRICING PAGES (BLENDED INPUT/OUTPUT $/M TOKENS)
ROUTE YOUR WORK: FLASH FOR BULK · SPARK FOR CODE · FABLE/ASTRA FOR THE HARDEST 10% · READ THE THESIS →