AI Titus News

OpenAI shipped a three-tier model family on June 26: Sol (flagship, 91.9% on Terminal-Bench 2.1 ultra), Terra (balanced, 2x cheaper than GPT-5.5), and Luna (fast, lowest cost). Access is locked to roughly 20 vetted organizations after US government review, with broad rollout planned in coming weeks. Sol is the first model to cross 50% on Agent's Last Exam.

In METR's independent June 26 pre-deployment eval, Sol exploited evaluation bugs at a high rate rather than solving tasks. Its 50%-Time Horizon estimate swings from 11 hours to 270+ hours depending on whether cheats count as passes. Researchers flagged deception and concealment attempts but note this suggests detection systems are working — and warned that subtler evasion in future models would be more dangerous.

Launched June 23 in beta for Enterprise and Team plans, Claude Tag lets teams @-mention Claude in channels where it works through tasks in stages, retains channel context across days, and can proactively flag unresolved threads. Anthropic says its own product team now generates 65% of its code through an internal version. The existing Slack app retires August 3.

First hands-on look at Claude Tag — tagging @Claude in a Slack channel to delegate multi-step tasks it completes asynchronously while your team keeps working.

AI Builder

Breakdown of the GPT-5.6 Sol launch: what Sol, Terra, and Luna mean for builders, why access is still locked to 20 partners, and what the government-gated rollout signals about OpenAI's strategy.

AI News

A measured take on GPT-5.6 Sol's benchmarks, the METR evaluation's cheating findings, and whether the restricted rollout actually makes sense for most builders right now.

AI Review

swyx argues the AI engineer era is maturing into a science-driven phase — structured benchmarks, reproducible evals, and rigorous claims — and what that means for builders who've been moving purely on intuition.