2026-07-09

1. GPT-5.6 Sol, Terra, and Luna — General Availability (today)

• OpenAI flipped the switch July 9: all three GPT-5.6 tiers are now publicly available globally via API and ChatGPT subscriptions, following DoC/CASI clearance. (source)
• Pricing unchanged from preview: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per Mtok. New at GA: Sol Fast (~750 tok/s, $12.50/$75) for latency-critical agentic pipelines.
• New caching model: explicit cache breakpoints (developer-controlled), 30-min minimum cache lifetime guaranteed, cache writes billed at 1.25× uncached input rate, reads still 90% off.
• Why it matters: GPT-5.6 Sol is the first model to complete the full White House AI EO voluntary pre-release government review cycle — June 26 limited preview → CASI testing → July 9 public clearance. Every future US frontier model launch will follow this template. The 13-day government window appears to be the new release norm.
• → GPT-5.6 Sol (and Terra, Luna)

2. Grok 4.5 — Public Launch

• xAI launched Grok 4.5 publicly July 8/9 (Musk announced July 8). Available in Grok Build (now default), Cursor (all plans), SpaceXAI console. Not yet in EU. (source)
• Pricing: $2/$6 per Mtok, ~80 tok/s. Built on V9 1.5T + Cursor training (SpaceX acquired Cursor ~$60B June 2026).
• Published benchmarks vs Claude Opus 4.8: DeepSWE 1.0 62% vs 55.75% ✓, Terminal Bench 2.1 83.3% vs 78.9% ✓, DeepSWE 1.1 53% vs 59% ✗, SWE-Bench Pro 64.7% vs 69.2% ✗. Token efficiency: 4.2× fewer output tokens per task than Opus 4.8.
• Why it matters: "Opus-class at Sol-class price" ($2/$6 vs $5/$30 for Sol) — Grok 4.5 is the lowest-cost frontier-tier coding model currently available. Combined with the Cursor integration across all plans (not just premium), this directly threatens Claude Sonnet 5's market position in developer workflows. If xAI's token-efficiency claim holds under real-world load, the ROI advantage is substantial.
• → Grok 4.5

3. Mistral Robostral Navigate — SOTA Robotics Navigation, Single Camera

• Mistral released Robostral Navigate (July 8): 8B robotics navigation model requiring only a single RGB camera (no depth sensors, no LiDAR). Sim-only training. (source)
• Benchmark: 76.6% R2R-CE validation unseen — beats prior best single-camera system by +9.7 pts and prior best depth/multi-camera system by +4.5 pts.
• Method: pointing-based prediction (outputs image coordinates of target + arrival orientation); online RL with CISPO (+3.2% success rate). Hardware-agnostic across robot platforms.
• Why it matters: a sim-trained 8B model beating depth-sensor approaches means the hardware bar for robot navigation deployment just dropped to "any robot with an RGB camera." This is Mistral's first physical AI product and expands their axis beyond language/coding into industrial AI — consistent with the Emmi AI (physics simulation) acquisition in May. European industrial customers now have an open-friendly path to robotics without NVIDIA or DeepMind.
• → Robostral Navigate