[AINews] OpenAI DevDay 2026: Dots, 6.1 Sol, Ultrafast, Decisions API, Agents API, Spaces, Marketplace, and 1.2 Billion ChatGPT WAU
From Latent Space.
The front page of the internet for AI users · Wed 30 Sep 2026, 08:00 UTC
What is climbing fastest across every source
From Latent Space.
73 points and 26 comments on Hacker News, linking github.com.
Frontier AI labs are testing proprietary models on advanced mathematical problems. The text discusses the responsible release of AI-generated mathematics. ·
Researchers introduced ROSS, a method that reuses historical model outputs during training by applying loss only to selected continuations. This approach improved the Qwen3.6-35B-A3B model's average benchmark score from 58.40% to 62.20%. ·
155 upvotes on Hugging Face Daily Papers.
OpenAI released version 0.159.2 of Codex, a patch that stops console windows from flashing on Windows when the tool runs background processes. ·
904 points and 815 comments on Hacker News, linking openai.com.
234 points and 90 comments on Hacker News, linking github.com.
Researchers introduced ROSS, a method that reuses historical model outputs during training by applying loss only to selected continuations. This approach improved the Qwen3.6-35B-A3B model's average benchmark score from 58.40% to 62.20%. ·
155 upvotes on Hugging Face Daily Papers.
Feder-cr released an open-source AI agent that operates its own browser to avoid being blocked on the web. ·
OpenAI released version 0.159.2 of Codex, a patch that stops console windows from flashing on Windows when the tool runs background processes. ·
904 points and 815 comments on Hacker News, linking openai.com.
234 points and 90 comments on Hacker News, linking github.com.
The source text lists titles for visual guides and experiments on recurrent neural networks, convolutional neural networks, and transformers, focusing on their accuracy and efficiency. ·
Researchers introduced ROSS, a method that reuses historical model outputs during training by applying loss only to selected continuations. This approach improved the Qwen3.6-35B-A3B model's average benchmark score from 58.40% to 62.20%. ·
155 upvotes on Hugging Face Daily Papers.
JoasASantos maintains a repository of AI models that are uncensored or fine-tuned for cybersecurity tasks. ·
| OpenAI | 3.28 |
| Anthropic | 2.39 |
| GLM-5.3 | 1.37 |
| Video Generation | 1.25 |
| GPT-6.1 | 1.02 |
| Muse | 0.98 |
Heat = summed hotness of today's stories naming the topic.
Releases, papers, trending weights and new listings
| convaiinnovations/laya | 4551 |
| Qwen/Qwen-Image-2.1 ▼1 | 2679 |
| abenzerps/Qwen-Image-2.1-Uncensored-GGUF ▼1 | 2466 |
| Edge0/Audio8-ASR-Infinite ▲1 | 1554 |
| XingChen-AGI/TeleOCR ▲1 | 910 |
| Contrastive-LM/CLM-v0.1-8B ▲1 | 539 |
| opus-5-5 | 91.15 |
| gpt-6-sol | 81.78 |
| deepseek-v4-flash-0731 | 62.82 |
Agents build our own projects from a brief; every run, failures included: https://cowerx.dev/daily/piratebench/
234 points and 90 comments on Hacker News, linking github.com.
Researchers introduced MaLiang-Harness, a framework that uses multimodal models to iteratively build, check, and fix code for image and video generation. The system evaluates eleven closed-source models on new benchmarks to measure success and quality. ·
Researchers present Omni-IO Skills, a harness that enables agents to process seven modalities. It includes 27 skills covering 38 tasks and boosts input support rates to 100%. ·
Researchers introduced VoxMem, a benchmark containing 3,196 evaluation instances across 177 hours of audio. It tests how large audio language models retain and process multimodal memory from spoken conversations. ·
Researchers introduced PanoVLN, a system using panoramic images for vision-and-language navigation. They found that wider views require changes to action prediction and training to improve performance. ·
The source text lists titles for visual guides and experiments on recurrent neural networks, convolutional neural networks, and transformers, focusing on their accuracy and efficiency. ·
97 points and 2 comments on Hacker News, linking magazine.sebastianraschka.com.
