Arena.ai(@lmarena_ai)

In the Frontend Code Arena, Hy3 is #2 for open-weight models! -#16 overall - Top 20 in Reference-Ba...

Arena.ai (@arena) — Tue, 21 Jul 2026 22:40:23 GMT

In the Frontend Code Arena, Hy3 is #2 for open-weight models!

-#16 overall
- Top 20 in Reference-Based Design, Gaming, Simulations and Content Creation Tools

💬3🔄1❤️27👀6235📊7

Hy3 by Tencent is #5 in Agent Arena for open-weight models (#25 overall)! It also ranks as the #2 op...

Arena.ai (@arena) — Tue, 21 Jul 2026 22:40:21 GMT

Hy3 by Tencent is #5 in Agent Arena for open-weight models (#25 overall)! It also ranks as the #2 open model in the Frontend Code Arena (#16 overall)!

In Agent Arena: Hy3 lands at #25 overall (net -2.2%). Hy3 has strengths in tool-use (recovering well from CLI/bash errors, +2.6% and #25), its biggest weakness is steerability as it struggles to course-correct when users push back, coming in at -7.1% (#30).

Agent Arena measures models on millions of real-world, long-horizon agentic tasks. Models get web search, filesystem, and terminal tools to complete complex workflows: writing code, creating slide decks, researching the web, building apps, and analyzing documents.

We use causal tracing methodology to measure a model's net improvement, which indicates how much it improves outcomes relative to the average model.

Below we break down how Hy3 scored across 5 key signals, drawn from tasks submitted by a global community of users. Here’s an overview on the signals:

User-satisfaction proxies
- Confirmed Success: an explicit "yes that worked" feedback from the user
- Praise vs. Complaint: implicit sentiment in users reactions
- Steerability: can the model course-correct when you push back?

Tool-use proxies
- Bash Recovery: how it recovers from CLI errors (primary signal for tool use)
- Tool Hallucination: does it call tools that don't exist

Congrats to the @TencentHunyuan team on this release!

💬15🔄32❤️315👀65413📊53

Dive into all the leaderboard details at: https://t.co/PjWOaDEXWR

Arena.ai (@arena) — Tue, 21 Jul 2026 15:48:07 GMT

Dive into all the leaderboard details at: arena.ai/leaderboard

💬0🔄0❤️9👀11343📊1

Gemini 3.6 Flash has landed in the Text Arena Top 20 at #12 with 1485 pts. In categories it ranks: ...

Arena.ai (@arena) — Tue, 21 Jul 2026 15:48:06 GMT

Gemini 3.6 Flash has landed in the Text Arena Top 20 at #12 with 1485 pts.

In categories it ranks:
- #10 Instruction Following
- Within Occupational: #4 in mathematics, #8 in Legal & Government, and #9 in Medical & Healthcare

Gemini 3.5 Flash-Lite is also available in the Text Arena, with a 26-point improvement over Gemini 3.1 Flash-Lite (#81->#42).

💬3🔄3❤️83👀14696📊12

Gemini 3.6 Flash has landed #12 with 1537 pts in the Frontend Code Arena. This release is a signifi...

Arena.ai (@arena) — Tue, 21 Jul 2026 15:48:05 GMT

Gemini 3.6 Flash has landed #12 with 1537 pts in the Frontend Code Arena.

This release is a significant improvement from Gemini 3.5 Flash (#21->#12).

Across domains it ranks:
- #8 Reference-Based Design
- #9 Content Creation Tools
- #10 Brand & Marketing
- #12 Simulations
- #13 Gaming
- In the top 20 for Data & Analytics and Consumer Product

Click into the thread to see the results for Gemini 3.6 Flash and Gemini 3.5 Flash-Lite in the Text Arena.

Congrats to the @GoogleDeepMind team on this launch!

Google DeepMind@GoogleDeepMind

We’re rolling out three new models to make AI agents faster, smarter, and cheaper at scale:

🔵 Gemini 3.6 Flash: It uses fewer tokens than 3.5 Flash to deliver higher quality work at the exact same cost.

🔵 Gemini 3.5 Flash-Lite: A fast, cost-effective option for everyday tasks like processing documents and agentic search.

🔵 Gemini 3.5 Flash Cyber: A cybersecurity model built to find and patch critical software vulnerabilities.

