2026-09-28 · Anthropic
Claude Sonnet 5.5
Anthropic’s faster, lower-cost Sonnet refresh. Terminal-Bench 4.0 70.6%, GDPval-AA v2.1 1844, at $2 / $10 per million input / output tokens.
SourceAs of 2026-10-08 · XBench average
Ours · average of labs
XBench average.
Our board. Not theirs.
XBench scores are the mean of independent and aggregator labs after each lab is scaled 0–100. Rank here will not match Artificial Analysis or Arena.
XBench #1 is Opus 5.5. AA #1 is Opus 5.5. Arena #1 is Fable 5.
2026-10-07 · Anthropic
Claude Haiku 5.5
Anthropic’s newest small model for fast, high-volume work. $0.10 / $0.50 per million tokens up to 100k context; $0.50 / $2.50 above 100k. No independent board score recorded yet.
Source2026-09-22 · Anthropic
Claude Opus 5.5
First of the Claude 5.5 line. AA Intelligence v4.3.2 58, clear #1 by five points. $4 / $20 (~20% under Opus 5), cache reads $0.20. GDPval-AA v2.1 1846.
SourceClick a release row to highlight that model on the charts and score table. Click again to drop it. Purple is selected, not “this week.”
Left to right is the 2026 release timeline. Purple dots are models you picked in the table.
| Release | Company | Model | Rough positioning |
|---|---|---|---|
| Feb 2026 | Google DeepMind | Gemini 3.1 Pro | Frontier multimodal/reasoning |
| Apr 2026 | OpenAI | GPT-5.5 | Frontier reasoning + agents |
| Apr 2026 | Anthropic | Claude Opus 4.7 | Frontier coding/reasoning |
| Apr–May 2026 | SpaceXAI | Grok 4.3 / 4.20 | Frontier reasoning/agents |
| Apr 2026 | DeepSeek | DeepSeek-V4 | Frontier open/low-cost challenger |
| May 2026 | Anthropic | Claude Opus 4.8 | Frontier reasoning/coding |
| May 2026 | Alibaba | Qwen3.7-Max | Frontier Chinese model |
| Jun 2026 | Anthropic | Claude Fable 5 | New Claude 5 frontier tier |
| Jun 2026 | Anthropic | Claude Mythos 5 | New Claude 5 frontier tier |
| Jun 2026 | Anthropic | Claude Sonnet 5 | Frontier general-purpose/agentic |
| Jun 2026 | Zhipu AI | GLM-5.2 | Frontier/open-weight coding |
| Jun 2026 | ByteDance | Doubao Seed 2.1 Pro | Frontier reasoning |
| Jun 2026 | Meituan | LongCat-2.0 | Frontier open-weight agentic coding |
| Jul 2026 | SpaceXAI | Grok 4.5 | Frontier coding/agents |
| Jul 2026 | OpenAI | GPT-5.6 Sol | Top-tier flagship reasoning |
| Jul 2026 | OpenAI | GPT-5.6 Terra | Balanced GPT-5.6 |
| Jul 2026 | OpenAI | GPT-5.6 Luna | Cost-efficient GPT-5.6 |
| Jul 2026 | Google DeepMind | Gemini 3.6 Flash | Frontier-class fast reasoning |
| Jul 2026 | Moonshot AI | Kimi K3 | Frontier open-weight |
| Jul 2026 | Anthropic | Claude Opus 5 | Anthropic’s flagship frontier model |
| Aug 2026 | SpaceXAI | Grok 4.6 | Latest Grok frontier model |