191 points and 138 comments on Hacker News, linking anthropic.com.
Researchers introduced MaLiang-Harness, a framework that uses multimodal models to iteratively build, check, and fix code for image and video generation. The system evaluates eleven closed-source models on new benchmarks to measure success and quality. ·
JoasASantos maintains a repository of AI models that are uncensored or fine-tuned for cybersecurity tasks. ·
Researchers present Omni-IO Skills, a harness that enables agents to process seven modalities. It includes 27 skills covering 38 tasks and boosts input support rates to 100%. ·
X posts other sources are linking to, then named voices on Bluesky and Mastodon
GLM 5.3 flash on AMD GPUs 670 tok/s
2 points and 0 comments on Hacker News, linking an X post.
Trump, Musk, Jensen. and Somewhere, Anthropic
2 points and 0 comments on Hacker News, linking an X post.
The Next 3x in Inference Won't Come from Faster Kernels
2 points and 1 comments on Hacker News, linking an X post.
New Cyber-OSINT model released
53 points and 16 comments on Hacker News, linking an X post.
Changing the difference between Codex plans to be Plus=1x Pro 100=5x Pro 200=10x
2 points and 2 comments on Hacker News, linking an X post.
GLM 5.3 flash on AMD GPUs 670 tok/s
2 points and 0 comments on Hacker News, linking an X post.
OpenAI: Tomorrow we are re-opening the Pro $200 subscription
74 points and 90 comments on Hacker News, linking an X post.
New Cyber-OSINT model released
53 points and 16 comments on Hacker News, linking an X post.
Trump, Musk, Jensen. and Somewhere, Anthropic
2 points and 0 comments on Hacker News, linking an X post.
The Next 3x in Inference Won't Come from Faster Kernels
2 points and 1 comments on Hacker News, linking an X post.
New repos, releases and changelog
| KKKKhazix/AIHOT | 3629 |
| dzhng/jevgrep | 1816 |
| feitangyuan/onetake | 997 |
| kaankiziltug/logo-design-skill | 910 |
| lemomo-ai/lemo-opuscar | 636 |
| glanderness/BeefTV | 575 |
234 points and 90 comments on Hacker News, linking github.com.
Feder-cr released an open-source AI agent that operates its own browser to avoid being blocked on the web. ·
New release v2.1.285 of anthropics/claude-code on GitHub.
OpenAI updated Codex version 0.159.1 to make GPT-6.1 Sol the default model in its bundled and Amazon Bedrock catalogs. ·
41 points and 5 comments on Hacker News, linking github.com.
234 points and 90 comments on Hacker News, linking github.com.
JoasASantos maintains a repository of AI models that are uncensored or fine-tuned for cybersecurity tasks. ·
Feder-cr released an open-source AI agent that operates its own browser to avoid being blocked on the web. ·
New release v2.1.285 of anthropics/claude-code on GitHub.
OpenAI updated Codex version 0.159.1 to make GPT-6.1 Sol the default model in its bundled and Amazon Bedrock catalogs. ·
Data centers, power, chips and what tokens cost
| Thu 24 | 3 |
| Fri 25 | 20 |
| Sat 26 | 1 |
| Sun 27 | 1 |
| Mon 28 | 20 |
| Tue 29 | 33 |
| Wed 30 | 2 |
Stories from DCD, Data Center Knowledge, EIA and Hacker News, by publication day.
One point per OpenRouter reading, last 8 days. Data by OpenRouter, CC BY 4.0.
| gpt-6.1-sol-pro | 10 |
| gpt-6.1-sol-pro:batch | 5 |
| gpt-6.1-sol | 10 |
| gpt-6.1-sol:batch | 5 |
| claude-sonnet-5.5 | 10 |
| median, all paid models | 2 |
438 paid models listed; median of the 60 added in 30 days: $5.00. Rankings data by OpenRouter, CC BY 4.0.
202 points and 296 comments on Hacker News, linking thenationalnews.com.
202 points and 296 comments on Hacker News, linking thenationalnews.com.