💬132🔄222❤️2010👀351518📊325

Dive into the Inkling scores in the Agent Arena at: https://t.co/5PhJhhhUYI

Arena.ai (@arena) — Mon, 20 Jul 2026 19:00:26 GMT

Dive into the Inkling scores in the Agent Arena at: arena.ai/leaderboard/ag…

💬0🔄1❤️7👀6016📊1

Inkling is the #9 open model, ranking #30 overall in the Agent Arena (+9.6%) - #18 Bash Recovery (+6...

Arena.ai (@arena) — Mon, 20 Jul 2026 19:00:25 GMT

Inkling is the #9 open model, ranking #30 overall in the Agent Arena (+9.6%)
- #18 Bash Recovery (+6.4%)
- #29 Confirmed Task Success (-5.6%)
- #35 Tool Hallucination (-0.9%) 
- #35 Steerability (-10.4%)
- #38 Praise vs. Complaint (-18.0%)

💬3🔄2❤️13👀6942📊5

Inkling by @thinkymachines debuts at #9 in Agent Arena among open-weight models, and #30 overall. Th...

Arena.ai (@arena) — Mon, 20 Jul 2026 19:00:24 GMT

Inkling by @thinkymachines debuts at #9 in Agent Arena among open-weight models, and #30 overall. This makes the best open model from the US for agentic long running tasks and also the only US model in the top 10 open model field. If Kimi K3's weights are released on schedule by July 27, it will become the #1 open-weight model shifting all others down by 1 rank.

Inkling's weak spot is user sentiment: people complain more than they praise it (-18.0%, #38), and it doesn't take feedback well, ranking #35 on steerability and #29 on task completion. Where it does shine is recovering from CLI errors (+6.4%, #16).

In Agent Arena, we measure models on millions of real-world, long-horizon agentic tasks. Models get web search, filesystem, and terminal tools to complete complex workflows: writing code, creating slide decks, researching the web, building apps, and analyzing documents.

Congrats to the @thinkymachines team on this strong debut!

Arena.ai@arena

Inkling by @thinkymachines debuts at #10 (1434 pts) in Frontend Code Arena for open weight models, and #37 overall. This makes it not only the best open model from the US, but also the only US model in the top 15 open model field.

Within open weight models, Inking is strong with #4 in Brand & Marketing, #7 in Data & Analytics, and within the top 20 across all remaining domains.

💬40🔄40❤️582👀102305📊84

Learn more about the causal tracing methodology for Agent Arena on our blog: https://t.co/bpIkMhEeKL

Arena.ai (@arena) — Mon, 20 Jul 2026 17:16:17 GMT

Learn more about the causal tracing methodology for Agent Arena on our blog: arena.ai/blog/agent-are…

💬0🔄1❤️6👀7451📊1

See the full Agent Arena leaderboard at https://t.co/sE9q4FSYAt

Arena.ai (@arena) — Mon, 20 Jul 2026 17:12:52 GMT

See the full Agent Arena leaderboard at arena.ai/leaderboard/ag…

💬1🔄2❤️12👀8687📊3

Exciting update: Kimi K3 has landed at #4 on the Agent Arena leaderboard, matching Claude Opus 4.8 a...

Arena.ai (@arena) — Mon, 20 Jul 2026 17:12:50 GMT

Exciting update: Kimi K3 has landed at #4 on the Agent Arena leaderboard, matching Claude Opus 4.8 and GPT-5.6 Sol. If Kimi K3's weights are released on schedule by July 27, it will become the #1 open-weight model.

This release marks a major leap in agentic performance over Kimi K2.7 Code (#23 to #4). Based on 8K+ live agentic sessions, Kimi K3 leads on confirmed task success rate (#1). It also posts a strong +20.6% on praise vs. complaint (#3). It currently lags the field in steerability (#14) and bash recovery (#17).

Agent Arena measures models on millions of real-world, long-horizon agentic tasks. Models get web search, filesystem, and terminal tools to complete complex workflows: writing code, creating slide decks, researching the web, building apps, and analyzing documents.

We use causal tracing methodology to measure a model's net improvement, which indicates how much it improves outcomes relative to the average model.

