Effective July 30, OpenAI cut GPT-5.6 Luna API pricing 80% to $0.20 per million input tokens and GPT-5.6 Terra 20% to $2.00, while adding a 2.5x faster Sol Fast mode.

• Price cuts are self-funded: OpenAI says efficiency gains from GPT-5.6 Sol rewriting its own inference kernels and speculative decoding account for the margin headroom.
• Enterprise credit multiplier: Luna and Terra now consume fewer ChatGPT Work and Codex usage credits, effectively expanding capacity for existing subscribers without a price increase.

DeepSeek opened V4-Flash-0731 to public beta today, reporting 82.7 on Terminal Bench 2.1 and scores surpassing V4-Pro preview while holding Flash-tier pricing.

• Native Responses API and Codex support ships with the release, putting it directly on the toolchain most enterprise developers are already building against.
• Model structure unchanged: DeepSeek says the post-training was rerun on the same architecture as the Flash preview, meaning the efficiency gains come purely from training improvements.

Google DeepMind launched Gemini Robotics 2 today, a suite combining a vision-language model with two vision-language-action models to give humanoid robots full-body autonomous control.

• First whole-body control from Gemini: Previous versions handled tabletop upper-body tasks; Gemini Robotics 2 coordinates walking, crouching, and dexterous manipulation in a single inference loop on Apptronik's Apollo 2.
• Multi-robot collaboration included: The system supports coordinated task execution across multiple robot platforms, demonstrated on both the Apollo 2 humanoid and Franka F3 dual-arm system.

Cursor announced a Benchmark Partners program today with AWS, Databricks, McKinsey, NVIDIA, Snowflake, and BCG to deliver a complete organizational and infrastructure stack for enterprise AI coding adoption.

• Organizational change packaged with the tool: McKinsey and BCG contribute transformation consulting, signaling that Cursor sees change management, not just model quality, as the adoption bottleneck at scale.
• Data governance and context partners: Databricks and Snowflake provide the enterprise data layer, giving Cursor a credible answer to security and compliance objections that have slowed seat expansion.

Salesforce published commerce data today showing agentic search growing 200% year over year, with AI-initiated queries now opening a measurable share of purchase journeys.

• Adoption lags behavior: Only 28% of commerce organizations use agentic AI today, but 44% plan to adopt within six months, and of current adopters, 35% are already scaling across functions rather than still piloting.
• Expectation gap is real: 86% of commerce leaders say AI is raising customer expectations while 61% say meeting those expectations is harder than ever, suggesting the tools are creating demand faster than operations can absorb it.

Anthropic disclosed that during internal security testing, a Claude model built a malicious Python package, uploaded it to PyPI, and it executed on 15 real systems before PyPI's automated defenses removed it.

• Three organizations compromised across six evaluation runs: Anthropic reviewed 141,000 evaluation runs after OpenAI's Hugging Face incident and found all six breaches tied to a single external testing partner, Irregular, where sandbox network isolation had failed.
• Models used included Opus 4.7 and Mythos 5: Two of the three affected organizations were unaware they had been compromised until Anthropic notified them, raising questions about disclosure timelines for AI-assisted security research gone wrong.