Nov 12, 2026ModelsResearchModelsLung cancer study links AI-mapped cell groups to higher risk of spreadLuCaS+ connects millions of cell profiles with patient histories. Independent-cohort validation supports its findings, but proposed therapies still need laboratory testing.3 min
Nov 12, 2026ModelsOpen releaseModelsThinking Machines publishes Inkling 2 Small weights after hosted-only debutThe download ends the model’s initial closed-weight phase. Hosted access remains an alternative, with Together AI charging $0.30 per million input tokens and $1 per million output tokens.3 min
Nov 12, 2026ModelsLaunchModelsIBM releases experimental AI models trained to separate instructions from dataThe LLMON previews pair a new text format with specialized training. IBM reports sharply lower attack success rates, but its tests cover static attacks—not an enforced security boundary.3 min
Nov 12, 2026ModelsResearchModelsOpenAI study links training choices to harder-to-monitor AI reasoningExperiments identify several routes to stronger silent reasoning and greater control over reasoning logs. But those warning signs do not reliably predict successful deception.2 min
Nov 12, 2026ModelsBenchmarkModelsScale and IMDA find AI models skip needed questions in Singapore legal testSG-LegalClarify evaluates 19 models on Singapore law. Web search improves scores, but gathering all the missing facts still does not reliably produce the right conclusion.3 min
Nov 12, 2026ModelsLaunchModelsBaidu releases ERNIE 5.5 Flash Preview, with model weights due November 25The model is available through Qianfan and free on OpenRouter. Baidu pairs it with two specialist models, while its speed and efficiency claims lack numerical results in the announcement.3 min
Nov 12, 2026ModelsResearchModelsOpenAI finds training on AI-generated attacks improves jailbreak resistanceControlled experiments showed gains against both fixed prompts and attackers that adapt mid-conversation. The results measure the model layer, without production safeguards.3 min
Nov 12, 2026ModelsLaunchModelsGoogle releases Gemini 4 Hydrogen for coding, Search and the Gemini appAPI rates start at $1.50 per million input tokens. Google’s performance pitch combines adjustable reasoning with automatic model switching in Antigravity.3 min
Nov 12, 2026ModelsResearchModelsMicrosoft and Terradot’s AI cuts soil sampling by up to 75% in carbon-removal simulationsThe model combines predictions with physical measurements rather than replacing them. Its potential savings target an expensive obstacle to scaling enhanced rock weathering.3 min
Nov 12, 2026ModelsBenchmarkModelsTrellis Data introduces 220-billion-parameter AI model for Australian businessesThe company promises a locally developed model that can handle more professional tasks, but its reported global ranking comes without an identified benchmark.2 min
Nov 12, 2026ModelsLaunchModelsAlibaba opens Qwen4-Max preview with a million-token context and a fixed end dateThe new model accepts text, images and video, with support for tool use. API access requires a subscription, and its reasoning shares an output budget with the final answer.3 min
Nov 11, 2026ModelsBenchmarkModelsSingle AI coordinators beat personal-agent teams in four simulations, researchers findThe study exposes a gap between better group scores and respecting each user’s wishes. Fixes worked differently for shared budgets and group bookings.3 min
Nov 11, 2026ModelsResearchModelsMIT researchers improve climate AI training with one optimized emissions scenarioThe method trains fast climate simulators on deliberately unusual emissions patterns. Results show better handling of unfamiliar scenarios, but full-scale Earth-system validation remains ahead.3 min
Nov 11, 2026ModelsResearchModelsAI Futures moves its forecast for superhuman AI researchers to 2027The forecasters are urging a development pause. Their biggest input is METR’s assessment of Anthropic, whose acceleration estimates carry substantial measurement and attribution uncertainty.4 min
Nov 11, 2026ModelsModelsStanford backs five projects linking AI to infant learning, therapy and brain biologyThe research will pair models with brain recordings, infant-viewpoint video and genetics. Better treatments and more human-inspired AI are goals, not delivered results.3 min
Nov 11, 2026ModelsLaunchModelsElevenLabs adds live speaker labels with its Scribe v3 transcription releaseThe release targets interrupted callers, crowded meetings and specialized vocabulary. Its accuracy claims rest on internal evaluations using different measures of transcription errors.3 min
Nov 11, 2026ModelsOpen releaseModelsRunway releases Praxis-1 robot-control model with open weights and commercial-use rightsThe release includes inference and fine-tuning code under Apache 2.0. Runway’s reported simulation results lead its model category, but four leaderboard entries score higher.3 min
Nov 11, 2026ModelsLaunchModelsRadical Numerics introduces offline AI to flag pathogens, with evaluation still under wayOmnii Defense predicts biological threats from genetic sequences. A company experiment suggests some ability to flag unfamiliar viruses, but a warning is not proof of danger—or engineering.3 min