AI models, services and robotics
Anthropic launched its small model Haiku 5.5 and cut prices for requests of up to 100,000 tokens by 90% from its predecessor. It scored 72.4% on OSWorld 2.1, ahead of GPT-6 Luna in the same price tier, and is the first Haiku model with an adjustable effort setting.
By 김현국
Anthropic launched its small model Haiku 5.5 and cut prices for requests of up to 100,000 tokens by 90% from its predecessor. It scored 72.4% on OSWorld 2.1, ahead of GPT-6 Luna in the same price tier, and is the first Haiku model with an adjustable effort setting.
OpenAI has published a case study on how quantitative trading firm Jump Trading uses GPT-6 Astra to hand days-long market research to agents. Even trading signals produced by agents are treated as inputs that may be wrong, and humans perform the final validation.
Google DeepMind has made SynthID Detector, which checks images, video and audio for AI-generation watermarks, publicly available worldwide in English. One portal now checks output from OpenAI, NVIDIA and Kakao tools as well, with Apple joining soon.
OpenAI has posted 722 mathematical manuscripts produced by an unreleased internal model to GitHub, grouped into 372 result families. The collection includes claims on the quasi-Riemann hypothesis and the Hodge conjecture, but many results are not yet formalized, and the model's name and prompts were left out.
Google DeepMind has released EmbeddingGemma 2, a 740-million-parameter open model that places text, code, images, audio and video in a single embedding space. A modular design that loads only the encoders an app needs brings multimodal search to smartphones.
Media and software company Every has released the shared company agent it built on Claude Managed Agents after about six months of internal use. The agent works inside Slack threads and saves corrections as skills, turning one person's know-how into something the whole company can use.
OpenAI has opened its Decisions API, which uses GPT-6 Luna to return only probabilities, choices, and scores, to all developers in public beta. Input costs $0.10 per 1M tokens with no charge for output tokens, and the company said it expects general availability within weeks.
Anthropic has opened its startup program for founders building on Claude to more companies, offering a free year of the Team plan, $1,000 in API credits, and partner benefits worth up to $45,000. The credits can be used only on its own API within six months, a design that binds early-stage startups to its ecosystem.
Anthropic has released a public beta of a Claude add-on that reads files in a sidebar next to Google Docs, Sheets, and Slides and edits them in place, available on every paid plan. By default, each edit requires the user's approval, and the Google permissions it requests are limited to the single open file.
Five case studies apply Google Earth AI's Population Dynamics Foundation Model to measles vaccination, cardiovascular disease, dengue, postpartum depression and cholera. Adding a single input to existing epidemiological models improved results where statistics lag or stop at borders.
U.S. medical AI startup Nolla Health has launched a pilot in which AI directly issues first-time acne prescriptions under a regulatory mitigation agreement with the state of Utah. The 50-page agreement limits the AI to eight medications, sets three stages of physician oversight and bars liability waivers in the terms of service.
Google Research released a 116-page workshop report laying out 51 open research problems in the privacy and security of AI agents. Drawing on the theory of Contextual Integrity, it proposes a policy engine that screens agent actions against their context.
OpenAI released a demo video of Codex for the terminal showing developers assigning tasks by voice and managing agents across several projects from a single screen. The features in the video shipped over five version updates in the ten days starting September 22.
On-device AI advisor trained on naval gun manuals supports equipment operation
Supports large-scale data ingestion, storage and analytics without building or managing dedicated servers… built on open standards to reduce vendor lock-in and the cost of moving data between clouds
On October 5, Cohere unveiled North 2, an enterprise AI agent platform with a redesigned orchestration system, memory and caps on token spending. It also opened the platform to models from other companies, including Anthropic, OpenAI and Google.
Wayfair has used OpenAI models to validate more than 11 million product specs and now automates 41,000 supplier support tickets a month, OpenAI said. In an A/B test, product impressions and clicks rose significantly.
OpenAI will add an invisible text watermark called textGrain to ChatGPT and Codex output in line with the EU AI Act, and offer it as an opt-in feature to API customers. It published figures showing detection falls to 17% when a quarter of the words are swapped, and opened its detector only to researchers.
OpenAI will test a new visual ad format that appears during image generation in ChatGPT, starting later this month in the US. It also expanded its measurement partners and unveiled brand suitability evaluation pilots.
Anthropic opened a training institution that will spend $100 million to develop 10,000 engineers who deploy AI inside enterprises by the end of 2027. Accenture, McKinsey, Morgan Stanley and Novo Nordisk are among those in the first cohorts.
OpenAI released an official model guide on choosing and instructing GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna for different kinds of work. It covers four reasoning levels, caching, steering for long-running tasks and four customer cases.
Google Research has unveiled a federated learning system that decrypts device data only inside server-side trusted execution environments and lets outsiders verify the code through a public log. A Gboard English model trained on it in three weeks with a privacy budget three times smaller.
The Den, a Denver social club for parents, cut its leadership team's workload by 10 to 15 hours a week with ChatGPT Work while preparing a second location. Grant applications dropped from three days to two hours, and liquor-license applications from four days to three hours.
U.S. supermarket chain Albertsons is broadening its work with OpenAI from employee tools to grocery shopping in ChatGPT, promotional analysis and advertising. A retailer that serves 36 million customers a week across more than 2,200 stores has put its operations and commerce on a single AI provider.
