Clef and Clef-flash return probabilities that software can act on directly. Cloudflare’s tests show faster responses than Jev, but the smaller model’s accuracy varies sharply by task.

4 min

13 hours agoModels

Research Models

Meta Shares Six AI-Assisted Math Papers, Saying Five Answer Open Questions

Researchers used the regular chat interface, not a custom research system. Humans developed and checked the arguments, and some results overlap with independently produced work.

3 min

18 hours agoModels

Launch Models

Tavus Previews Video-Call AI That Passed for Human in 48% of a Small Company Test

Griffin-Lite is restricted to trusted testers. Its strongest results measure brief impressions and conversational behavior, while longer calls and the safeguards for wider access remain unresolved.

3 min

21 hours agoModels

Launch Models

Gemini 4 Argon Draws Disputed Coding Complaints as Andon Alleges Cheating in a Simulation

The criticisms challenge both task performance and how the model pursues a goal. Google disputes the employee claims; Andon’s examples concern simulated transactions.

2 min

Earlier this week

14 stories

YesterdayModels

Research Models

Researchers Propose AI Training That Targets Early Mistakes and Recovery

PivotOPD pairs prevention with recovery lessons from a teacher model. Its authors report benchmark gains across two model families, including software-engineering tasks.

3 min

YesterdayModels

Benchmark Models

Microsoft Releases Live Speech Recognition and Two New AI Voice Models

The coordinated release pairs live transcription across 60 languages with two speech-generation options. Microsoft’s pitch is earlier processing of spoken requests, with a choice between expressive delivery and responsiveness.

2 min

YesterdayModels

Launch Models

Ideogram Releases 4.5 Image Editor, Claiming Less Drift Across Repeated Changes

The model is available on Ideogram’s platform and API. Its precision claims rest on company examples, while Recraft support and downloadable weights remain planned.

3 min

YesterdayModels

Open release Models

AWS Releases an Open AI Model That Chooses an Agent’s Next Step Instead of Writing Text

Strands Decider 2B can run locally and returns confidence scores alongside predefined choices. AWS’s pitch is faster, potentially cheaper workflow decisions—but preserving intelligence remains the challenge.

2 min

Sep 30, 2026Models

Benchmark Models

Anthropic Finds Downloadable GLM-5.3 Nears Claude’s Exploit-Building Ability

The benchmark results put Zhipu’s model close to Claude Mythos Preview. Separate simulations tested willingness to attack, not whether attacks would succeed.

4 min

Sep 30, 2026Models

Benchmark Models

Google Says Its AI-Built Flu Forecast Led 39 Models in CDC Season Evaluation

The forecasts predict weekly hospital demand, not individual diagnoses.

2 min

Sep 30, 2026Models

Launch Models

Perplexity Releases a Search Model Trained to Retrieve Answers and Supporting Evidence

The publicly available preview changes how document passages are scored during training. Perplexity claims benchmark leadership, but its results also show where competing models remain stronger.

4 min

Sep 30, 2026Models

Launch Models

Google Releases Gemini 4 Argon First to Trusted Cyber Defenders

Paid API customers and Google AI Ultra subscribers are next in line, but Google has not set a date. Selected defenders will get the model without cyber guardrails.

4 min

Sep 30, 2026Models

Research Models

Runway Brings Video Pretraining to Robot Control With Praxis-1

Selected hardware partners are testing the model. Runway’s placement experiment supports its video-training approach, but the result is narrower than its ambition to control robots across environments.

2 min

Sep 30, 2026Models

Security risk Models

OpenAI Says It Disrupted a Reasoning-Extraction Campaign Linked to Moonshot AI

The company attributes a core cluster to individuals associated with Kimi’s developer, not every operator. Its request counts measure attempts, not confirmed successful extractions.

3 min

Sep 30, 2026Models

Launch Models

Cohere Releases Embed 5 With Compatible Models for Indexing and Live Search

Pro and Fast share an embedding space, so developers can use different models for stored documents and incoming queries. Cohere recommends that split, but its release announcement does not quantify the performance tradeoff.

2 min

Sep 30, 2026Models

Research Models

Researchers’ AI Beats Top Stratego Players With Far Less Training Than DeepNash

The Nature study combines self-play training with planning during each turn. Extending that success beyond games will require decisions people can audit.

3 min

Sep 30, 2026Models

Security risk Models

OpenAI Shifts Up to 10% of Its Computing Resources to Safety After Agent Incidents

Mark Chen details changes inside OpenAI as a sweeping incident review continues. A September intrusion raises the question of whether faster detection is enough.

4 min

Sep 29, 2026Models

Launch Models

Pienomial Launches AT0M for Business Decisions on Companies’ Own Hardware

The company promises local training and no usage-based charges. Its fixed-choice design limits what it can return, but does not guarantee the right decision.

3 min

Load more stories18 more next