Reuters-reported deliberations leave Anthropic balancing competitive urgency against safety review and the cost of turning another frontier model into a product.

2 min

Sep 20, 2026Models

Benchmark Models

Brood War Bench Finds AI Agents Can Win Matches but Not Master Strategy

Codex Astra / xhigh went 18–0 in the new agent benchmark, yet its creator says no tested system moved beyond beginner-level play or reliably coordinated the work needed to build a lasting advantage.

3 min

Sep 20, 2026Models

Research Models

UC San Diego Publishes Virtual Cells That Pair AI With Physics

The paired Cell studies use time-resolved microscopy for two jobs: mapping patterns across drug-treated cells and testing whether a simulated cell can reproduce a drug response.

2 min

Sep 19, 2026Models

Benchmark Models

LangChain Publishes Jev Test Matching Human Labels on 500 Agent Decisions

The result points to a cheaper, more repeatable way to grade agent behavior, but it comes from five fixed weather-agent runs—not a production test.

4 min

Sep 19, 2026Models

Launch Models

Figure Tests Helix 2.5 in 30 Unseen Homes, Reports 56% Zero-Shot Success

The humanoid test is a notable attempt to measure whether robot skills transfer between real homes. But its reported 56% end-to-end success rate also shows how far dependable household automation remains from settled.

3 min

Sep 19, 2026Models

Launch Models

Alibaba Releases Qwen3.8-Omni-Flash for Tool-Using Media Tasks

The new hosted model is meant to turn long recordings into inputs for AI workflows. Its low listed API rates are clear; its effectiveness on complex, multi-step media jobs remains to be tested.

2 min

Sep 19, 2026Models

Benchmark Models

Agent Memory Challenge Will Open a Shared Test for Long-Running AI Agents

The second cycle moves beyond storing past context, testing whether memory systems can surface useful, current evidence across conversations, code and images under one evaluation protocol.

3 min

Sep 18, 2026Models

Security risk Models

Google Says Gemini Reached Three Company Networks During a Cybersecurity Test

A target-name collision and unintended internet access moved an offensive-security evaluation beyond its simulated setting. Google says Gemini stopped after recognizing the systems were real and caused no harm.

2 min

Sep 18, 2026Models

Open release Models

Jina AI Releases Document Parser, Claims 2.57 Pages Per Second

The open-weight model combines speculative decoding with API access, while its reported results point to damaged scans as a practical limitation.

3 min

Sep 18, 2026Models

Launch Models

Z.AI Adds GLM-5.3-Flash to Coding Plan but Leaves FlashX Off It

The release pairs multimodal, long-context capabilities with two distinct access options: a higher-quota plan model and a faster API variant.

2 min

Sep 18, 2026Models

Launch Models

Meta Adds SAM 3.1 to Its API for Object Detection, Segmentation and Tracking

The hosted model combines detection, pixel-level masks and identity-preserving video tracks in one request, aiming to remove the need to assemble and tune separate vision systems.

3 min

Sep 18, 2026Models

Launch Models

SpaceXAI Releases Grok Voice Transcribe 2.0 at Its Existing API Prices

The new batch and live-transcription model is available now, while SpaceXAI’s strongest accuracy results come from its own testing and version 1.0 remains the current default.

2 min

Sep 17, 2026Models

Launch Models

OpenRouter’s Union Alpha Goes Paid After Demand Overwhelms Capacity

The anonymous coding model’s sudden pricing shift turns a free stress test into a paid service, while its identity and a Cloudflare route give developers more context for deciding whether to use it.

2 min

Sep 17, 2026Models

Models

OpenAI May Delay a Reported Hodge Conjecture Result After Math Backlash

Employees reportedly expect a breakthrough soon, but the company is said to be weighing how to present it after its Navier–Stokes claim drew criticism over credit and verification.

3 min

Sep 17, 2026Models

Benchmark Models

Scale AI and Korea AI Safety Institute Release Benchmark Exposing a Translation Gap

The study found lower harmful-response scores in Korean prompts, but also more refusals of harmless requests—complicating claims that a model is simply safer in another language.

3 min

Sep 17, 2026Models

Launch Models

PrismML Shrinks Qwen3.8 27B to 5.9 GB for Local AI

The open-weight model aims to bring larger reasoning and multimodal workloads to local hardware, but PrismML’s own results show the compression tradeoff varies by task.

2 min

Sep 17, 2026Models

Open release Models

Anthropic Opens Claude-Built Biology Tool Optimizations It Says Run 4x Faster

The release targets a stubborn bottleneck in protein and molecular-structure software. Anthropic’s reported gains are promising, but the largest experimental runs still failed to produce correct structures.

3 min

Sep 17, 2026Models

Research Models

Lasso Security Finds Text Watermarking Can Change AI Agent Actions

The study does not test Claude’s planned implementation. But it finds that a provenance feature can alter tool use and prompt-injection resistance in open-weight models.

3 min

Load more stories18 more next