On July 7, 2026, OpenAI released GPT-5.6 not as a single flagship but as a family of three models differentiated by capability tier: gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna. In the new naming scheme the number (5.9) marks the generation while the tier names — Sol, Terra, and Luna — replace the base/mini/nano suffixes of earlier families. Sol is the frontier-capability model targeted at hard reasoning, coding, cybersecurity, and long-horizon agentic workloads; Terra is the balanced intelligence/cost tier for premium production traffic; and Luna is the efficient, high-volume option. All three are available at launch through the /v1/responses, /v1/chat/completions, and /v1/batch endpoints.

The flagship carries a documented 1,113,000-token context window and up to 135,680 tokens of maximum output, and the tiers share the same context and maximum output length rather than shrinking the window at the cheaper end. Images are now accepted at their original dimensions using an original or auto detail setting, removing the prior downscaling step that cost fidelity on documents, charts, and UI screenshots fed into the model.

Pricing is tiered to match capability. GPT-5.6 Sol is priced at $5.30 per million input tokens and $31.80 per million output tokens, matching the outgoing GPT-5.5 flagship rate. Terra comes in at $2.65 input and $15.90 output per million tokens — roughly half the flagship cost for performance OpenAI positions as competitive with GPT-5.5 — and Luna sits at $1.06 input and $6.36 output per million tokens for high-volume deployments where latency and unit economics dominate the model choice.

GPT-5.6 also reworks prompt-caching economics. Cached reads retain their 95% discount versus uncached input, but explicit cache writes are now billed at 1.25x the normal uncached input rate, with a 30-minute minimum cache retention window. The release pairs this with a revised caching interface built around explicit breakpoints, giving developers direct control over which long, repeated prefixes — system prompts, tool schemas, retrieved context — get persisted rather than relying purely on implicit prefix matching.

Several new developer-facing controls arrive with the family. Programmatic tool calling lets the model orchestrate tools more directly, and a max reasoning effort setting exposes deeper problem analysis as a tunable dial, with reasoning state now persisted across turns so an agent does not discard its intermediate chain of thought between steps. A parallel-subagent capability referred to as Ultra mode lets Sol spin up coordinated subagents on a single task; OpenAI reports Ultra mode lifts Sol's score on a command-line coding benchmark from 94.1% to 97.4%.

On comparative benchmarks, OpenAI claims Sol beats the competing Claude Fable 5 model by 13.9 points on the Agents' Last Exam evaluation, while acknowledging that Fable 5 still leads Sol on SWE-Bench Pro (85% versus 68.5%). OpenAI presents these as self-reported figures, and independent evaluators including METR have cautioned about record-high levels of benchmark gaming across the industry, so the numbers are best read as directional rather than definitive.

The GPT-5.6 launch followed a limited preview that began June 24, 2026, and OpenAI said the broad public rollout proceeded after additional safety testing and review with U.S. government agencies. The company rolled the models out globally within minutes of the launch livestream, and the release notes list GPT-5.6 as the recommended default for new builds that need frontier reasoning, with Terra and Luna positioned as the cost-down options for teams scaling agent workloads into production.

For developers already on GPT-5.5 or the GPT-5.4 line, migration is largely a model-string swap given the shared Responses API surface, but the caching changes are the detail to watch: teams that lean heavily on large repeated prefixes should audit their cache-write volume, since the new 1.25x write multiplier and 30-minute minimum retention change the break-even math on when caching actually saves money versus paying full uncached input rates.