AI Daily · 2026-09-30

OpenAI's DevDay 2026 anchored the day: the company launched GPT-6.1 Sol, offering near-Astra intelligence at about one-fifth the cost, along with the personal assistant Dots, ChatGPT Space for team collaboration, the high-speed Ultrafast tier, Pro 500, and a preview of the Decisions API; it also partnered with SBDC to train small-business advisors. Evaluation and infrastructure are shifting toward standardization and automation: Hugging Face introduced the Open TTS Leaderboard, using objective metrics such as WER, RTFx, and speaker similarity instead of human preference votes; LangChain added Jev as a fast, low-cost judge, Google's Diffusion Controller improved prompt alignment, and Replit explored letting models delegate to subagents dynamically. Anthropic routed inbound sales to Claude Managed Agents, roughly doubling conversion to opportunities, and Simon Willison built a local face-blurring tool called Photo Scrubber with Astra.

North America · First-hand

Anthropic

⭐⭐ [Product Update] How Anthropic's sales team rebuilt inbound with Claude Managed Agents

Anthropic's sales team built a buying agent on Claude Managed Agents (beta) to handle prospects coming from its Contact Sales form. Previously, buyers waited days for replies to simple questions about plan pricing, seat minimums, or HIPAA contract requirements that were already covered in the docs. The agent holds thousands of conversations a day, asks follow-up questions, answers pricing, security, and data questions, and recommends a plan and seat count; each conversation ends in a purchase, a hand-off to a rep with the full conversation details, or a quick answer. Anthropic says customers escalated this way become opportunities more than twice as often as leads from the old form and close about five days faster, and the experience is opt-in so customers choose whether to talk to an agent or a rep. The team chose Managed Agents because the platform handles hosting, session management, and tool orchestration; one engineer built the initial agent in a few weeks, and sales and content leads edit the system prompt directly in the Console, with changes going to a staging agent first.

OpenAI

⭐⭐⭐ [Model Release] Introducing GPT-6.1 Sol

OpenAI introduced GPT-6.1 Sol, an upgrade to GPT-6 Sol that nearly matches GPT-6 Astra's intelligence on agentic coding, computer use, and professional work at one-fifth of Astra's standard input and output token prices; cached input costs $0.10 per million tokens, 95% less than standard input pricing. On DeepSWE v1.1 it matches GPT-6 Astra at roughly one-fifth the cost and beats GPT-6 Sol's best score by 6.4 percentage points at lower reasoning effort and cost, while on GDP.pdf it scores higher than Opus 5.5 with fallbacks at less than half the cost per task and approaches Astra at about one-fifth the cost per task. On AutomationBench it scores 2.2 points above Opus 5.5 at medium reasoning effort at roughly one-third the cost, and 4.8 points above GPT-6 Sol at the same setting; on OSWorld 2.0's offline set it outperforms GPT-6 Sol by seven points at maximum reasoning effort at less than half the cost, coming within 2.1 points of Astra at roughly one-seventh the cost per task. On Terminal-Bench Science 0.1 it more than doubles GPT-6 Sol's score at maximum effort, averaging $5.47 per task versus $23.21 for Opus 5.5 and $23.80 for Astra, while Astra still holds the highest tested score at 68.1%.

⭐⭐ [Other] Helping small businesses put AI to work

OpenAI announced a partnership with America's SBDC, the national network of small business development centers, to help small businesses use AI at work. In this initial phase, through OpenAI Academy's Community Trainer Program, OpenAI plans to train around 150 SBDC advisors and bring hands-on workshops to SBDC networks nationwide, aiming to reach at least 1,000 small businesses through in-person training. The plan includes training and credentialing advisors, hosting at least one three-hour Academy workshop per participating network, and jointly creating practical AI guides plus an industry-specific companion guide covering potential topics such as marketing, operations, manufacturing, exporting, and government contracting. Alongside the partnership, OpenAI released the report "Small Businesses, Bigger Capabilities," which reports that in one week from September 9-15 about 4 million small-firm employees worldwide used OpenAI's tools, nearly one in five of them at businesses with fewer than 10 employees, and that in August agentic output tokens for products like ChatGPT Work and Codex made up two-thirds.

⭐⭐ [Product Update] DevDay 2026 Recap

OpenAI published a recap of DevDay 2026, calling it its biggest yet, with more than 20 announcements spanning ChatGPT, Codex, its models, and new ways of working with AI. The directions mentioned include agents that can take on ongoing responsibilities, new ways for people and AI to work together, and opening ChatGPT as a shared surface where humans and agents collaborate and where developers can launch native experiences directly to a collective 1.2B weekly users. The listed items include building with Codex and the API, customizing ChatGPT with Plugins, doing more with a ChatGPT subscription, and the new product dots. The page gives no version numbers, parameters, or benchmark data for the individual announcements.

