GPT-5.6 went generally available on July 9, 2026, and unlike a normal point release it's a family of three models — Sol, Terra, and Luna — plus a batch of new developer primitives in the Responses API. If you're wiring it into an app or an agent, the first decision is which tier, and the second is which of the new features to actually use.
This is a working reference, not a launch recap. It covers what each tier is for, current pricing, and each new API capability with a one-line "why it matters." Illustrative snippets are marked as such — always confirm exact parameter names against the official OpenAI API docs, since betas move.
The three tiers
| Model | Built for | Input $/M | Output $/M |
|---|---|---|---|
| GPT-5.6 Sol | Frontier capability — hardest reasoning, agentic planning, Pro mode | $5.00 | $30.00 |
| GPT-5.6 Terra | Balance of intelligence and cost — the sensible default | $2.50 | $15.00 |
| GPT-5.6 Luna | Efficient, high-volume workloads — classification, extraction, routing | $1.00 | $6.00 |
Rule of thumb: prototype on Terra, escalate the genuinely hard calls to Sol, and push repetitive high-volume traffic down to Luna. Don't default everything to Sol — at $30/M output it's the fastest way to a surprising bill.
New Responses API features
Programmatic Tool Calling
The model emits tool calls your code executes and feeds back, tightening the agent loop instead of relying on free-form text the model hopes you'll parse. Cleaner control flow, fewer parsing hacks.
Multi-agent orchestration (beta)
In beta on the Responses API, GPT-5.6 can run concurrent subagents and synthesize their work within a single request — parallelizing a task (research + draft + critique) without you hand-rolling the fan-out/fan-in.
Max reasoning effort & Pro mode
An explicit reasoning-effort control lets you dial depth up for hard problems and down to save tokens on easy ones. Pro mode pushes the flagship to its highest-effort setting for the toughest tasks.
Explicit prompt caching controls
You can now control caching of repeated context explicitly — a real lever for cost and latency when every request shares a big system prompt or knowledge blob.
Persisted reasoning
Reasoning state can persist across turns, so multi-step agent sessions keep their chain of thought instead of restarting cold each call.
An illustrative request (confirm exact fields in the docs):
# Illustrative only — verify params at developers.openai.com
resp = client.responses.create(
model="gpt-5.6-sol",
reasoning={"effort": "max"}, # dial depth for hard tasks
input="Refactor this service and add regression tests",
# programmatic tool calling + prompt caching configured here
)Picking a tier fast
- Latency/cost-sensitive, simple task? Luna.
- General product feature or agent step? Terra.
- Hard reasoning, planning, or a task that failed on Terra? Sol (optionally Pro mode).
Migrating from GPT-5.5
- Map your current GPT-5.5 calls to Terra first — it targets that quality tier at lower cost — then selectively promote to Sol where evals show a gap.
- Audit any home-grown multi-agent or tool-dispatch code; the new orchestration and programmatic tool calling may let you delete glue.
- Add explicit prompt-cache directives anywhere you resend large shared context.
Availability & caveats
- GPT-5.6 rolled out after a customer-by-customer government review — access timing can vary by account.
- Multi-agent orchestration is beta; don't build a critical path on it without a fallback.
- Prices above are launch GA prices — re-check before modeling unit economics.
The through-line: GPT-5.6 turns "which model" into "which tier + which primitives." Spend five minutes matching tier to workload and you'll get most of the value — and avoid most of the bill.