Here's a primer on the 5 signals:

User-satisfaction proxies
- Confirmed Success: an explicit "yes that worked" feedback from the user
- Praise vs. Complaint: implicit sentiment in users reactions
- Steerability: can the model course-correct when you push back?

Tool-use proxies
- Bash Recovery: how it recovers from CLI errors (primary signal for tool use)
- Tool Hallucination: does it call tools that don't exist

Below we break down how Kimi K3 scored across the 5 signals, drawn from tasks submitted by a global community of users.

Congrats @Kimi_Moonshot on another big milestone!

Arena.ai@arena

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.

This is a 17-place jump from Kimi-k2.6 (#18 -> #1).

In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.

The full model weights will be released by July 27.

Congrats to the @Kimi_Moonshot team on this major milestone!

💬93🔄264❤️2165👀299244📊307

Watch the speedrun of a second example using a real engineering task: https://t.co/IJpOwvCbbm

Arena.ai (@arena) — Mon, 20 Jul 2026 15:01:58 GMT

Watch the speedrun of a second example using a real engineering task: youtu.be/niAql9hW664

💬0🔄1❤️6👀6955📊2

STEP-CHANGE ECONOMICS: Six months ago, he spent $8,000 and 2 weeks of a senior engineer's time. Now ...

Arena.ai (@arena) — Mon, 20 Jul 2026 15:01:57 GMT

STEP-CHANGE ECONOMICS: Six months ago, he spent $8,000 and 2 weeks of a senior engineer's time. Now the same job takes $20 and 45 minutes.

Your browser does not support the video tag.

💬8🔄8❤️85👀20776📊20

Let us know what you think, and visit https://t.co/DYnDhxb9y7 to generate and share your own.

Arena.ai (@arena) — Fri, 17 Jul 2026 22:08:18 GMT

Let us know what you think, and visit arena.ai/code to generate and share your own.

💬2🔄2❤️13👀9521📊4

0:00 Ramen Service Night Website - Kimi K3: https://t.co/R2E24neog9 - Fable 5: https://t.co/9yDYbb0H...

Arena.ai (@arena) — Fri, 17 Jul 2026 22:08:18 GMT

0:00 Ramen Service Night Website
- Kimi K3: …-b853-81a7741caf77.staging.arena.site
- Fable 5: …-a869-a94b040a5112.staging.arena.site

0:06 Deep-Sea Research Station Dashboard
- Kimi K3: …-bde7-55e3188be393.staging.arena.site
- Fable 5: …-a25b-7c20e4310ca8.staging.arena.site

0:12 Theater Stage Rehearsal Site
- Kimi K3: …-9c48-11eeb74638ca.staging.arena.site
- Fable 5: …-a35d-6e783194075b.staging.arena.site

0:17 Record Pressing Plant Website
- Kimi K3: …-9d65-df2101eab307.staging.arena.site
- Fable 5: …-9379-219c4b5618d4.staging.arena.site

0:22 Haunted Hotel Booking System
- Kimi K3: …-a1fe-2e514902b111.staging.arena.site
- Fable 5: …-aed3-7a81d1ec167e.staging.arena.site

0:28 Children’s Science Physics Table
- Kimi K3: …-abc4-011c296a9ae9.staging.arena.site
- Fable 5: …-ada5-5fc75509dbd3.staging.arena.site

0:34 The Shape Of An American Day
- Kimi K3: …-b669-85ee85f7b04b.staging.arena.site
- Fable 5: …-b252-c817eeed7e0c.staging.arena.site

0:40 A Thousand Years Of Cherry Blossoms
- Kimi K3: …-92d4-a1ae3e44480b.staging.arena.site
- Fable 5: …-91a7-a2b19d4b1996.staging.arena.site

0:45 Print Shop Poster Press
- Kimi K3: …-81f8-183d9a7971c1.staging.arena.site
- Fable 5: …-af02-f8ba19e773cf.staging.arena.site

0:50 Billion Dollar Disasters Become More Frequent 
- Kimi K3: …-b1d1-8c09c862857d.staging.arena.site
- Fable 5: …-bf29-689ec1273fd3.staging.arena.site

0:55 The New Rush To Orbit
- Kimi K3: …-baf8-25aaf1ab4229.staging.arena.site
- Fable 5: …-8e0f-1cc29be3a82f.staging.arena.site

