The Arena Alignment Index has been updated!
We've refined how our signals detect model failures, providing broader coverage and more reliable judgements. Based on the updated methodology:
- @OpenAI continues to dominate, holding all seven top positions.
- @Anthropic's latest release, Claude Fable 5.2, debuts at #8 with a score of 78.2. Claude Opus 5.5, which ranked #6 on the previous index, is now #17.
- Kimi K3.1 Pro by @Kimi_Moonshot is the biggest upgrade over its predecessor (more info in thread).
- Most open source models are trailing closed releases. The best open model - Kimi K3.1 Pro - occupies #9, and the second best, Mimo-V2.6-Flash, is #26 on the Index, followed by Hy4, Beam, Inkling, DeepSeek and GLM.
The Top-6 lab:
- #1 @OpenAI GPT-6.2 Sol
- #2 @AnthropicAI Claude Fable 5.2
- #3 @Kimi_Moonshot Kimi K3.1 Pro
- #4 @GoogleDeepmind Gemini 4 Hydrogen
- #5 @SpaceXSI Grok 4.7
- #6 @AIatMeta Muse Ember
Alignment Index measures how agents behave in real user sessions across three dimensions:
- Unauthorized Action: Doing what the user didn’t ask for or approve
- Deceptive Completion: Saying work is done, checked or working when it isn't
- False Attribution: Crediting the user with something they never said or did
More insights in the thread🧵
Rapid Router retweeted
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
Today, @AnthropicAI is the #1 lab in the Code Arena: WebDev. Its best model, Claude Fable 5.2, is scoring 1824 points, +27 points ahead of the next-best lab.
See how @AnthropicAI’s ranking among labs has changed over the last year, and also see where it lands on the Agent Arena Index at the link in thread.
World models are now an API call, not a demo.
Genie 4 just hit the Gemini API:
720p at 24fps with native stereo audio
~₹9.40 per second, ~₹564 per minute
Input at ₹140 per 1M tokens
Prototype fast, but always call complete(). The meter runs until you do.
2
174
Voice agents just got cheaper to leave running.
GPT-Live-1.1 is live in the OpenAI API:
Active: ₹4.66/min
Standby: ₹1.87/min, 60% less than active
Same price as GPT-Live-1, sharper transcription, steadier accents across languages
Idle time stops eating your voice bill.
1
133
GPT-6.2 Astra beats GPT-6 Astra, and costs less. Sol is the real bill cut.
On the Agent Arena Index, ranks checked live on 19 Nov, matching Arena's post earlier today:
- GPT-6.2 Astra (Max) is #6 at 1291, +9 vs GPT-6 Astra (Max) at #7 with 1282, and OpenAI's best model on this board
- GPT-6.2 Sol (Max) is #11 at 1241, matching #12 GPT-6.1 Sol (Max), and on the cost Pareto frontier
Live median cost per task, 19 Nov: ₹297 for 6.2 Astra vs ₹459 for GPT-6 Astra. Sol is ₹90 vs ₹98 for GPT-6.1 Sol, and about 70% below 6.2 Astra. Arena's post used ₹275 and ₹84 for the 6.2 medians; the live board has moved.
Results for the 6.2 models are based on 7.1K+ real agentic sessions. Arena's post also had 6.2 Astra at #5 in Praise vs Complaint (1462 pts). Time per task was not published for the 6.2 models.
If you want OpenAI's best score on this index, 6.2 Astra is the new default, and it undercuts GPT-6 Astra. If 1241 is enough, Sol does that work at ₹90 a task.
₹ at 93.51/USD, 19 Nov. Source: Arena Agent Arena Index. Not a Rapid Router test.
1
275
Gemini 4 Neon almost caught Opus, at less than half the task cost.
Live Agent Arena Index, 19 Nov: Gemini 4 Neon (High) is #3 at 1333, three points behind Claude Opus 5.5 (Max) at 1336. Claude Fable 5.2 (Max) still leads at 1448. Median cost per task: ₹243 vs ₹551 vs ₹856 (Neon, Opus, Fable). Time: 8.4 vs 17.2 vs 19.5 min. Yesterday's Arena post had Neon at #2 with 1336; the board has moved.