| 13 Aug 2026 | Google DeepMind | Gemini 3.7 Flash | Latest Flash workhorse |
| 14 Aug 2026 | Z.ai | GLM-5.3 | Latest GLM coding / cyber-defense |
| 26 Aug 2026 | Z.ai | GLM-5.3-Flash | Open-weight GLM, MIT |
| 26 Aug 2026 | Alibaba | Qwen3.8-Flash-Next | Qwen4-architecture open preview |
| 28 Aug 2026 | Tencent | Hy4 Preview | Open Apache-2.0 frontier-adjacent |
| 1 Sep 2026 | Anthropic | Claude Fable 5.1 | New Claude 5 frontier tier |
| 2 Sep 2026 | Meta | Muse Spark 1.3 | Meta reaches the frontier; agent-native |
| 2 Sep 2026 | Google DeepMind | Gemini 3.8 Flash | Latest Flash workhorse |
| 3 Sep 2026 | OpenAI | GPT-6 Astra | OpenAI GPT-6 flagship |
| 10 Sep 2026 | DeepSeek | DeepSeek V4.1 Flash | Open MoE successor to V4 Flash/Pro |
| 21 Sep 2026 | SpaceXAI | Grok 4.7 | Latest Grok frontier model |
| 21 Sep 2026 | Xiaomi | MiMo-V2.6-Pro | Open MIT; new #1 open-weights on AA |
| 22 Sep 2026 | Anthropic | Claude Opus 5.5 | Anthropic’s 5.5 frontier flagship |
| 22 Sep 2026 | OpenAI | GPT-6 Sol | GPT-6 flagship tier |
| 22 Sep 2026 | OpenAI | GPT-6 Luna | Cost-efficient GPT-6 |
| 28 Sep 2026 | Anthropic | Claude Sonnet 5.5 | Fast general-purpose and coding model |
| 7 Oct 2026 | Anthropic | Claude Haiku 5.5 | Fast, cost-efficient small model |
XBench average
Each lab is min-max scaled to 0–100 on its own board, then averaged. Vendor first-party numbers are not in the average. XBench place needs at least two pillars (Intelligence, Agentic, Coding) so one SWE-Pro row cannot take #1. Placement is XBench’s, not a copy of AA or Arena.
Line · higher is better · same order as the table
Hover a column. Click a legend chip to hide a series. Purple dots are models picked in the release table.
Line · lower is better
Rank and cost. Top of the chart is better.
Line · API price USD / 1M
Input and output token price. Lower is cheaper.
| 1 | Opus 5.5 Anthropic · closed Anthropic’s 5.5 frontier flagship | 100XBench · 2 labs | 100Intel · 1 lab | — | 100Coding · 1 lab | $5.98Lab | AA #1Arena —BenchLM #1 | $4 / $20 |
| 2 | Fable 5.1 Anthropic · closed New Claude 5 frontier tier | 92.1XBench · 3 labs | 87.6Intel · 2 labs | 100Agentic · 1 lab | 88.6Coding · 2 labs | $7.63Lab | AA #2Arena #4BenchLM #2 | $10 / $50 |
| 3 | GPT-6 Astra OpenAI · closed OpenAI GPT-6 flagship | 87.4XBench · 3 labs | 87.8Intel · 1 lab | 86.4Agentic · 1 lab | 88Coding · 1 lab | $3.26Lab | AA #2Arena —BenchLM #6 | $10 / $50 |
| 4 | Fable 5 Anthropic · closed New Claude 5 frontier tier | 86.8XBench · 3 labs | 90.2Intel · 2 labs | 85.3Agentic · 1 lab | 85Coding · 3 labs | $8.75Lab | AA #5Arena #1BenchLM #4 | $10 / $50 |