Google

⭐⭐ [Research] How Diffusion Controller unifies and simplifies AI image generation

Google Research introduces Diffusion Controller, a lightweight 「steering damper」 network that steers image generation more precisely for significantly better prompt alignment, and can be attached to access-restricted closed-source models to boost image quality without breaking baseline stability. Rather than treating generation as a sequence of isolated steps, the framework reframes the whole denoising process as a smooth continuous control problem: the add-on network dynamically adjusts the generation trajectory while the base model's weights stay frozen. The reported results show the lightweight add-on network outperformed the industry standard for matching human preferences, and its fully unlocked version — a fine-tuned model with unrestricted access to internal weights — reached a 90% win rate over the baseline. The post uses a 「lizard wearing sunglasses」 prompt to illustrate the trade-off between prompt adherence and image quality, and shows a side-by-side visual matrix comparing base pre-trained models, LoRA, and Diffusion Controller outputs on complex prompts.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Product Update

⭐⭐ [Product Update] Jev is now available in LangSmith Evals

LangChain announced that Jev is now available as a judge for evaluations in LangSmith, usable as an online evaluator from the Evaluators tab of any tracing project. The post recaps the two main approaches to agent evals: code-based evaluators are fast and deterministic but cover only behavior that can be fully specified in advance, while LLM-as-a-judge handles open-ended behavior but is slower, more expensive, non-deterministic, and introduces an extra error source when converting free text into structured output. Jev, described by TypeSafe AI as a System One model, does not generate text; it evaluates a state and returns typed answers and probabilities, answering three question types: noul (yes/no probability), choice (pick one option from a set), and score (rate on an ordered scale). According to TypeSafe AI, Jev is up to roughly 450x cheaper and 200x faster than comparable LLMs on classification tasks and can evaluate multiple questions about the same state in parallel.

⭐⭐ [Product Update] Free the models: Harness design at the frontier

The post describes the design of the Replit Agent harness: model routers, whether based on heuristics or a small model that picks an LLM each turn, are always less capable than the model they select for, so Replit Agent lets the model decide instead — the main agent (core loop) picks its subagents' tier and effort and adjusts its own as the task unfolds. Given that freedom, GPT-6 Astra hands routine implementation to cheaper subagents and decides where its tokens are worth spending; on DeepSWE and Terminal-Bench, Replit Agent is Pareto-efficient against Astra on its own, and beats a sidekick architecture (the same setup with one long-lived worker) by 11 and 16 points. The authors argue every model release invalidates assumptions baked into the harness: as models get stronger at long-horizon tasks they need less harness-layer scaffolding and lean more toward delegation on their own, using subagents for context management and parallelism. But the frontier is jagged — the strongest coding model is not necessarily best at UI design, slides, or emails — so the harness is built to let each model work its own way, with the guardrails it still needs and quality at minimum cost as the goal.

Research

⭐⭐⭐ [Research] Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning

A Hugging Face blog post introduces the Open TTS Leaderboard. As of September 30, 2026, the Hub hosts more than 8K TTS models, yet evaluation remains fragmented and unstandardized. Existing arena-style leaderboards such as TTS Arena v2 and Artificial Analysis Voice Arena rely on human preference votes and Elo scores, which cannot scale with the pace of TTS releases—only 16 of the 92 models on Artificial Analysis are open-weights—and voter consistency is another limitation. The new leaderboard instead uses objective metrics: intelligibility via WER/CER between the prompt and the transcript of generated audio using Qwen3 ASR, speed via inverse real-time factor for batched offline inference on an H200 GPU and time-to-first-audio for streaming batch size 1 on H200 and CPU, and speaker similarity via cosine similarity between WavLM speaker embeddings of generated audio and the reference clip, cutting evaluation from weeks to hours. The authors stress it does not replace human preference ranking: ASR-based WER is only a proxy for intelligibility and speaker similarity estimates voice identity preservation, neither of which directly measures naturalness, expressiveness, or listener preference. The default view ranks models by macro-average WER on the English splits of Seed TTS Eval and CV3 Eval (zero shot), with hexgrad/Kokoro-82M, Supertone/supertonic-3 and fishaudio/s2-pro leading, alongside Pareto plots of WER, batched inference (RTFx) and model size, and toggles for multiple languages.

Other

⭐⭐⭐ [Other] OpenAI DevDay 2026 live blog

Simon Willison live-blogged OpenAI's DevDay 2026 keynote from Fort Mason in San Francisco. Sam Altman announced Dots, a personal AI agent powered by the Astra model that users name and that is represented by a blob-like avatar, with heavy voice interaction in the demo; OpenAI also introduced ChatGPT Space for teams to collaborate with their Dots, available to ChatGPT Pro and Enterprise customers the same day, with specialist Dots planned for legal, finance and other areas. On the model side, OpenAI launched GPT-6.1 Sol, which Altman described as near-Astra intelligence at a fifth of the price, available that day, plus Ultrafast, an 8x faster mode up to 300 tokens/second priced at 6x the standard rate, available for Astra 6 and soon for Sol 6.1. OpenAI also introduced a Pro 500 subscription with Ultrafast access and 25x the usage of Plus, put its $200/month plan back on sale for new subscribers, and previewed a Decisions API in which the Luna model chooses from a predefined set of options to respond in a fraction of a second.

⭐⭐ [Other] Photo Scrubber — local face blur & metadata removal

Simon Willison released an experimental tool called Photo Scrubber for locally blurring faces and removing photo metadata. The idea came after he took a photograph of protesters and decided he did not want to share images of strangers with identifiable faces, so he had GPT-6 Astra build a tool that detects faces and blurs them automatically. The tool uses Google's MediaPipe C++ library, compiled to WebAssembly via @mediapipe/tasks-vision, along with the BlazeFace face detection model.