Artificial Analysis, 18 Nov:
- Intelligence Index: Neon (xhigh) 61 vs Carbon (high) 60, 3rd overall
- ₹602 vs ₹726 per task (~17% less, Carbon at standard non-discounted price), ~21 vs ~36 min, ~110 vs ~57 output tokens/sec
- Coding Agent Index, Antigravity CLI: Neon (xhigh) 59 vs 58 for Opus 5.5 (max) and Fable 5.2 (xhigh), at ₹642 vs ₹1,179 and ₹1,447
API list price: ₹280.54 input and ₹1,402.70 output per 1M tokens, 1M context.
Neon is a small bump over Carbon with a lower bill. On coding agents, it sits with the Claude leaders for a lot less per task.
₹ at 93.51/USD, 19 Nov. Sources: Arena and Artificial Analysis. Not a Rapid Router test.
1
1
249
Today, we’re excited to introduce the Agent Arena Index, our new headline measure of overall agent performance, alongside six new focused indexes: Agentic Work, Coding, Chat, Task Outcomes, Tool Use, and Steering.
Together with the Alignment Index, these indexes provide a comprehensive view of how agents perform across real-world tasks, from completing complex workflows to using tools and adapting to user feedback.
Arena measures agentic performance from millions of real-world tasks by our global community. With the launch of the new indexes, we are updating our scores to provide a more stable and interpretable way to understand how agents perform in the real world.
Claude Fable 5.2 (Max) leads the Agent Arena Index by 134 points with a score of 1463. Top 5 labs on the Agent Arena Index:
- #1 @AnthropicAI: With Claude Fable 5.2 (Max) at 1463 pts
- #2 @GoogleDeepMind: With Gemini 4 Carbon (High) at 1307 pts
- #3 @OpenAI: With GPT-6 Astra (Max) at 1274 pts
- #4 @AIatMeta: With Muse Wonder Full (Max) at 1200 pts
- #5 @Alibaba_Qwen: With Qwen4-Max (xHigh) at 1191 pts
The Agent Arena Index combines four key signals of real-world agent performance:
- Task Completion: explicit confirmation of task success from users
- Praise vs. Complaint: user satisfaction with agent's outcome
- Bash Reliability: how well an agent uses terminal tools
- Steerability: how well an agent responds to user corrections
Alongside capability, the Alignment Index measures how faithfully agents act in accordance with user intent, based on three key signals: Deceptive Completion, Unauthorized Action, and False Attribution.
As agents take on increasingly complex, long-horizon work, measuring their real-world capabilities and alignment is more important than ever.
Explore the rankings and see what each index measures in the thread below.
The cheap everyday agent is now open weight. The hard benches are not.
Hy4 flash is open weight: 220B params, only 12B active, 1M context. These are Tencent's own evals, not an independent lab test.
On Tencent's public agent suite:
- DeepSWE v1.1: 64.6 vs 73.8 for DeepSeek-V4.1-Flash and 64.3 for Hy4 preview
- Terminal-Bench 4.0: 19.2 vs 40.0 and 26.3
- APEX-Agents v1.1: 63.1 vs 72.9 and 53.9
- WorkIndex: 42.1 vs 39.7 for Hy4 preview and 51.4 for Hy4, at 13.8x and 24.5x lower task cost
Blind test (85 users, 259 tasks): 2.79/4 vs 2.81 for Claude Haiku 5.5 and 3.09 for DeepSeek-V4.1-Flash. Cost per task: ₹3.66 vs ₹66.65 vs ₹16.91. External models ran at the high tier.
TokenHub list price, per 1M tokens: ₹7.05 input, ₹28.18 output, ₹0.28 cached.
For everyday work, Hy4 flash is the cheap open-weight default. For harder coding agents, Tencent still points you to Hy4, and DeepSeek still leads several of these benches.
₹ at 14.09/¥, 18 Nov. Source: Tencent Hy4-flash model card and TokenHub list prices.
1
1
543
Rapid Router retweeted
MAI Realtime 1 outperforms GPT-Realtime 2.1 and starts at just $0.01 per audio minute. msft.it/6018xEmwC
Rapid Router retweeted
make it yours: introducing custom domains in ai studio
turn your idea into a live, branded app with a custom domain, all within ai studio build
ai.studio/build to get started
Rapid Router retweeted
We updated the API usage dashboard to make it easier to monitor your usage and investigate issues.
