Large Language Models

Today's Stories

1. 1

   Progressive Disclosure: From Training Wheels to Week-Long AI Agents

   Progressive Disclosure: From Training Wheels to Week-Long AI Agents

2. 2

   Mistral Robotics model puts France's industrial AI push in focus

   Mistral Robotics model puts France's industrial AI push in focus

3. 3

   OpenAI launches GPT-5.6 in three sizes, beats Claude on agent benchmarks

   OpenAI released three models of its GPT-5.6 family—Luna, Terra, and Sol (smallest to largest)—with general availability starting today. On the Agents' Last Exam benchmark, GPT-5.6 Sol scored 53.6, beating Claude Fable 5's 53.6-point lead by 13.1 points, while Luna and Terra outperformed Fable 5 at around one-sixteenth the estimated cost. The new models deliver agentic performance (long-running professional workflows) more efficiently than existing competitors. Smaller models like Terra and Luna achieving Fable 5-level results at a fraction of the cost may appeal to businesses looking to deploy AI agents in real-world applications without prohibitive spending.

   Pricing is Luna $1/$6, Terra $2.50/$15, and Sol $5/$30 (per 1M input/output tokens). New API features include Programmatic Tool Calling (for orchestrating tool calls), Multi-agent (spinning up subagents in parallel), and Prompt cache breakpoints (explicit cache control). OpenAI also published findings that approximately 30% of SWE-bench Pro tasks are broken, after Claude Fable 5 outperformed GPT-5.6 Sol on that benchmark (80% vs. 64.6%).

4. 4

   Announcing OpenAI GPT 5.6 on Snowflake Cortex AI

   Announcing OpenAI GPT 5.6 on Snowflake Cortex AI

5. 5

   OpenAI may have made a fatal misstep in copyright fight with news orgs

   OpenAI may have made a fatal misstep in copyright fight with news orgs

6. 6

   69% of enterprises expose AI agents via shared API keys

   VentureBeat's June 2026 survey of 107 enterprises found that 69% run AI agents with shared API credentials in their deployments. When one API key is shared across multiple agents, a compromise of any single agent grants an attacker access to all workflows the key touches, and audit trails cannot show which agent performed which action. This credential-sharing vulnerability represents a major security gap that enterprise vendors are now racing to address. Palo Alto Networks, CrowdStrike, and Cisco have collectively bet more than $22 billion（約3.5兆円） on securing this exact layer in the past year, signaling that enterprises recognize the risk but have not yet built adequate defenses.

   Palo Alto Networks completed its acquisition of CyberArk on February 11 for $21.1 billion（約3.4兆円） in total consideration—the largest deal in this buying spree—announced last July at roughly $25 billion（約4兆円）. This consolidation may reshape how enterprises manage credentials across AI workloads.