
Tencent releases Hy4 with vision and computer use; weights come later
Hy4 is API-only for now, scores 35 on Artificial Analysis, and trails Kimi K3.1 Pro by a wide margin.
42 stories, newest first · one card per story

Hy4 is API-only for now, scores 35 on Artificial Analysis, and trails Kimi K3.1 Pro by a wide margin.
NorthCUA turbocharges computer-use agents through an experience flywheel and staged learning achieving state-of-the-art results among open-source models.

Version 1.0.2 under Apache-2.0 drives native apps on macOS, Windows and Linux without screenshots or a bundled model.
Mage-VL-Cowork is a 4B parameter hybrid computer use model achieving 66.0% on OSWorld-Verified.
HCU combines programmatic, semantic (AX), and visual (GUI) interaction channels to enable flexible channel selection and combination in computer-use agents.
The 4B hybrid CUA is trained through a five-stage curriculum using skill-conditioned generation workflows that synthesize 293806 tasks.

Standard API prices are half those of GPT-5.6 Terra and GPT-6.1 Sol, and cybersecurity is rated Critical.
OpenAI calls it a step up from GPT-5.6 Terra for proposals, technical drawings, forms and video editing.

API pricing is $1 per million input tokens and $5 per million output, half of GPT-5.6 Terra, and it is available now.
Standard API pricing is $1 per million input tokens and $5 per million output tokens, 50% below prior Terra and Sol rates.
Seven frontier agents hit up to 91.6% action completion but only 47.7% full task success.
Across 64 benchmarks, Act is explicitly evaluated, while only 5 assess Anticipate and 3 assess Verify.
Across 64 benchmarks, only five assess Anticipate and three assess Verify, so both remain largely hidden behind task success.

The visible page is a cookie notice; only the title states a 67.59 score for computer-use agents.
Qwen3.8-27B with evolved interfaces scores 21.3%, above GPT-5.5 and Claude Opus 4.8 at far lower cost.

OpenAI's ChatGPT Chrome extension can now read open tabs and act on them with Computer Use.
OpenGCUA open-sources a generalist computer-use stack that lifts OSWorld-v2 scores and cuts steps.
CLASHarness benchmarks 600 multi-agent computer-use conflicts; frontier models fail most closing actions.

Anthropic released Claude Haiku 5.5, about 75 percent cheaper than Haiku 4.5, with much stronger computer use.

Startup Hark widely released Hark Pro, a privacy-focused AI assistant trained for computer use, with a device planned for 2027.

OpenAI partnered with Ironclad to train GPT-6 Astra on complex contracting workflows, beating GPT-5.6 Sol.

Unreleased Gemini Desktop Full Access could let the assistant control files, apps, and network activity on a Mac.

Google is testing Gemini Desktop controls that could access Mac files, apps, and the web with fewer prompts.
AutoTrust released GEV-26B-Decide, an open 26B Gemma-4 decision model scoring 62.48 on Decision Index 0.2.1.
Today 1.5K readers · 2.4K views · 754 reads · all time 189K views