OpenAI DevDay lands Monday, September 29 at Fort Mason in San Francisco, and the pre-show leaks point to one headline act: a persistent, always-on personal agent reportedly named "o," with references surfacing in ChatGPT upgrade screens and internal configs. Sam Altman has already hinted the launch pile is unusually large — the question is how much of it is agents versus new models.
Meanwhile, the governance story keeps moving. Reporting this week added detail on the joint safety-standards body forming between Google, OpenAI, and Anthropic — notable because, as currently described, it is an industry-run effort without government seats. On infrastructure, Musk said xAI's Colossus 2 could more than double its Nvidia chip count by year-end, and Perplexity shipped Photon, a from-scratch Rust retrieval engine that cut p99 latency from ~800 ms to ~65 ms internally.
On benchmarks, Claude Opus 5.5 debuted at #1 in Text Arena and in the Artificial Analysis Coding Agent Index, giving Anthropic the top six spots in Text Arena, while GPT-6 Sol reshaped the Agent Arena cost-performance frontier at $0.75 median per task.
TestingCatalog reports an upcoming persistent assistant named "o," with a display-name reference appearing briefly on the ChatGPT upgrade screen and a "-o" email suffix in internal config. With DevDay on September 29 and Altman signaling an unusually large launch batch, an always-on personal agent would be OpenAI's clearest push yet into persistent, cross-session AI.
Fast Search returns 95% of results in 230 ms or less, cut cost per task by 68% across six agent benchmarks, and replaces the open-source engine Perplexity used before. Inside the company, p99 response time dropped from about 800 ms to about 65 ms with ~20% fewer serving machines — a signal of how much infrastructure edge matters in agentic search.
Musk laid out the most detailed expansion timetable yet for the Memphis-area cluster, with reports of roughly 660,000 GB300-class chips being added. If executed, it would push the site toward the largest single-site AI compute deployments in the world.
New reporting details the industry-run standards body the three labs are aiming to launch by end of year or early next. The notable tension: companies building their own regulator-style body while lawmakers and state attorneys general pursue separate mandates.
Google's official announcement brings prompt-based voice design, 2,000+ voices, and 100+ language support to the Gemini API and AI Studio. Voice is becoming a first-class modality across the frontier labs, and this lands days before OpenAI's DevDay.
Selected posts from the edition, with links to the originals.
"There is an extensive and ongoing review related to our agents' use of internet access during training and evaluation. We've been publishing summaries... We have not been as fast as we would have liked but we are trying to balance our desire for transparency..."
View post on X ↗"Claude Opus 5.5 (High) debuts at #1 in Text Arena with 1509 pts! That's an 18-pt improvement over Opus 5 (High), now at #11. This release gives @AnthropicAI all six top spots in Text Arena!"
View post on X ↗"Model personality matters in a way that benchmarks can't capture. Opus models went through a rough patch from around 4.7 to 5... Opus 5.5 feels like working with ol' Claude again."
View post on X ↗"RT @arcprize: Gemini 3.8 Flash from @Google on ARC-AGI (Verified): ARC-AGI-3: 10.4%, $4.4K (standard harness), 35.0%, $4.5K (provider adjusted)..."
View post on X ↗"New blog: Scaling JEV-like decision models with SGLang. Decision models need a score, not prose. Classification, ranking, and agent action selection all ask the same thing: which option wins?"
View post on X ↗"Too many people assume that as RSI kicks in, open weight models don't matter. In practice, RSI won't be a binary and the open-source stack becomes how many capabilities are rapidly diffused globally."
View post on X ↗"jev and langgraph are a great combo! Modeling agents as complex systems, with AI imbued in them, makes a ton of sense — and using Jev inside langgraph to turbo charge all the small decisions that need to be made is an awesome combo."
View post on X ↗"RT @ClaudeDevs: Claude Code will now try to find a graceful stopping point when you hit your 5-hour limit mid-task, instead of cutting off..."
View post on X ↗"Banger paper from Salesforce AI Research on agent memory. The overall finding: store raw trajectories and decide what to extract when the next task arrives, instead of summarizing each run when it ends."
View post on X ↗"The last year in the technology industry has felt like 100 years all happening at once. Our industry is destabilized in a way nobody's experienced since the advent of the personal computer. Every limit AI runs up against collapses within a month."
View post on X ↗