1:00 More Than A Century Of Nobel Prizes
- Kimi K3: …-a6b6-8452df9a9c90.staging.arena.site
- Fable 5: …-acfd-7b2b8d06bbed.staging.arena.site

1:09 How MoMA Built Its Collection
- Kimi K3: …-af0c-d1efd5ae5311.staging.arena.site
- Fable 5: …-a7fe-816c083466a6.staging.arena.site

1:12 The Seating Chart Solver
- Kimi K3: …-b9e8-d602c3567441.staging.arena.site
- Fable 5: …-b32a-811ec7cc4b36.staging.arena.site

1:18 City Parade Route Planner
- Kimi K3: …-88d3-d7c7d223d151.staging.arena.site
- Fable 5: …-87d9-917c1b79942b.staging.arena.site

1:23 Orbit Travel Booking Console
- Kimi K3: …-9271-3a2aaf07aa30.staging.arena.site
- Fable 5: …-b3a7-6a299ffe33d9.staging.arena.site

💬1🔄4❤️34👀9438📊8

Big news from @Kimi_Moonshot this week. Check out Kimi K3 head-to-head with Fable 5 on identical pro...

Arena.ai (@arena) — Fri, 17 Jul 2026 22:08:17 GMT

Big news from @Kimi_Moonshot this week. Check out Kimi K3 head-to-head with Fable 5 on identical prompts in Frontend Code Arena.

Your browser does not support the video tag.

Arena.ai@arena

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.

This is a 17-place jump from Kimi-k2.6 (#18 -> #1).

In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.

The full model weights will be released by July 27.

Congrats to the @Kimi_Moonshot team on this major milestone!

💬29🔄110❤️1045👀118959📊148

See the full Frontend Code Arena leaderboards at https://t.co/GFZ3FCC7Cl

Arena.ai (@arena) — Fri, 17 Jul 2026 20:01:44 GMT

See the full Frontend Code Arena leaderboards at arena.ai/leaderboard/co…

💬0🔄2❤️23👀13594📊4

For the first time, China has taken the lead over the US in Frontend Code Arena with the launch of K...

Arena.ai (@arena) — Fri, 17 Jul 2026 20:01:43 GMT

For the first time, China has taken the lead over the US in Frontend Code Arena with the launch of Kimi-K3 by @Kimi_Moonshot.

The last time a Chinese model came close was in early 2025, with DeepSeek-R1.

💬98🔄513❤️3099👀467301📊566

Subscribe to the Arena YouTube channel for more researcher deep dives: https://t.co/cukQBwkjua

Arena.ai (@arena) — Fri, 17 Jul 2026 15:41:56 GMT

Subscribe to the Arena YouTube channel for more researcher deep dives: youtube.com/@ArenaAIOffici…

💬0🔄0❤️6👀5838📊1

Factuality and human preference are complementary signals. 00:00 Why we're adding factuality as a s...

Arena.ai (@arena) — Fri, 17 Jul 2026 15:40:32 GMT

Factuality and human preference are complementary signals.

00 Why we're adding factuality as a signal
14 Human preference vs. factuality: complementary, not the same
06 How this differs from style control
32 The math: reviewing the existing Bradley-Terry objective
27 Extending the objective to factuality
23 The composite objective: weighting human preference + factuality
14 Why rankings shift between human preference and factuality
34 Defining the factuality label
48 Why models aren't penalized for abstaining from claims
44 Confidence intervals via M-estimation and asymptotic normality
27 Wrap-up: what's next for factuality at Arena

Your browser does not support the video tag.

Arena.ai@arena

Introducing factuality in the Arena: a new ranking of models according to a weighted combination of human preference and factuality.

Model rankings are now viewable according to a weighted combination of human preference and factuality. Factuality is live in our Text and Search Arenas as a non-default toggle.

We audit model responses by randomly sampling battles and extracting web-verifiable claims. We then verify these claims and compare the average correctness between model responses.

To power these rankings, we’ve labeled over 2 million claims made by LLMs in real-world conversations, 1.3+ million from Text Arena, and 700k+ from Search Arena.

Notable highlights with factuality enabled in the Text Arena:
- Claude Fable 5 moves down slightly to spot #2
- GPT-5.5 saw the largest increase, moving up 13 spots into the #7 spot
- Muse Spark dropped the most from #7 to #20 (-13pt)

By labs, Meta saw the largest drop from #2 to #5, while Anthropic overall held the #1 spot. Looking at only open model providers, Xiaomi saw the largest improvement, jumping from #9 to #6.

