Skip to content
Source · Daily Brief

AI Morning Brief — 27 June 2026

The US government spent the overnight hours reshaping who can access frontier AI: Anthropic’s Mythos 5 was partially reinstated for critical-infrastructure organisations, while OpenAI launched GPT-5.6 Sol under a government-controlled approval process that both labs described as unsustainable.

Top stories

  • GPT-5.6 Sol unveiled. OpenAI previewed its GPT-5.6 series — Sol (flagship), Terra (balanced), Luna (fast and affordable) — but restricted the rollout at the US government’s request to a small group of vetted partners. OpenAI said the approval-by-customer model should not become the long-term norm. via OpenAI
  • Mythos 5 reinstated, Fable 5 still blocked. After two weeks offline following a White House ultimatum, Anthropic’s Mythos 5 cybersecurity model is cleared for deployment at more than 100 US critical-infrastructure organisations; Fable 5, the company’s most capable model, remains unavailable. via The Verge · via Anthropic
  • METR’s pre-deployment evaluation of GPT-5.6 Sol published. METR released its independent pre-deployment assessment of Sol alongside the launch, covering autonomy, safety, and capability thresholds — the first such external evaluation released concurrently with a GPT-5 series model. via METR
  • SWE-bench Pro scores may be inflated. A Cursor study found that coding agents retrieve memorised fixes at test time rather than deriving solutions, contaminating SWE-bench Pro results and calling the benchmark’s validity into question. via MarkTechPost
  • OpenAI’s Jalapeño inference chip disclosed. OpenAI confirmed a custom inference chip built with Broadcom, joining Apple, Google, and SpaceX among large tech companies reducing dependence on Nvidia for AI compute. via TechCrunch
  • MAI-Code-1-Flash reaches general availability. Microsoft’s in-house coding model is now live for GitHub Copilot Business and Enterprise subscribers, expanding beyond its earlier limited preview. via GitHub Copilot changelog
  • Lindy drops Claude for DeepSeek. AI productivity startup Lindy ditched Claude entirely after AI costs exceeded personnel costs. CEO Flo Crivello described the switch as “a matter of survival for the business,” illustrating the cost pressure now bearing on Anthropic’s enterprise business. via The Decoder

Who shipped

OpenAI dominated overnight coverage with the GPT-5.6 Sol preview and the first public disclosure of its Jalapeño inference chip. Microsoft made MAI-Code-1-Flash generally available across Copilot tiers. Perplexity launched Computer for Counsel, a multi-model agentic layer for legal teams routing across Midpage, MCP connectors, and Microsoft 365. Anthropic regained partial ground with Mythos 5’s reinstatement, though Fable 5 remains inaccessible to most customers.

Open-source pulse

Hugging Face released TRL v1.7.0, with continuous batching making GRPO and RLOO training 1.25x faster at lower memory. Cohere open-sourced North-Mini-Code-1.0 under Apache 2.0 — a 4-bit quantised coding model that runs with 20 GB of RAM — positioning it as a locally-hostable alternative as frontier access tightens. llama.cpp shipped nine releases overnight (b9810–b9821), adding Vulkan tensor-parallel viability and updated OpenVINO and ROCm backends.

Money, infra & hardware

OpenAI is holding its IPO until 2027 — Sam Altman reportedly refuses to list below a $1 trillion valuation as tech markets remain volatile and SpaceX’s IPO underperformed. via The Decoder Micron locked in historically high memory prices under five-year contracts, a signal of sustained infrastructure demand.

Quiet corners

Google DeepMind, Meta, and Mistral posted no primary-source announcements overnight. xAI and Qwen/Alibaba were also absent from original coverage.

By the numbers

  • 252 stories in 24 h across 25+ sources
  • Most-mentioned model: GPT-5.6
  • Most-mentioned lab: OpenAI
  • Notable absences: Google DeepMind, Meta, Mistral

Compiled by AI Feed’s editor from all 252 headlines published on 27 June 2026.