AI Models & Tools

Gemini API CLI lets coding agents preview requests before using a key

Google's beta Gemini API CLI gives existing coding agents a schema and offline request preview. The dry run cannot establish live model access, cost or results.

Why OpenAI moved Habitat storage back to shared tenancy

OpenAI says Habitat returned to shared Cosmos DB storage after dedicated containers made product launches hard to provision. Hot-key limits, request priority and caching carry the tradeoffs.

SoL-Pi reduces coding-agent token traffic through four harness mechanisms

On 51 public EdgeBench tasks, SoL-Pi's authors report 44.7–49.0% less token traffic and about one-third lower estimated API cost than Pi. At xhigh, the full stack's mean score fell from 44.833 to 42.003 with GPT-5.6 Sol and from 44.756 to 42.224 with Opus 5.

OpenAI Decisions API is generally available: answers, pricing and limits

OpenAI’s Decisions API is generally available with JSON inputs, question images and caching. Its typed answers, prices and limits define what builders can evaluate now.

Continues OpenAI DevDay 2026: launches, availability and what builders can use

Community hub organizes revisions to OpenAI’s math papers

A community fork of OpenAI’s mathematics collection is recording manuscript edits, reviews and follow-up work. Its growing notes do not certify the proofs.

Continues From unit distances to Navier-Stokes: AI mathematics reaches a new phase

Claude-designed proteins bound four scorpion toxins in lab tests

Anthropic reports that 17 of 211 selected protein designs met its toxin-binding criteria. The study did not test whether they neutralise venom in cells or animals.

AWS's TorchNeuron brings a native PyTorch backend to Trainium

AWS’s Neuron 2.33 release introduces a public-beta PyTorch backend for Trainium2 and Trainium3. Its supported versions, operator coverage and compilation limits define what engineers need to check in an evaluation.

Harness-Delta Attribution separates agent gains from benchmark shortcuts

A new research method separates detected benchmark overfitting and extra inference compute from the remaining gain in optimized agent harnesses.

FLARE tests AI grid restoration decisions under degraded reports

A new simulation benchmark finds that AI agents often match a grid-restoration reference plan, but their worst cases and unsafe switch proposals keep independent safety checks central.

HybridAgent tests latent reasoning alongside text-based tool use

A new preprint and released code show how HybridAgent alternates latent reasoning with text tool use. Its measured gains come from the authors' benchmark setup.

Liquid’s d1-ultra adds a hosted model for decisions without generated text

Liquid AI released d1-ultra as a hosted model that returns probabilities for structured text and image decisions without generating answer tokens. Its benchmarks and price come from the company.

vLLM’s Kimi K3 decode path reaches 842 tokens a second in a specific GB300 test

vLLM reports 842 versus 488 aggregate decode tokens a second for Kimi K3 on 16 GB300 GPUs under a specific DSpark setup. Prefill is excluded.

UfiSpace shows liquid-cooled 1.6T switch at OCP summit

UfiSpace demonstrated an early-access 102.4 Tbps switch with liquid cooling at OCP. The public material establishes a live demo, not a deployed or independently measured data-center system.

Replit names OpenAI its preferred Agent partner while Claude remains available

Replit CEO Amjad Masad says OpenAI is now the preferred partner and default for Agent, while Claude remains an option. Replit's model controls show how that choice varies by mode and plan.

FLUX 3 Image Klein: open weights, API access and license limits

Black Forest Labs has released FLUX 3 Image Klein weights and a hosted API. Its model package, runtime memory and commercial permissions differ in ways teams need to check.

StudentSim tests AI tutor feedback without measuring student learning

StudentSim compares AI student simulators with learner records and tests responses to guided corrections. Its chess tutor study rates advice, not learning gains in real students.

Claude proposes quantum circuits that reduce estimated gate and qubit costs

An Anthropic and Harvard preprint uses Claude to propose quantum circuits and a separate evaluator to check them. Seven of eight tasks show lower estimated gate and qubit counts under the paper's comparisons.

How BlenderWorld evaluates editable 3D scenes made by AI agents

QwenLM's BlenderWorld benchmark asks agents to submit native Blender projects. Its task rubrics and visual judges test different aspects of a 3D scene, with limits for real production use.