Learn more about the findings and methodology in this thread.

💬10🔄14❤️121👀33225📊22

Kimi-K3 just topped the Frontend Code Arena with a 76% pairwise win rate. When its output was comp...

Arena.ai (@arena) — Thu, 16 Jul 2026 23:11:16 GMT

Kimi-K3 just topped the Frontend Code Arena with a 76% pairwise win rate. 

When its output was compared head-to-head against other models on the same task, it was picked as the better output 76% of the time on average.

For reference: Claude Fable 5 (63%), GPT-5.6 Sol (58%). 50% is baseline, a model winning and losing equally often.

Arena.ai@arena

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.

This is a 17-place jump from Kimi-k2.6 (#18 -> #1).

In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.

The full model weights will be released by July 27.

Congrats to the @Kimi_Moonshot team on this major milestone!

💬103🔄346❤️2841👀380354📊405

See the full Text Arena leaderboards at https://t.co/q0J9NAytLa

Arena.ai (@arena) — Thu, 16 Jul 2026 20:41:41 GMT

See the full Text Arena leaderboards at arena.ai/leaderboard/te…

💬1🔄1❤️14👀11778📊3

In the Text Arena, Kimi-K3 by @Kimi_Moonshot landed #9, with 1486 pts. This is another significant ...

Arena.ai (@arena) — Thu, 16 Jul 2026 20:41:40 GMT

In the Text Arena, Kimi-K3 by @Kimi_Moonshot landed #9, with 1486 pts.

This is another significant improvement from Kimi-k2.6 (#38 -> #9).

- Top 10 in Creative Writing, Coding and Instruction Following
- #1 in three occupations: Physical & Social Science, Legal & Government, and Medicine & Healthcare. Plus #3 Business, Management, & Financial Ops

The full model weights will be released by July 27.

💬16🔄41❤️643👀88902📊94

Check out the Pareto Frontier for Code Arena: Frontend https://t.co/I584K4VT5z

Arena.ai (@arena) — Thu, 16 Jul 2026 19:41:36 GMT

Check out the Pareto Frontier for Code Arena: Frontend arena.ai/leaderboard/co…

💬86🔄1❤️13👀10767📊71

Kimi K3 from @Kimi_Moonshot has moved the Pareto Frontier for Code Arena: Frontend.

Arena.ai (@arena) — Thu, 16 Jul 2026 19:41:35 GMT

Kimi K3 from @Kimi_Moonshot has moved the Pareto Frontier for Code Arena: Frontend.

Arena.ai@arena

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.

This is a 17-place jump from Kimi-k2.6 (#18 -> #1).

In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.

The full model weights will be released by July 27.

Congrats to the @Kimi_Moonshot team on this major milestone!

💬17🔄57❤️816👀90507📊87

Check out @Kimi_Moonshot's Kimi-K3 official release: https://t.co/AnJUchkGxf

Arena.ai (@arena) — Thu, 16 Jul 2026 19:16:14 GMT

Check out @Kimi_Moonshot's Kimi-K3 official release: x.com/Kimi_Moonshot/…

Kimi.ai@Kimi_Moonshot

Introducing Kimi K3: Open Frontier Intelligence

🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows

Kimi K3 is now live onKimi.comxZUo, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.

🔗 platform.kimi.aiXAqMw
🔗 Tech kimi.com/blog/kimi-k3MSNM1f

💬4🔄11❤️252👀316744📊27

See the full Frontend Code Arena leaderboards at https://t.co/GFZ3FCC7Cl

Arena.ai (@arena) — Thu, 16 Jul 2026 18:33:48 GMT

See the full Frontend Code Arena leaderboards at arena.ai/leaderboard/co…

💬4🔄16❤️242👀243173📊35

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing C...

Arena.ai (@arena) — Thu, 16 Jul 2026 18:33:47 GMT

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.

This is a 17-place jump from Kimi-k2.6 (#18 -> #1).

In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.

The full model weights will be released by July 27.

Congrats to the @Kimi_Moonshot team on this major milestone!

