AI

Pony.ai will supply the autonomy stack, Uber will integrate rides into its network, and local partners may own and operate fleets in four cities that remain unnamed.

In a 50-task run of Newsroom Reliability v0.2, SpaceXAI’s Grok 4.6 ranked first with a score of 0.79 at an estimated $0.0056 per task. OpenAI’s GPT-5.6 Sol followed at 0.77, while several GPT-5.6 variants and Anthropic’s Claude Opus 4.8 clustered close behind.

The model would sit alongside Qwen and Baidu technology in a China-only stack expected to reach Apple devices in the coming months.

Z.ai reports coding and vulnerability-exploitation gains, but its product documentation identifies GLM-5.1 as the latest release and provides no GLM-5.3 access path.

James Padolsey's guide explains how statistical text marks survive copying, while Declaude tests full rewrites against open watermark schemes.

Alibaba Cloud's June 30 Qwen Live recap describes an agent-first platform as Eddie Wu pushes paid model access, agent tooling and cloud adoption.

This matchup splits cleanly by strength: Kandinsky5 is better when the shot depends on believable physical evolution, while LTX 2.5 Image to Video Pro is stronger at staging busy multi-subject scenes and holding them together.

Nova beat social simulation tool Signal and brain-sensing speech prototype ThinkVoice in the 2026 competition.

The New York software maker is targeting service firms that need authentication, CRM data, permissions and payments built into generated apps.

Tulsee Doshi's team improved coding and workflow scores, though Google's full evaluation table shows gains were uneven.

The opt-in feature turns screen activity into reusable context for Pro, Business, and Enterprise users, with a timeline and privacy controls.

One model looked prettier in spots; the other actually did the assignment. Across all eight image-generation tests, Grok Imagine Image 2.0 was the more reliable, more obedient system—and the numbers leave no room for debate.

Cerebras hardware drives the API tier to 750 output tokens per second, with access restricted to selected customers while capacity grows.

Yan Junjie's model splits long-range composition from acoustic rendering and accepts detailed sectional controls, though its official weights remained unavailable at verification.

OpenAI's customer telemetry links heavier use to Plugins, reusable skills and Codex, offering startup operators a deployment playbook without proving those tools improve results.

Joy Wu's Shopify app creates machine-readable product pages, tracks AI-attributed orders and charges merchants when those sales happen.

The local workflow rewrites marked text, but Anthropic has not published the detector needed to independently verify the cleaner.

Jensen Huang says CUDA keeps older hardware useful as CoreWeave tries to extract revenue beyond its GPUs' initial contracts.

Bloomberg reports that Anthropic is weighing an unsigned deal for Decart's low-latency video, world-model and inference technology.

The live San Francisco presentation puts H3 before Magnific's claimed 1 million subscribers, though the event materials do not establish a product rollout.

The expected revision would end an exemption for open models once they reach the capabilities of the strongest closed systems.

One model kept landing the brief; the other kept drifting into adjacent, less specific scenes. Across all four tests, MiniMax H3 Reference to Video separated itself on prompt fidelity and multi-subject coherence, turning this into a decisive result rather than a stylistic debate.

The integration gives xAI distribution inside an established engineering workflow and strengthens Cognition's pitch that orchestration, not any single foundation model, is the durable product.

Product chief Mike Minton said "nobody would opt in," making creator inertia central to Amazon's access to a native multimodal dataset.