In Analytics, you can now:
• Better track usage and spend
• Check rate limits and rate-limited requests
• Explore errors by model, project, or API key
Rapid Router retweeted
Frontier performance. A fraction of today’s prices. Built to run 24/7.
Meet the new MAI models, launching today at #MSIgnite: aka.ms/MAI-Ignite-2026
Rapid Router retweeted
Upgrades, people. Upgrades.
The wait is over, you can now use GPT-6.2 Astra and GPT-6.2 Sol in Codex and the API.
Tencent just open-sourced a Haiku-class model you can run yourself.
Hy4 flash: 220B params, only 12B active, 1M context
Tencent's own blind test (85 users, 259 tasks): 2.79/4 vs Claude Haiku 5.5 at 2.81/4
WorkIndex: 42.1 vs 39.7 for Hy4 preview
About ₹3.7 per task, 1/8 the cost of DeepSeek-V4.1-Flash, 1/13 of Hy4 preview
Cheap agents just went open weight.
Start building with Rapid Router, all the latest models: rapidrouter.ai/
Hy4 flash is now open-weight, including Hy4 flash-FP8 and Hy4 flash-Base.
224B (12B active), 1M context, natively multimodal.
Fast, efficient and built for everyday productivity.
It's yours now. And stay tuned for Hy4 coming up next!
🤗: huggingface.co/tencent/Hy4-f…
🐙: github.com/Tencent-Hunyuan/H…
📄: hy.tencent.ai/research/hy4-f…
Released under the Tencent Hy License.
See full terms here: huggingface.co/tencent/Hy4-f…
1
166
Rapid Router retweeted
In AA-PersonalityBench, Claude Fable 5.2 is more direct than Claude Fable 5.1 and writes notably shorter responses with almost no em dashes
Claude Fable 5.2 ranks higher than Claude Fable 5.1 on Direct and Answer First, but lower on Informativeness and Conviction. Its communication characteristics changed even more than its personality traits, with characteristics moving an average of 38 percentile points.
Microsoft wants your coding budget.
These are Microsoft's own numbers from its Ignite launch, not an independent test.
MAI Code 2, averaged over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1:
- 98% of GPT-6.1 Sol's coding score at 40% of the price
- 96% of Claude Opus 5.5's score at 5% of the price
MAI Code 2 Flash, on DeepSWE 1.1, Terminal Bench 2.1 and SWE-Atlas:
- Matches GPT-6 Luna at 65% of the price
- Matches Claude Haiku 5.5 at 20% of the price
Microsoft also says developers preferred MAI Code 2 over GPT-6 Sol 64% of the time in a blind frontend evaluation.
Microsoft has not published a token price, so there is no ₹ figure here. Microsoft says these models are tuned for everyday enterprise coding, not yet for harder science or very long-horizon agent work.
Until an independent bench lands, treat the quality claims as the lab's own. The question worth testing is whether 98% of Sol at 40% of the price holds up outside Microsoft's tests.
Source: Microsoft AI, Ignite 2026 announcement, 17 Nov 2026.
3
1
427
Microsoft just took the image editing crown. At a discount.
Artificial Analysis has MAI-Image-2.7 at #1 on AA-Image-Editing v2.1, at one sixth the price of the four models behind it. The official AA chart is tight at the top: MAI-Image-2.7 at 1215 Elo, just ahead of GPT Image 2.6 Sunburst (max) at 1211. MAI-Image-2.7-Flash sits at 1179, in 7th.
Price per 1,000 images:
- MAI-Image-2.7: ₹3,256
- MAI-Image-2.7-Flash: ₹958
On AA, 2.7 ranks #1 for Text or Symbol, Scene & Style, and Enhancement & Restoration edits. Flash is #3 for Scene & Style. Both also sit on AA's Quality vs Price frontier for Text to Image.
Arena's ranks for the same model are lower: #3 Multi-Image Edit, #4 Single Image Edit, #4 Text-to-Image. Arena also ranks Microsoft AI as the #2 lab across those three image arenas.
If editing quality is the job, AA says Microsoft is now first, and cheaper than the four behind it. Arena still has it in the top four, not first.
₹ at 93.51/$, 18 Nov. Source: Artificial Analysis and Arena, 17 Nov 2026.
1
135






