Kimi.ai@Kimi_Moonshot

Meet Kimi K3

Your browser does not support the video tag.

💬1075🔄3956❤️26362👀19154119📊4962

Dive into the Inkling scores across both leaderboards at: https://t.co/PjWOaDEXWR

Arena.ai (@arena) — Thu, 16 Jul 2026 17:19:41 GMT

Dive into the Inkling scores across both leaderboards at: arena.ai/leaderboard

💬0🔄1❤️6👀5393📊1

In the Text Arena, Inkling also landed at #10 for open source (1447 pts), and #58 overall. It is the...

Arena.ai (@arena) — Thu, 16 Jul 2026 17:19:40 GMT

In the Text Arena, Inkling also landed at #10 for open source (1447 pts), and #58 overall. It is the #2 open source model from the US, second to GoogleDeepMind’s Gemma-4-31B. 

Highlights across key categories:
- #10 Hard prompts
- #12 Longer Query
- #14 Coding
- #15 Instruction Following
- Top 15 across occupations: Writing, Literature & Language (#9), Entertainment, Sports & Media (#10), Software & IT Services (#12), Life, Physical, & Social Science (#14) and Business, Management & Financial Ops (#14).

💬3🔄0❤️24👀6561📊6

Inkling by @thinkymachines debuts at #10 (1434 pts) in Frontend Code Arena for open weight models, a...

Arena.ai (@arena) — Thu, 16 Jul 2026 17:19:39 GMT

Inkling by @thinkymachines debuts at #10 (1434 pts) in Frontend Code Arena for open weight models, and #37 overall. This makes it not only the best open model from the US, but also the only US model in the top 15 open model field.

Within open weight models, Inking is strong with #4 in Brand & Marketing, #7 in Data & Analytics, and within the top 20 across all remaining domains.

💬11🔄19❤️234👀136281📊39

Kimi K3 by @Kimi_Moonshot is available in the Agent Arena! Agent Arena evaluates long-running agent...

Arena.ai (@arena) — Thu, 16 Jul 2026 17:06:18 GMT

Kimi K3 by @Kimi_Moonshot is available in the Agent Arena!

Agent Arena evaluates long-running agents. It measures models on millions of real-world, long-horizon agentic tasks from a global community of users. Models can access web search, filesystem, and terminal tools to complete complex workflows. The leaderboard measures model performance on outcomes relative to the average model using a causal tracing methodology.

In addition to Agent Arena, Kimi K3 is also available for: Text, Vision, Document and Frontend Code Arena.

Head over and check it out, your votes drive the leaderboards. Scores incoming!

💬16🔄27❤️437👀29514📊61

Test out Kimi K3 in Battle Mode and Agent Mode at: https://t.co/yZiJuG8ica. Scores incoming.

Arena.ai (@arena) — Thu, 16 Jul 2026 17:06:18 GMT

Test out Kimi K3 in Battle Mode and Agent Mode at: arena.ai. Scores incoming.

💬0🔄1❤️31👀5830📊3

Test out Inkling in Battle Mode and Agent Mode at: https://t.co/yZiJuG8ica

Arena.ai (@arena) — Wed, 15 Jul 2026 19:33:08 GMT

Test out Inkling in Battle Mode and Agent Mode at: arena.ai

💬0🔄0❤️7👀5041📊1

Inkling by @thinkymachines is now in the Agent Arena! Agent Arena evaluates long-running agents. It...

Arena.ai (@arena) — Wed, 15 Jul 2026 19:33:07 GMT

Inkling by @thinkymachines is now in the Agent Arena!

Agent Arena evaluates long-running agents. It measures models on millions of real-world, long-horizon agentic tasks from a global community of users. Models can access web search, filesystem, and terminal tools to complete complex workflows. The leaderboard measures model performance on outcomes relative to the average model using a causal tracing methodology.

In addition to Agent Arena, Inkling is also available for: Text, Vision, and Code Arena.

Head over and check it out, your votes drive the leaderboards. Scores incoming!

Thinking Machines@thinkymachines

Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.

thinkingmachines.ai/news/introduci…

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵

💬5🔄7❤️159👀19862📊17

👀

Arena.ai (@arena) — Wed, 15 Jul 2026 18:36:42 GMT

👀

Haleemah@Haleeeemahh

Another day Another AI release   

a chinese openweight model  kimi k3 is already live on @arena under the name 'KIVINE' 

official release coming soon but here's some of the best results from AI creators testing it   

i am yet to test it, if you know how i can share tips in the CS   

let me know who's test you live best :)  

 - jazii's race game

Your browser does not support the video tag.

