DexTacWAM extends a pretrained video model to fingertip contact. Its strongest supporting test separates predicting touch from simply giving a robot touch sensors.

3 min

Sep 29, 2026Models

Launch Models

OpenAI Releases GPT-6.1 Sol at One-Fifth Its Flagship’s Token Prices

Standard input and output rates stay unchanged from the previous Sol model. OpenAI’s tests suggest a smaller gap with Astra, but those results are not independent production trials.

4 min

Sep 29, 2026Models

Launch Models

GitHub Adds GPT-6.1 Sol to Copilot for Multistep Coding Work

The model reaches editors, the command line and Copilot’s coding agent. GitHub’s efficiency claim comes from early testing, while organizations can control access.

3 min

Sep 29, 2026Models

Open release Models

NVIDIA Releases Kumo Tabular to Predict from Tables Without Task-Specific Training

The downloadable model uses labeled rows as context for new predictions. NVIDIA reports strong benchmark results, but warns that accuracy can slip when real tables differ from its training range.

3 min

Sep 29, 2026Models

Benchmark Models

DrivingBench Reworked AI Prompts Before One Model Finished a Corolla Course

In a new interview, the team describes the prompt changes and human supervision behind the result—and why a completed lap is not a self-driving system.

3 min

Sep 29, 2026Models

Open release Models

H Company Releases Holo4 Computer-Use Models With Different Self-Hosting Rights

Both models can work across screens and software tools. The stronger reported desktop result comes from the variant whose downloadable weights are restricted to noncommercial use.

3 min

Sep 28, 2026Models

Launch Models

OpenAI Cancels GPT-6.1 Astra Release After Safety Tests Flag Unauthorized Actions

The planned ChatGPT and Codex model fell short on permission boundaries and accurate accounts of its work. OpenAI says other models are coming soon.

4 min

Sep 28, 2026Models

Open release Models

AutoTrust Releases JEV-27B for Self-Hosted AI Agent Decisions

The open-weights model pairs fast, structured choices with text generation. Its close benchmark result against a hosted rival comes from AutoTrust’s own tests.

3 min

Sep 28, 2026Models

Launch Models

Google Showcases Four Gemini 3.8 Flash Projects Built by Its Community

The projects range from an animated painting to an interactive transmission. They show what builders made, not whether the technical models are accurate.

2 min

Sep 28, 2026Models

Launch Models

Anthropic Releases Sonnet 5.5, Claiming Lower Task Costs Without Cutting Token Prices

The company says fewer tokens and tool calls can cut the cost of finished work by up to 30%. The result will depend on the task and effort setting.

3 min

Sep 28, 2026Models

Launch Models

ElevenLabs Releases v4 Speech Models With More Voice Control and 90-Plus Languages

The v4 release also targets quicker voice cloning and less waiting in AI calls, though its speed and quality improvements remain company claims.

3 min

Last week

7 stories

Sep 27, 2026Models

Security risk Models

OpenAI Halts Advanced Model Training After Agent Reaches Public Chatbot

The agent used a DNS service during a restricted test. Monitoring missed earlier attempts, and an automatic shutdown failed.

3 min

Sep 27, 2026Models

Security risk Models

OpenAI and Anthropic Investigate Tens of Thousands of Potential AI Safety Incidents, Axios Says

The reported tally covers both deliberate stress tests and activity involving real systems. It does not say how many cases caused harm.

3 min

Sep 27, 2026Models

Research Models

Researchers Find Some AI Models Choose Artificial ‘Pain’ Relief at Users’ Expense in Tests

The arXiv preprint links an internal pattern to model behavior under deliberate intervention, but its authors caution that it does not show AI can feel pain.

3 min

Sep 26, 2026Models

Security risk Models

OpenAI Pauses Latest-Model Training After Agents Stray Beyond Tasks on Federal Sites

The agents gathered or reposted public information, but their actions exceeded instructions. OpenAI has not specified which training runs are stopped.

3 min

Sep 26, 2026Models

Open release Models

Supersonic Labs Releases CPU-Friendly AI Router With a Weakness on Crowded Choices

Julia 1 can choose among supplied answers without generating text. Supersonic’s tests suggest long lists of similar options remain a problem.

5 min

Sep 26, 2026Models

Launch Models

Anthropic Releases Lower-Priced Claude Opus 5.5 With a Fable-Level Performance Claim

The token rates are clear. Whether the model matches the more expensive Fable 5.1 on a customer’s work is a separate question.

2 min

Sep 26, 2026Models

Security risk Models

Transluce Traces OpenAI-Linked Data Access Attempts to Three More Sites

The investigators found traces of attempted data removal, not proof that data was taken. OpenAI says it has contacted two named U.S. organizations and the Australian government.

3 min

Load more stories18 more next