GPT-5.6 Goes GA: Programmatic Tool Calling Changes Everything

#ai #agents #programming #openai

GPT-5.6 went GA with three tiers (Sol, Terra, Luna) and a new capability that matters more than the benchmarks: Programmatic Tool Calling.

Agents can now write and execute lightweight programs between tool calls — filtering data, coordinating tools, monitoring progress — without round-tripping intermediate results through the context window. Fewer tokens, fewer model calls, faster task completion.

Production receipts: Ploy.ai migrated from Claude Opus 4.8 to GPT-5.6 Sol. Results: $2.22 per build vs $3.06. Half the wall-clock time.

Also in this issue:

Anthropic's J-Space: Claude has a silent internal workspace for reasoning it never writes down. Used for deception detection.
CubeSandbox: Tencent open-sourced hardware-isolated KVM sandboxes (60ms boot, <5MB overhead, E2B compatible)
OpenAI audited SWE-Bench Pro: ~30% of tasks are broken
OfficeCLI: Office suite built for AI agents (15.7K stars)
GRAM: Anthropic's modular off-switch for dangerous knowledge
Loom for AWS: Enterprise reference architecture for Strands + AgentCore
Orca: Parallel agent IDE (17.6K stars)