💬6🔄10❤️173👀15896📊20

Factuality signals - a new way to measure whether models are making true verifiable claims. The Are...

Arena.ai (@arena) — Wed, 15 Jul 2026 17:33:04 GMT

Factuality signals - a new way to measure whether models are making true verifiable claims.

The Arena team breaks down how factuality is measured, why it matters for real-world reliability, and what it reveals about how frontier models compare.

Your browser does not support the video tag.

💬2🔄3❤️57👀12360📊8

Watch the full walkthrough: https://t.co/85BseOncSY

Arena.ai (@arena) — Wed, 15 Jul 2026 17:33:04 GMT

Watch the full walkthrough: youtube.com/watch?v=YDvV9x…

💬0🔄0❤️9👀6962📊1

Some highlights in Search Arena with the Factuality leaderboard today include: - GPT-5.5-search mov...

Arena.ai (@arena) — Wed, 15 Jul 2026 16:37:13 GMT

Some highlights in Search Arena with the Factuality leaderboard today include:

- GPT-5.5-search moved up by one to take the top Factuality leaderboard spot
- Larger improvement in ranking for GPT-5.2-search (jumping from #11 to #3)
- Drops in model rankings includes claude-sonnet-4-6-search (falling from #6 to #9) and gemini-3.1-pro-grounding (falling from #7 to #13)

💬2🔄1❤️25👀4140📊5

For a deeper look at our factuality findings and methodology, visit our blog, where you can interact...

Arena.ai (@arena) — Wed, 15 Jul 2026 16:37:13 GMT

For a deeper look at our factuality findings and methodology, visit our blog, where you can interact with the weighted charts and see how scores and rankings shift as the weighting increases: arena.ai/blog/factualit…

💬0🔄0❤️12👀3588📊2

Introducing factuality in the Arena: a new ranking of models according to a weighted combination of ...

Arena.ai (@arena) — Wed, 15 Jul 2026 16:37:10 GMT

Introducing factuality in the Arena: a new ranking of models according to a weighted combination of human preference and factuality.

Model rankings are now viewable according to a weighted combination of human preference and factuality. Factuality is live in our Text and Search Arenas as a non-default toggle.

We audit model responses by randomly sampling battles and extracting web-verifiable claims. We then verify these claims and compare the average correctness between model responses.

To power these rankings, we’ve labeled over 2 million claims made by LLMs in real-world conversations, 1.3+ million from Text Arena, and 700k+ from Search Arena.

Notable highlights with factuality enabled in the Text Arena:
- Claude Fable 5 moves down slightly to spot #2
- GPT-5.5 saw the largest increase, moving up 13 spots into the #7 spot
- Muse Spark dropped the most from #7 to #20 (-13pt)

By labs, Meta saw the largest drop from #2 to #5, while Anthropic overall held the #1 spot. Looking at only open model providers, Xiaomi saw the largest improvement, jumping from #9 to #6.

Learn more about the findings and methodology in this thread.

💬12🔄45❤️404👀133813📊72

Meta’s Muse Spark 1.1 ranks at #17, above Gemini 3.1 Pro and Qwen-3.7 Plus, but lower than Grok 4.5 ...

Arena.ai (@arena) — Tue, 14 Jul 2026 20:50:06 GMT

Meta’s Muse Spark 1.1 ranks at #17, above Gemini 3.1 Pro and Qwen-3.7 Plus, but lower than Grok 4.5 or GLM 5.2.

💬2🔄3❤️24👀5324📊5

Check out the Agent Arena leaderboard to see the details: https://t.co/qNgm4b5vwT In Agent Arena, w...

Arena.ai (@arena) — Tue, 14 Jul 2026 20:50:06 GMT

Check out the Agent Arena leaderboard to see the details: arena.ai/leaderboard/ag…

In Agent Arena, we measure models on millions of real-world, long-horizon agentic tasks from a global community of users. Models can access web search, filesystem, and terminal tools to complete complex workflows. The leaderboard measures model performance on outcomes relative to the average model using a causal tracing methodology.

