OpenAI said on July 29, 2026, that its GPT-5.6 model family was designed to balance capability and cost across common model tasks by optimizing models, inference, and an agentic harness. The family includes GPT-5.6 Sol, a flagship model with maximum reasoning; Terra, which OpenAI says performs as well as GPT-5.5 on intelligence benchmarks at half the price; and Luna, described as the fastest and most affordable option, priced 80% less than Sol.

GPT-5.6 Sol outperforms Claude Fable 5 on the Artificial Analysis Coding Agent Index at less than half the cost, according to OpenAI. The company said GPT-5.6 reached its strongest intelligence-per-token efficiency by training models to complete more work per token, optimizing both task success and efficiency so the model takes a more direct path through tasks.

OpenAI said GPT-5.6 Sol improves inference, the process of running a trained model, by serving more tokens on the same hardware while preserving intelligence, latency, availability, and reliability. Reported gains came from routing, scheduling, kernels, caching, and model implementation. Load balancing routes requests by geography, capacity, and accelerator type, while cluster-level distribution uses load, context length, and cache availability. GPT-5.6 Sol in Codex analyzes production traffic to find imbalances and tune heuristics.

OpenAI said GPT-5.6 Sol autonomously rewrote and optimized production kernels, the GPU code that runs mathematical operations. The model was trained to write and improve kernels in Triton and Gluon, 2 open-source GPU programming languages maintained by OpenAI, and those efforts reduced end-to-end serving costs by 20%. OpenAI also uses FpSan, or Floating-Point Sanitizer, to validate kernel correctness.

Speculative decoding runs a smaller “speculator” model beside the primary model to propose and verify multiple tokens in parallel. OpenAI said GPT-5.6 Sol improved its own draft model by designing and running hundreds of experiments on its architecture, increasing token-generation efficiency by more than 15%. GPT-5.6 Sol in Codex also analyzes production workloads and generates optimized configurations for batching, sharding, and KV cache management.

OpenAI said ChatGPT Work and Codex rely on an agentic harness, a Rust orchestration layer connecting models, tools, and a user’s environment. The harness reduces repeated work through deferred discovery, where integrations, custom MCP tools, and plugins appear only when needed, and output capping, where tool output is capped at 10,000 tokens by default unless the model requests another limit. Prompt caching treats model-visible history as append-only, keeps tools in deterministic order, and applies runtime settings during execution rather than embedding them in tool definitions.

OpenAI said on July 29, 2026, that its GPT-5.6 model family was designed to balance capability and cost across common model tasks by optimizing models, inference, and an agentic harness. The family includes GPT-5.6 Sol, a flagship model with maximum reasoning; Terra, which OpenAI says performs as well as GPT-5.5 on intelligence benchmarks at half the price; and Luna, described as the fastest and most affordable option, priced 80% less than Sol.

GPT-5.6 Sol outperforms Claude Fable 5 on the Artificial Analysis Coding Agent Index at less than half the cost, according to OpenAI. The company said GPT-5.6 reached its strongest intelligence-per-token efficiency by training models to complete more work per token, optimizing both task success and efficiency so the model takes a more direct path through tasks.

OpenAI said GPT-5.6 Sol improves inference, the process of running a trained model, by serving more tokens on the same hardware while preserving intelligence, latency, availability, and reliability. Reported gains came from routing, scheduling, kernels, caching, and model implementation. Load balancing routes requests by geography, capacity, and accelerator type, while cluster-level distribution uses load, context length, and cache availability. GPT-5.6 Sol in Codex analyzes production traffic to find imbalances and tune heuristics.

OpenAI said GPT-5.6 Sol autonomously rewrote and optimized production kernels, the GPU code that runs mathematical operations. The model was trained to write and improve kernels in Triton and Gluon, 2 open-source GPU programming languages maintained by OpenAI, and those efforts reduced end-to-end serving costs by 20%. OpenAI also uses FpSan, or Floating-Point Sanitizer, to validate kernel correctness.

Speculative decoding runs a smaller “speculator” model beside the primary model to propose and verify multiple tokens in parallel. OpenAI said GPT-5.6 Sol improved its own draft model by designing and running hundreds of experiments on its architecture, increasing token-generation efficiency by more than 15%. GPT-5.6 Sol in Codex also analyzes production workloads and generates optimized configurations for batching, sharding, and KV cache management.

OpenAI said ChatGPT Work and Codex rely on an agentic harness, a Rust orchestration layer connecting models, tools, and a user’s environment. The harness reduces repeated work through deferred discovery, where integrations, custom MCP tools, and plugins appear only when needed, and output capping, where tool output is capped at 10,000 tokens by default unless the model requests another limit. Prompt caching treats model-visible history as append-only, keeps tools in deterministic order, and applies runtime settings during execution rather than embedding them in tool definitions.