| 5 | Opus 5 Anthropic · closed Anthropic’s flagship frontier model | 86.6XBench · 3 labs | 76.4Intel · 2 labs | 96.2Agentic · 1 lab | 87.2Coding · 3 labs | $5.86Lab | AA #4Arena #7BenchLM #5 | $5 / $25 |
| 6 | Spark 1.3 Meta · closed Agent-native Meta model; reaches the frontier | 85.3XBench · 3 labs | 75.6Intel · 1 lab | 95.2Agentic · 1 lab | 85.2Coding · 1 lab | — | AA #6Arena —BenchLM #12 | $1.25 / $4.25 |
| 7 | GLM-5.3 Z.ai · staged Latest GLM coding / cyber-defense | 79.7XBench · 3 labs | 65.9Intel · 2 labs | 90.4Agentic · 1 lab | 82.7Coding · 1 lab | $2.01Lab | AA #11Arena #10BenchLM #13 | — |
| 8 | Kimi K3 Moonshot AI · open Frontier open-weight | 78.9XBench · 3 labs | 66.3Intel · 2 labs | 84.5Agentic · 1 lab | 86Coding · 2 labs | $2.00Lab | AA #13Arena #9BenchLM #10 | $3 / $15 |
| 9 | Grok 4.6 SpaceXAI · closed Latest Grok frontier model | 75.8XBench · 3 labs | 43.2Intel · 2 labs | 90.4Agentic · 1 lab | 93.9Coding · 2 labs | $1.86Lab | AA #13Arena #22BenchLM #7 | $2 / $6 |
| 10 | GPT-5.6 Sol OpenAI · closed Top-tier flagship reasoning | 75.6XBench · 3 labs | 68.3Intel · 2 labs | 84.3Agentic · 1 lab | 74.3Coding · 3 labs | $1.99Lab | AA #8Arena #10BenchLM #24 | $5 / $30 |
| 11 | Qwen3.8 Max Alibaba · closed Hosted 2.4T flagship; weights promised | 73.9XBench · 3 labs | 64.3Intel · 2 labs | 82.4Agentic · 1 lab | 75.1Coding · 1 lab | $5.41Lab | AA #11Arena #13BenchLM #17 | $2 / $6 |
| 12 | 3.8 Flash Google DeepMind · closed Latest Flash workhorse | 73.4XBench · 3 labs | 69Intel · 2 labs | 64.6Agentic · 1 lab | 86.5Coding · 1 lab | $1.24Lab | AA #18Arena #5BenchLM #9 | $0.75 / $3.75 |
| 13 | Spark 1.2 Meta · closed Agent-native Meta model | 73.3XBench · 3 labs | 73.3Intel · 2 labs | 70.6Agentic · 1 lab | 76.1Coding · 1 lab | $0.97Lab | AA #21Arena #3BenchLM #16 | $1.25 / $4.25 |
| 14 | GPT-5.6 Terra OpenAI · closed Balanced GPT-5.6 | 68.7XBench · 3 labs | 48.7Intel · 2 labs | 70Agentic · 1 lab | 87.5Coding · 1 lab | $1.40Lab | AA #15Arena #20BenchLM #8 | $2 / $12 |
| 15 | 3.7 Flash Google DeepMind · closed Latest Flash workhorse | 68.3XBench · 3 labs | 64.1Intel · 2 labs | 54.7Agentic · 1 lab | 86Coding · 1 lab | $0.93Lab | AA #24Arena #6BenchLM #11 | $0.75 / $3.75 |
| 16 | Flash Next Alibaba · open Qwen4-architecture open preview | 67.1XBench · 2 labs | 56.1Intel · 1 lab | — | 78.1Coding · 1 lab | — | AA #21Arena —BenchLM #14 | — |
| 17 | Opus 4.7 Anthropic · closed Frontier coding/reasoning | 64.5XBench · 3 labs | 76.1Intel · 2 labs | 61.2Agentic · 1 lab | 56.1Coding · 3 labs | — | AA #18Arena #2BenchLM #25 | $5 / $25 |
| 18 | Opus 4.8 Anthropic · closed Frontier reasoning/coding | 63.7XBench · 3 labs | 60.6Intel · 2 labs | 67.7Agentic · 1 lab | 62.7Coding · 3 labs | $4.08Lab | AA #15Arena #13BenchLM #19 | $5 / $25 |
