Skip to content
Source · Daily Brief

AI Daily Brief — 30 June 2026

Anthropic’s release of Claude Sonnet 5 dominated Monday’s cycle, claiming near-Opus performance at Sonnet pricing; simultaneously, Meituan confirmed that a consumer internet company can train a near-frontier 1.6 trillion-parameter model entirely on Chinese-made chips and ship it publicly.

Top stories

  • Claude Sonnet 5 ships. Anthropic released its most capable Sonnet model, with benchmarks placing it close to Opus 4.8 on knowledge and agentic work. Available from day one on the API, Amazon Bedrock, GitHub Copilot, Cursor (which reported a jump from 49% to 57% on its internal benchmark), and Perplexity. via TechCrunch
  • Claude Science launches. In parallel, Anthropic announced a researcher-focused workbench with more than 60 preconfigured skills spanning genomics, computational chemistry, and related fields. A verification agent automatically checks citations and calculations; the app runs locally or on HPC clusters, keeping sensitive data on-premises. via TechCrunch
  • Meituan’s LongCat-2.0 goes public. The Chinese food-delivery giant released a 1.6 trillion-parameter MoE model with roughly 48 billion active parameters, a one-million-token context window, and training entirely on domestic Huawei Ascend ASICs — no Nvidia hardware. The model is live on OpenRouter. via THE DECODER
  • Google ships two generative media models. Nano Banana 2 Lite generates images in under four seconds at $0.034 each; Gemini Omni Flash brings video generation and editing to the Gemini API for the first time at $0.10 per second. Google suggests chaining both for rapid image-to-video workflows. via THE DECODER
  • Etched exits stealth. The inference-chip startup announced $800 million raised, more than $1 billion in customer contracts after a successful first tapeout, and a $5 billion valuation. First racks are expected to ship this summer, with claimed state-of-the-art throughput and power efficiency on inference workloads. via TechCrunch
  • OpenAI halves inference costs. Engineers applied an optimisation that cut inference costs by more than half on models where it was deployed. For logged-out ChatGPT traffic, peak GPU requirements dropped to a few hundred, down from significantly higher levels before the change. via THE DECODER

Who shipped

Anthropic had its largest simultaneous launch in some time, releasing both Claude Sonnet 5 and Claude Science to the public on the same day. Google DeepMind shipped Nano Banana 2 Lite and Gemini Omni Flash, expanding its generative media API surface. Meituan made LongCat-2.0 publicly accessible on OpenRouter, while DeepSeek published DSpark, a speculative-decoding framework claimed to boost per-user throughput by 60–85 percent on existing hardware, with community integration into SGLang reported within hours.

Open-source pulse

Huawei open-sourced OpenPangu-2.0-Flash (92B total parameters, 6B active, 512K context, Ascend-trained), with a 505B Pro variant promised for July. Meituan LongCat-2.0 is now accessible on OpenRouter under open weights. DeepSeek DSpark was released with public code; its speculative-decoding approach is reported to reduce time-per-output-token to 2.9–5.2 ms on supported inference stacks.

Money, infra & hardware

Etched reported a $5 billion valuation, $800 million raised, and $1 billion-plus in customer commitments as it exited stealth. Amazon announced a $1 billion field deployment engineering organisation to embed purpose-built agents directly at enterprise customers. Moonshot AI (Kimi) reportedly reached $300 million ARR and is raising at a $31.5 billion valuation. Taiwanese authorities raided Super Micro offices in a probe over suspected Nvidia chip smuggling to China.

Quiet corners

Mistral, Cohere, and Stability AI were absent from the day’s primary news. Meta appeared mainly in a report about its contractors covertly probing competitors’ chatbots with minor-perspective crisis prompts — a story about testing practice, not a product launch.

By the numbers

  • 330+ stories today across 40+ sources
  • Most-mentioned model: Claude
  • Most-mentioned lab: Anthropic
  • Notable absences: Mistral, Cohere, Stability AI

Compiled by AI Feed’s editor from all 330+ headlines published on 30 June 2026.