💬1🔄0❤️10👀4324📊2

Muse Spark 1.1 by @AIatMeta lands in the Agent Arena, marking the introduction of Meta to the leader...

Arena.ai (@arena) — Tue, 14 Jul 2026 20:50:05 GMT

Muse Spark 1.1 by @AIatMeta lands in the Agent Arena, marking the introduction of Meta to the leaderboard at #5 across all labs (and #17 on the model leaderboard). 

Congrats to @AIatMeta on its initial entry into Agent Arena!

AI at Meta@AIatMeta

We’re excited to introduce Muse Spark 1.1, a significant upgrade from the first Muse Spark model we released earlier this year.

Along with this release, we are launching a public preview of the new Meta Model API where developers can access Muse Spark 1.1.

The model is also available now in "Thinking" mode in the Meta AI app and on meta.ai.

Learn more: go.meta.me/ff8e2c

💬12🔄21❤️348👀37604📊39

Curious how different models actually perform before you build your next agent pipeline? Check out t...

Arena.ai (@arena) — Tue, 14 Jul 2026 15:44:39 GMT

Curious how different models actually perform before you build your next agent pipeline? Check out the Arena leaderboard: arena.ai/leaderboard

💬0🔄0❤️9👀4879📊1

Right before joining Arena, @melissapan (PhD candidate, UC Berkeley) presented research on cutting a...

Arena.ai (@arena) — Tue, 14 Jul 2026 15:44:38 GMT

Right before joining Arena, @melissapan (PhD candidate, UC Berkeley) presented research on cutting agent system costs by 89% — while matching 100% of the best static config's accuracy.

Picking the right LLM matters less than you think. Full system config >> LLM routing alone.

00 Old way: LLM routing
39 Q&A agent options
05 Cost vs accuracy tradeoffs
44 Insight #1: full config > LLM routing
12 Matei vs Melissa example
27 Introducing BRANE
00 Benchmarks covered
00 Results: 89% cost cut

Your browser does not support the video tag.

💬8🔄18❤️256👀45897📊65

In Agent Arena, Grok-4.5 ranks #13 overall (+5.0%). - #4 Hallucination (+1.3%) - #6 Confirmed Task S...

Arena.ai (@arena) — Mon, 13 Jul 2026 18:00:36 GMT

In Agent Arena, Grok-4.5 ranks #13 overall (+5.0%).
- #4 Hallucination (+1.3%)
- #6 Confirmed Task Success (+6.9%)
- #9 Bash Recovery (+9.7%)
- #12 Praise vs. Complaint (+6.1%)
- #15 Steerability (+0.9%)

💬1🔄1❤️19👀5062📊3

Check out the Agent Arena leaderboard to dive into the details: https://t.co/qNgm4b5vwT In Agent Ar...

Arena.ai (@arena) — Mon, 13 Jul 2026 18:00:36 GMT

Check out the Agent Arena leaderboard to dive into the details: arena.ai/leaderboard/ag…

In Agent Arena, we measure models on millions of real-world, long-horizon agentic tasks from a global community of users. Models can access web search, filesystem, and terminal tools to complete complex workflows. The leaderboard measures model performance on outcomes relative to the average model using a causal tracing methodology.

💬0🔄0❤️8👀4243📊1

Grok-4.5 from @SpaceXAI debuts at #13 on the new Agent Arena leaderboard, based on 9.8K live agentic...

Arena.ai (@arena) — Mon, 13 Jul 2026 18:00:35 GMT

Grok-4.5 from @SpaceXAI debuts at #13 on the new Agent Arena leaderboard, based on 9.8K live agentic sessions.

Compared to Grok 4.3, this release is a significant step forward in agentic performance (#29->#13). It makes major gains on Bash Recovery, catching up to Anthropic and OpenAI models, and shows a substantial increase in confirmed task success, making it far more effective in real-world use. Check out rankings by signal below.

Congrats to the SpaceXAI team on the strong Grok-4.5 release!

SpaceXAI@SpaceXAI

Announcing Grok 4.5, our first model trained specifically for coding and agents. It was trained with Cursor and offers frontier intelligence at leading speeds and cost efficiency.

x.ai/news/grok-4-5

Your browser does not support the video tag.