| 19 | Grok 4.5 SpaceXAI · closed Frontier coding/agents | 59.1XBench · 3 labs | 46.7Intel · 2 labs | 66.7Agentic · 1 lab | 63.8Coding · 2 labs | $1.04Lab | AA #24Arena #19BenchLM #15 | $2 / $6 |
| 20 | GPT-5.6 Luna OpenAI · closed Cost-efficient GPT-5.6 | 57.9XBench · 3 labs | 31.5Intel · 2 labs | 67.9Agentic · 1 lab | 74.3Coding · 1 lab | $0.18Lab | AA #28Arena #23BenchLM #18 | $0.20 / $1.20 |
| 21 | GLM-5.2 Zhipu AI · open Frontier/open-weight coding | 57.4XBench · 3 labs | 43.7Intel · 2 labs | 61Agentic · 1 lab | 67.4Coding · 1 lab | $0.96Lab | AA #32Arena #18BenchLM #21 | $1.40 / $4.40 |
| 22 | V4 Pro 0813 DeepSeek · open Current DeepSeek-V4 open challenger | 53.7XBench · 3 labs | 46.3Intel · 1 lab | 82.4Agentic · 1 lab | 32.3Coding · 3 labs | $0.67Lab | AA #30Arena —BenchLM #32 | — |
| 23 | Sonnet 5 Anthropic · closed Frontier general-purpose/agentic | 53.5XBench · 3 labs | 39.9Intel · 2 labs | 71.3Agentic · 1 lab | 49.3Coding · 3 labs | $5.09Lab | AA #27Arena #21BenchLM #26 | $2 / $10 |
| 24 | 3.6 Flash Google DeepMind · closed Frontier-class fast reasoning | 53.3XBench · 3 labs | 50.1Intel · 2 labs | 41.5Agentic · 1 lab | 68.4Coding · 1 lab | $0.93Lab | AA #32Arena #15BenchLM #19 | $1.50 / $7.50 |
| 25 | GPT-5.5 OpenAI · closed Frontier reasoning + agents | 49.6XBench · 3 labs | 41.9Intel · 2 labs | 56.6Agentic · 1 lab | 50.4Coding · 2 labs | $0.69Lab | AA #37Arena #12BenchLM #29 | $2.50 / $15 |
| 26 | V4 Flash DeepSeek · open Open price-performance pick | 45.7XBench · 3 labs | 43.9Intel · 1 lab | 65.8Agentic · 1 lab | 27.4Coding · 3 labs | $0.31Lab | AA #31Arena —BenchLM #34 | $0.14 / $0.28 |
| 27 | 3.1 Pro Google DeepMind · closed Frontier multimodal/reasoning | 39.4XBench · 3 labs | 50.8Intel · 2 labs | 0Agentic · 1 lab | 67.4Coding · 1 lab | $0.67Lab | AA #34Arena #7BenchLM #21 | $1 / $6 |
| 28 | MiniMax M3 MiniMax · open Open-weight frontier-adjacent | 34.3XBench · 3 labs | 29.3Intel · 1 lab | 43Agentic · 1 lab | 30.5Coding · 2 labs | $0.51Lab | AA #35Arena —BenchLM #29 | $0.60 / $2.40 |
| 29 | Qwen3.7 Max Alibaba · closed Frontier Chinese model | 33.7XBench · 3 labs | 38.4Intel · 2 labs | 28.5Agentic · 1 lab | 34.1Coding · 3 labs | $1.15Lab | AA #35Arena #17BenchLM #27 | $1.48 / $4.42 |
| 30 | Inkling Thinking Machines · open New open lab, Apache-2.0 | 33.3XBench · 3 labs | 58.5Intel · 1 lab | 29.4Agentic · 1 lab | 12Coding · 3 labs | $0.34Lab | AA #18Arena —BenchLM #33 | $0.50 / $2.02 |
| 31 | Hy3 Tencent · open Open Apache-2.0 coding/reasoning | 32XBench · 3 labs | 22Intel · 1 lab | 32.1Agentic · 1 lab | 42Coding · 1 lab | $0.04Lab | AA #37Arena —BenchLM #28 | $0.13 / $0.53 |
