OpenAI on June 28, 2026 previewed GPT-5.6, a new three-model family comprising Sol (flagship), Terra (capable lower-cost option), and Luna (fastest and most efficient), with initial access limited to a small group of trusted partners at the US government’s request (OpenAI). All three models carry “High” designations in OpenAI’s Preparedness Framework for both Cybersecurity and Biological and Chemical risk categories - the classification level that triggered the staged rollout rather than a direct public launch (OpenAI Deployment Safety Hub). Sol introduces a new “ultra mode” that coordinates multiple subagents for complex tasks and sets a new state of the art on Terminal-Bench 2.1, a benchmark that tests command-line workflows involving planning, iteration, and tool coordination (OpenAI). API pricing for the family spans $5 per million input tokens and $30 per million output tokens for Sol, $2.50/$15 for Terra, and $1/$6 for Luna (VentureBeat).
The US Commerce Department on June 26, 2026 partially reversed its export control order against Anthropic’s Claude Mythos 5 model, granting access to approximately 100 US companies and federal agencies (CNBC). The original block, imposed in mid-June after an external party reported a technique to bypass Mythos’s safeguards, had prompted Anthropic to disable both Mythos 5 and companion Fable 5 for all users (Axios). Commerce Secretary Howard Lutnick confirmed in a letter to Anthropic that “appropriate safeguards are in place to permit certain trusted partners” to proceed, with the authorization extending to foreign nationals employed at approved organizations (Semafor). Most cleared entities are participants in Project Glasswing, Anthropic’s defensive cybersecurity program launched in April 2026; Fable 5 remains restricted, though parties familiar with the discussions indicated movement toward a similar resolution (CNBC).
OpenAI and Broadcom on June 24, 2026 unveiled Jalapeño, OpenAI’s first custom AI accelerator - a reticle-sized ASIC built specifically for LLM inference (TechCrunch). The chip advanced from initial design to manufacturing tape-out in nine months, a pace described as potentially the fastest for any high-performance ASIC to date, with OpenAI’s own models used to accelerate parts of the design and optimization workflow (Tom’s Hardware). Engineering samples are running GPT-5.3-Codex-Spark workloads at production target frequency and power, with performance per watt reported as substantially better than current hardware (OpenAI). Initial production deployment is targeted by the end of 2026, with Broadcom providing networking and connectivity integration alongside the processor (Broadcom).