Outcome: On July 9, OpenAI removed the last waitlists. All three GPT-5.6 variants are fully open to every API tier, ChatGPT model picker, and Codex CLI channel.
On this page: Three rollout traps, a seven-row performance matrix, per-model highlights, six integration steps, citable July specs, and a MacPull path to benchmark every endpoint on isolated Apple Silicon before keys hit prod.
128K context
Voice + Realtime default
71.4% SWE-bench
Production default SKU
+4 pp alignment gain
Full-repo agent lane
Three traps when GPT-5.6 goes fully open on July 9
- Treating "fully open" as "pick Luna everywhere." Luna carries the highest input tier and slowest TTFT. Most chat, codegen, and eight-step agent loops run faster and cheaper on Terra—or Sol when latency is the bottleneck.
- Keeping July 1 Preview routing rules. Preview endpoints sunset for Tier-5 accounts on July 9. Hard-coded
gpt-5.6-previewstrings will 404. Migrate togpt-5.6-sol,gpt-5.6, andgpt-5.6-lunabefore CI pipelines fail silently. - Running three-model evals on your daily Mac. Full-open week invites parallel Sol latency tests, Luna full-repo dumps, and production vault access on one laptop. Security reviews replace launches when API keys collide on a shared keychain.
For the July 1 launch context, see our Sol vs Terra vs Luna comparison and OpenAI July 2026 roundup.
Performance highlights: what each GPT-5.6 variant does best
OpenAI did not ship three random SKUs. Each name maps to a compute profile inside the GPT-5.6 family—all sharing the July alignment fix, structured output GA, and parallel tool graph v2.
- Sol (
gpt-5.6-sol) — Optimized for time-to-first-token. Targets sub-180 ms TTFT on 4K prompts. Context capped at 128K. ChatGPT Advanced Voice and Realtime API v2 default here. Best for support bots, voice backends, and high-QPS chat surfaces. - Terra (
gpt-5.6/gpt-5.6-terra) — The GA default most teams should start with. 512K context, balanced reasoning, standard agent mode. ChatGPT Plus and Team tiers route general queries here unless you opt into Luna manually. - Luna (
gpt-5.6-luna) — The 1.5M-context lane. TTFT around 420 ms but highest instruction-following scores on full-repo and compliance workloads. Fully open to all API tiers from July 9—no Enterprise gate required.
Rule of thumb: Sol for speed, Terra for daily work, Luna only when the prompt truly needs more than 512K tokens or maximum alignment sensitivity.
Sol vs Terra vs Luna: July 9 full-open decision matrix
| Dimension | Sol | Terra | Luna | Pick when… |
|---|---|---|---|---|
| Access status (July 9) | All API tiers | All API tiers + ChatGPT default | All API tiers (waitlist removed) | No tier gating left |
| Context ceiling | 128K | 512K | 1.5M | Luna for full repos only |
| Typical TTFT | ~165 ms | ~275 ms | ~420 ms | Sol for voice and chat |
| SWE-bench Verified | 68.1% | 71.4% | 73.6% | Luna wins coding depth |
| Input / output (per 1M) | $1.80 / $7.20 | $2.50 / $10 | $4.00 / $16 | Terra for cost balance |
| Agent tool depth | Up to 8 steps | Up to 12 steps | Up to 16 steps | Match MCP graph size |
| July recommendation | Realtime and support bots | Staging + prod default | Audit and monorepo agents | Hybrid routing on M4 lab |
Quick verdict: Full open does not mean one model fits all. Start July 9 on Terra, add Sol for latency-sensitive surfaces, and gate Luna behind context allowlists plus spend caps.
Six steps to integrate fully open GPT-5.6 variants
- Audit Preview endpoint strings. Search repos for
gpt-5.6-previewand deprecated aliases. Replace with Sol, Terra, or Luna SKUs before July 9 cutover triggers 404s. - Tag workloads by context and latency. Under 32K plus low latency maps to Sol. Daily agents map to Terra. Full-repo or compliance jobs are Luna candidates only.
- Build a 50-case eval pack. Cover instruction following, eight- versus sixteen-step tool graphs, context boundaries at 128K, 512K, and 1.5M, plus alignment edge cases.
- Provision an isolated MacPull lab. Rent a Mac mini M4 (24 GB recommended) with separate API keys, OpenAI SDK 2.x, and routing scripts for all three endpoints—no production vault.
- Canary 5% traffic per variant. After July 9, replay identical agent traces on Sol, Terra, and Luna. Compare scores, bills, and p95 latency side by side.
- Lock routing rules and spend caps. Define automatic fallback Luna → Terra → Sol, daily token ceilings per SKU, and rollback to GPT-5.5 if any variant regresses.
Citable July specs (July 9, 2026)
- July 9, 2026 — GPT-5.6 fully open: Sol, Terra, and Luna available to all API tiers. Preview endpoints deprecated. ChatGPT model picker shows all three variants for Plus and Team subscribers.
- 128K / 512K / 1.5M — Context ceilings for Sol, Terra, and Luna respectively. Long-context surcharges apply above 128K on Terra and Luna only.
- 18% — Reported instruction-following gain vs GPT-5.5 on Terra at GA, with Luna adding approximately 4 percentage points on alignment-sensitive evals.
- $1.80–$4.00 — July input price range per 1M tokens across Sol to Luna. Budget Luna jobs separately from Terra defaults.
- 24 GB unified memory — MacPull recommended floor for running SDK eval scripts, Docker MCP sandboxes, and local MLX fallbacks on one M4 node while testing all three API variants.
Summary: full open means smart routing—not biggest SKU everywhere
July 9 closes the GPT-5.6 rollout arc. Sol, Terra, and Luna are no longer tier-gated experiments—they are production endpoints with distinct latency, cost, and reasoning profiles. Teams that route by workload adopt cleanly. Teams that default to Luna everywhere discover July invoices before they discover better answers.
The practical path: migrate off Preview strings, benchmark all three on isolated hardware, promote Terra to production, add Sol where latency matters, and gate Luna behind context allowlists plus spend caps.
Ready to validate Sol, Terra, and Luna before July routing goes wide? Rent MacPull Mac mini M4 (24 GB + 512 GB recommended for multi-variant agent sandboxes), compare plans on pricing, and follow SSH setup in help. Run side-by-side GPT-5.6 evals on dedicated Apple Silicon—then pick Sol, Terra, or Luna with data, not guesswork.
Benchmark fully open Sol, Terra & Luna on a Mac mini M4 lab
24 GB unified memory · three GPT-5.6 endpoints · SDK 2.x routing scripts · MLX local fallback · SSH + VNC · 24-hour delivery—validate every variant without touching your daily Mac.