| 32 | MiMo V2.5 Pro Xiaomi · open Open MIT, cheap per-task | 26XBench · 3 labs | 22Intel · 1 lab | 26Agentic · 1 lab | 29.9Coding · 2 labs | $0.03Lab | AA #37Arena —BenchLM #31 | $0.43 / $0.87 |
| 33 | Grok 4.20 SpaceXAI · closed Frontier reasoning/agents | 18.2XBench · 2 labs | 36.4Intel · 2 labs | — | 0Coding · 2 labs | — | AA #37Arena #16BenchLM #35 | $1.25 / $2.50 |
| 34 | Grok 4.3 SpaceXAI · closed Frontier reasoning/agents | 8.1XBench · 3 labs | 9.8Intel · 2 labs | 14.5Agentic · 1 lab | 0Coding · 1 lab | — | AA #41Arena #24BenchLM #36 | $1.25 / $2.50 |
| — | GPT-6 Sol OpenAI · closed GPT-6 flagship tier | — | 75.6Intel · 1 lab | — | — | — | AA #6Arena —BenchLM — | $2 / $10 |
| — | GPT-6 Luna OpenAI · closed Cost-efficient GPT-6 | — | 48.8Intel · 1 lab | — | — | — | AA #28Arena —BenchLM — | $0.10 / $0.50 |
| — | Grok 4.7 SpaceXAI · closed Latest Grok frontier model | — | 70.7Intel · 1 lab | — | — | — | AA #9Arena —BenchLM — | $2 / $6 |
| — | Qwen3.8 open Alibaba · open Open-weight Qwen3.8 (AA #2 open) | — | 56.1Intel · 1 lab | — | — | $2.16Lab | AA #21Arena —BenchLM — | — |
| — | GLM-5.3 Flash Z.ai · open Open-weight GLM, MIT | — | 61Intel · 1 lab | — | — | $0.25Lab | AA #15Arena —BenchLM — | $0.15 / $0.50 |
| — | V4.1 Flash DeepSeek · open Open MoE successor to V4 Flash/Pro | — | 53.7Intel · 1 lab | — | — | $0.27Lab | AA #24Arena —BenchLM — | $0.15 / $0.60 |
| — | Mythos 5 Anthropic · closed New Claude 5 frontier tier | — | — | — | 85.6Coding · 2 labs | — | AA —Arena —BenchLM #3 | $10 / $50 |
| — | MiMo V2.6 Pro Xiaomi · open Open MIT; #1 open-weights on AA | — | 70.7Intel · 1 lab | — | — | — | AA #9Arena —BenchLM — | $0.43 / $0.87 |
| — | Hy4 Tencent · open Open Apache-2.0 coding/reasoning | — | — | — | 36.5Coding · 1 lab | — | AA —Arena —BenchLM #23 | $0.83 / $2.50 |
| — | Doubao Seed ByteDance · closed Frontier reasoning | — | 0Intel · 1 lab | — | — | — | AA #43Arena —BenchLM — | — |
| — | LongCat 2.0 Meituan · open Frontier open-weight agentic coding | — | 4.9Intel · 1 lab | — | — | $0.06Lab | AA #42Arena —BenchLM — | — |
45 models on this board · # is this page’s rank · purple = picked in the release-dates table
Line graph of rank on this page versus XBench, AA, Arena, and BenchLM. Top of the chart is #1.
Place across boards. Top of the chart is #1. Purple dots are models picked in the release table.
XBench composite bars
Mean of Intelligence, Agentic, and Coding pillars when present. Click a release-date row to highlight that model on this page.
Models left to right by this metric. Purple dots are models picked in the release table.