AI News

Center for AI Safety releases HLE-Diamond, a refined subset of Humanity's Last Exam

HLE-Diamond follows a year-long cleaning and refinement process that included input from research communities, and the release is with ScaleAILabs.

OpenWiki 0.6.0 adds shared workspaces and concurrency controls

Colin Francis announced OpenWiki 0.6.0, which groups wikis into shared workspaces with OpenWiki Link and adds concurrency controls for generating code wikis. It can also be used with antigravity and oh my pi.

Cursor says it cut token costs by 7% with no drop in agent quality

Cursor attributes the savings to tighter prompts, selective tool loading, better caching, and compressed file reads.

Qwen3.8-Flash is free in Qoder through September 30

The offer includes free accounts and does not require a claim or Credits. Qoder says it may continue the offer if people keep using the model.

CoreWeave Remote Key Encryption keeps keys in the customer's own key store

CoreWeave says Remote Key Encryption stores keys for encrypted model weights in the customer's own key store and that CoreWeave only holds ciphertext.

Artificial Analysis prices Gemini 3.8 Flash TTS at $32.98 per million characters

Artificial Analysis lists Gemini 3.8 Flash-Lite TTS at $22.07 per million characters. Both are above Gemini 3.1 Flash TTS and SpaceXAI TTS, and below Eleven v3.

Gemini 3.8 Flash TTS ranks first on the Artificial Analysis Pronunciation Robustness Benchmark

Artificial Analysis says Google has released Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, Google DeepMind text-to-speech models with preset voices and voice replication. On its tests, Gemini 3.8 Flash TTS scores 89.5% in pronunciation robustness and an Elo of 1,263 on the Provider Voice Arena.

Polymarket says GPT-6 Astra finished a real Toyota Corolla obstacle course

Polymarket says GPT-6 Astra was the only AI tested by DrivingBench to complete the course.

DeepLearning.AI offers free course on building on-device AI memory

Building AI Assistants with On-Device Memory covers a local system for text, voice, and images and is taught with Qdrant. The course includes a new AI coding lab, and enrollment is free.

Gemini 3.8 Flash and Flash-Lite text-to-speech models are available in the Gemini API

Google for Developers is offering Gemini 3.8 Flash TTS for natural-language voice direction and Gemini 3.8 Flash-Lite TTS for high-volume audio, both through the Gemini API in Google AI Studio.

Google AI starts rolling out Gemini 3.8 Flash TTS and Flash-Lite TTS

Beginning September 23, 2026, developers can use both models in Google AI Studio and the Gemini API. Gemini 3.8 Flash TTS is also reaching Gemini Notebook, Gemini 3.8 Flash-Lite TTS is reaching Google Vids, and both are coming soon to Gemini Enterprise.

Google AI launches Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS

Google AI is launching two speech models that can create custom voices in more than 100 languages or use more than 2,000 ready-made voices, with conversation direction and line-by-line delivery control.

Google AI Studio introduces Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS

The two audio generation models can be used through the Gemini API and in AI Studio. Google AI Studio calls them its most expressive audio generation models yet.

Google says Gemini 3.8 Flash TTS ranks first on Hume AI voice tests

Google reports a 71.4 overall score and a 60.8 accent-modeling score for Gemini 3.8 Flash TTS on Hume AI Voice Design. It also places Flash TTS first and Gemini 3.8 Flash-Lite TTS second on Hume AI Overall Quality.

OpenRouter takes Space Bunny Alpha temporarily offline

OpenRouter has removed Space Bunny Alpha while its provider works through an issue, and says the model will return once it is stable.

Self-organizing team of three models averages 66.7% in Stanford and Together AI paper

DAIR.AI reports that a Stanford and Together AI paper found o3-mini, Claude Sonnet 4, and DeepSeek-V3 averaged 66.7% across five math and physics benchmarks as a self-organizing team, above 48.8% for the strongest member alone and 59.0% for a perfect router.