Skip to content
Source · Daily Brief

AI Daily Brief — 28 June 2026

Chinese AI providers are successfully drawing enterprise workloads away from Western incumbents on price, and a new cybersecurity benchmark suggests capability gaps are narrowing too. The pricing stress test is real: at least one major US company has already halved its AI spend by routing to Chinese alternatives.

Top stories

  • Coinbase routes to Chinese models, halves AI spend. Coinbase CEO Brian Armstrong switched the company to an automated routing system that selects from models including GLM 5.2 and Kimi 2.7 based on task and price. Better caching pushed cache-hit rates from 5% to 60%, cutting AI spending in half even as token usage kept rising. via THE DECODER
  • Semgrep finds GLM 5.2 beats Claude in its cyber benchmarks. Security tooling firm Semgrep published results showing Zhipu AI’s GLM 5.2 outperforms Claude on its internal cybersecurity evaluation suite — fuelling the routing discussions above and a broader conversation about Chinese model competitiveness. via Semgrep
  • Grok 4.5 enters private beta at SpaceX and Tesla. Elon Musk posted that Grok 4.5 — built on a 1.5T-parameter V9 foundation model with supplemental training on Cursor data — is now in private beta at the two companies, with early evals claiming performance “close to, perhaps exceeding Opus.” The claim has not been independently verified. via X · @elonmusk
  • Google restricts Meta’s access to Gemini models. According to the Financial Times, Google has placed limits on how Meta can use its Gemini AI models — a rare instance of one major tech company capping another’s access to its AI infrastructure. via CNBC
  • Austria lobbies EU to host Anthropic amid US access curbs. Bloomberg reported that Austria is actively courting Anthropic to establish EU-based operations, following US export controls that have limited European customers’ access to Claude models. via Bloomberg
  • Ford rehires experienced engineers after AI falls short on quality. Ford is bringing back so-called “gray beard” engineers it previously let go, after AI-assisted processes failed to maintain manufacturing quality. “Mistakenly we thought that by just introducing artificial intelligence … that would produce a high-quality product,” a Ford executive was quoted as saying. via TechCrunch
  • ChatGPT logs submitted as evidence in Palisades fire arson trial. Prosecutors in the Jonathan Rinderknecht arson case presented his ChatGPT history alongside location data and security footage — establishing a legal precedent for AI conversation logs as admissible evidence in a US criminal trial. via The Verge
  • Liquid AI ships LFM2.5-230M for on-device inference. Liquid AI released the 230M-parameter LFM2.5 model with native support for llama.cpp, MLX, vLLM, SGLang, and ONNX — targeting edge and on-device deployments where smaller efficient models are preferred over frontier scale. via MarkTechPost

Who shipped

xAI put Grok 4.5 into private beta at SpaceX and Tesla with unverified claims of near-Opus capability. Liquid AI released LFM2.5-230M for edge deployment, with broad inference framework support. Suno launched its Spark incubator programme for independent artists, offering grants, mentorship, and marketing support — a move to position the company as a music discovery platform rather than a generation tool alone.

Open-source pulse

Liquid AI’s LFM2.5-230M landed with support for llama.cpp, MLX, vLLM, SGLang, and ONNX. llama.cpp pushed multiple release builds today (b9829 through b9835). Nathan Lambert’s Interconnects newsletter highlighted Zyphra and Poolside as examples of how the open-model ecosystem is broadening beyond the dominant Chinese and US labs.

Money, infra & hardware

Coinbase cut its AI spend in half by switching from premium Western models to cheaper Chinese alternatives — a concrete data point on the financial pressure facing Western providers at the enterprise level. Micron featured in TechCrunch analysis on why Wall Street views it as a potential infrastructure beneficiary of sustained inference demand.

Quiet corners

OpenAI and Anthropic made no announcements during the cycle; both appeared only in third-party contexts — Anthropic in a benchmark and in European regulatory discussions, OpenAI in Coinbase’s routing decisions. Meta was notable primarily as the subject of Google’s Gemini access restrictions.

By the numbers

  • 160 stories in 24 h across 20+ sources
  • Most-mentioned model: Claude (8 references)
  • Most-mentioned lab: GLM / xAI (tied, each with 3–4 direct mentions)
  • Notable absences: OpenAI, Anthropic, Meta (no primary announcements)

Compiled by AI Feed’s editor from all 160 headlines published on 28 June 2026.