Arena.ai places Claude Opus 5.5 (Max) first in Code Arena WebDev at 1,818 points, 26 ahead of GPT-6 Astra (Max) and 126 above Opus 5 (Max). Claude says the model matches Claude Fable 5.1 on most tasks and costs 40% less to run than Opus 5.
Tencent Hy says learning-rate retuning can preserve learning per response across GRPO and PPO over a bounded range of batch sizes. On fixed hardware, larger batches improve PPO generation-stage throughput by up to 2.29×, and its best measured GRPO configuration reaches the same validation target in 29% less time.
Command Code is offering Space Bunny Alpha, which it calls a stealth model, at no charge on all plans. The model has 1M context and is multimodal.
Waabi said the system had not previously practiced the 300-mile drive in simulation and did not rely on new real-world data, calling the run a first.
ClaudeDevs says officially available Claude Code cloud sessions are billed on Pro or Max plans like the rest of Claude Code. Existing subscribers get an optional one-time credit of $100 on Pro or $250 on Max, which those sessions spend before normal plan usage.
In Suno Studio, arming a track and pressing record captures a performance straight onto the timeline.
Artificial Analysis says Claude Opus 5.5 uses 15.6 million tokens per Coding Agent Index task, compared with 11.4 million for Opus 5. Cached input rises from 10.9 million to 14.6 million, and output more than doubles from 137,000 to 333,000.
At max effort in Claude Code, Artificial Analysis measures Opus 5.5 at 66, above Opus 5 at 60 and Claude Fable 5.1 at 62. Its cost per task is $13.04, higher than Opus 5’s $10.79 despite lower token prices.
Salesforce AI Research put that collection at 35.8% clean and two others at 10.1% and 3.3%, after finding reward errors in both directions. The authors say RIVER, which filters defective environments, increases RL gains by 106% on Terminal-Bench-Lite and 30% on Terminal-Bench v2.1 while using fewer than 30% of TMax's environments.
Muse will talk back during the video chat, and users can prompt exactly how it sounds. Wang describes the feature only as coming soon, with no release date.
Polymarket describes Muse Charm as a Tamagotchi-style AI device designed to fit on a keychain, and says Zuck called it a “joyful little device.”
dax says that user count is now about 50% of TUI users.
On the Artificial Analysis Text to Image Leaderboard, the 6B Ming-Image-0.1-Design model places 17th of 81 in UI/UX design and first among open-weight models, compared with 45th of 160 overall. It is released under the MIT license with RGBA output and was evaluated at 2K resolution using 12 inference steps.
Shopify says Shop Pay can now be used as a wallet in Muse for agentic checkout.
ClaudeDevs has added local support to Projects in Claude Code, so threads can run on a user’s own machine. Projects were rolled out on desktop and web as conversations whose threads run in parallel cloud sessions.
ClaudeDevs is still processing the Projects early-access waitlist, and migration is not finished for people who previously used Claude projects.
Tenure-track faculty at U.S. universities can submit Fall 2026 proposals on the scientific foundations of AI evaluation. Each project may receive up to $50,000, and the deadline is October 30, 2026.
Jiantao Jiao says an NVIDIA team built SWE-Serve to test whether AI agents can develop an inference engine and serve real models. The benchmark has 53 tasks drawn from real sgl_project engineering work.
Wispr Flow users can now write Hinglish in English letters, in the same form they would use when typing.
The Tev1-4B-experimental weights are in the togethercomputer repository on Hugging Face. Together AI describes the model as an experiment meant to show how specialized decision models can be trained fairly easily.
