GPT-5.6: Agent Tool Orchestration Meets Cost Controls

OpenAI’s GPT-5.6 family is now available in three sizes, with published pricing, shared context and output limits, and benchmark results focused on long-running agent workflows. The update also describes new developer-facing API capabilities for tool orchestration, multi-agent execution, and prompt caching control. For teams building agent systems, the changes center on how models coordinate external actions and manage costs across varying reasoning effort.

1. OpenAI releases GPT-5.6 family (Luna, Terra, Sol) in general availability

• GPT-5.6 is available in three sizes: Luna, Terra, and Sol.
• All three models list a February 16, 2026 knowledge cutoff, a one-million token context window, and a maximum output of 128,000 tokens.
• Pricing is per 1M input/output tokens: Luna $1/$6, Terra $2.50/$15, and Sol $5/$30.
• Impact: Developers can select the GPT-5.6 model size based on published token pricing while targeting shared context and output limits.

2. GPT-5.6 evaluation results emphasize long-running agent performance

• OpenAI’s “On Agents’ Last Exam” evaluation spans long-running professional workflows across 55 fields.
• GPT-5.6 Sol scores 53.6, compared with Claude Fable 5 (adaptive reasoning) by 13.1 points.
• GPT-5.6 Sol also scores 64.6% on SWE-Bench Pro, while Claude Fable 5 scores 80%.
• Impact: Developers can use the published evaluation figures to benchmark long-running agent workflows and coding performance tradeoffs.

3. New GPT-5.6 API capabilities for agent workflows and tool use

• Programmatic Tool Calling lets models “compose and run JavaScript that orchestrates tool calls.”
• Multi-agent support lets the model “spin up subagents for parallel, focused work.”
• Prompt cache breakpoints allow explicit cache breakpoint placement rather than relying on automatic detection.
• Impact: Developers can build agent orchestration patterns with JavaScript tool composition, parallel subagent workflows, and controllable prompt caching.

4. GPT-5.6 usage controls include image detail and explicit reasoning effort costs

• Image requests can set detail: original to avoid resizing before processing.
• Token and calculated costs are shown per model and reasoning effort, with the least expensive being gpt-5.6-luna at effort none for 0.71 cents.
• The most expensive listed case is gpt-5.6-sol at max reasoning level for 48.55 cents.
• Impact: Developers can reduce preprocessing steps for images and align expected token costs with selected reasoning effort levels.

Sources

• SimonWillison: https://simonwillison.net/2026/Jul/9/gpt-5-6/#atom-everything