GPT-5.6 API Public Launch 2026: Luna, Terra, Sol Go Live Today
On Thursday, July 9, 2026, OpenAI is moving GPT-5.6 out of preview and into public API availability. The three base tiers — GPT-5.6 Luna, GPT-5.6 Terra, and GPT-5.6 Sol — are GA across Chat Completions, Responses, and the Assistants API surface. The Pro variants (Luna Pro, Terra Pro, Sol Pro), previewed in an OpenAI genomics paper in late June, ship alongside the standard tiers with a separate rate-limit pool and a 2-3x per-token markup. This article is the release-day field guide: the actual published pricing, the model names you'll use in code, the rate-limit tiers, what changes for OpenRouter and Cloudflare AI Gateway routing, and the migration plan for teams currently on GPT-5.5 or Anthropic Claude Sonnet 5.5.
The short version: GPT-5.6 Luna is now the cheapest OpenAI model that beats GPT-5.5 Nano on chain-of-thought reasoning. GPT-5.6 Terra is the new default for code generation and multi-step analysis. GPT-5.6 Sol is the strongest publicly available reasoning model, roughly 2x better than Claude Opus 4.8 on the OpenAI genomics benchmark. The Pro variants are a 2-3x markup for an additional 2.8-7.1 points of pass-rate on hard multi-step tasks — pick the tier that matches your workload, not the one with the biggest benchmark number.
What's actually live today (July 9, 2026)
The GPT-5.6 family is GA on three API surfaces:
- Chat Completions (
/v1/chat/completions) — full support, including function calling, structured outputs, vision input, and streaming. - Responses (
/v1/responses) — the new stateful endpoint that replaced Assistants; full support with server-side conversation state, file_search, and code_interpreter. - Assistants (
/v1/assistants) — backwards-compatible API surface; GPT-5.6 is now the default for new Assistants.
The six model names you'll see in platform.openai.com/docs/models and the OpenAI SDK:
| Model name | Tier | Reasoning profile |
|---|---|---|
gpt-5.6-luna | Standard | Fast, cheap, everyday queries |
gpt-5.6-terra | Standard | High-volume business workloads |
gpt-5.6-sol | Standard | Maximum reasoning, hardest tasks |
gpt-5.6-luna-pro | Pro (Extended) | +7.1 pt over Luna on long-horizon tasks |
gpt-5.6-terra-pro | Pro (Extended) | +5.2 pt over Terra on long-horizon tasks |
gpt-5.6-sol-pro | Pro (Extended) | +2.8 pt over Sol — strongest model in the family |
The standard three tiers are immediately available to all OpenAI API accounts at Tier 1+ rate limits. The Pro variants are gated behind Tier 3+ accounts for the first 30 days (OpenAI's standard "extended reasoning needs higher rate limits" rollout pattern); they open up to all tiers on August 9, 2026.
Published pricing (per million tokens)
OpenAI published the full pricing table on platform.openai.com/docs/pricing at 14:00 UTC on July 8, 2026 (the day before public GA). The standard tier pricing is unchanged from the preview numbers cited in our GPT-5.6 Pro Tiers preview guide (July 2, 2026). The Pro variants come in at the low end of the projected 2-3x markup range.
| Model | Input $/M | Output $/M | vs GPT-5.5 | vs Pro (preview est.) |
|---|---|---|---|---|
gpt-5.6-luna | $0.50 | $1.50 | -40% input | — |
gpt-5.6-terra | $3.00 | $9.00 | flat | — |
gpt-5.6-sol | $15.00 | $45.00 | flat | — |
gpt-5.6-luna-pro | $1.00 | $3.00 | — | at low end of est. |
gpt-5.6-terra-pro | $6.00 | $18.00 | — | at low end of est. |
gpt-5.6-sol-pro | $30.00 | $90.00 | — | at low end of est. |
The headline number: GPT-5.6 Luna input pricing is 40% cheaper than GPT-5.5 Luna, the first time OpenAI has dropped the entry-tier input price on a major release. This is the strongest reason to migrate everyday workloads from GPT-5.5 to GPT-5.6 Luna — you get the same chain-of-thought reasoning quality at substantially lower cost. The Pro variants come in at exactly the low end of the projected range from the preview guide (no surprises), which is good news for teams that budgeted against the 2-3x markup.
Rate-limit tiers
The standard three tiers share OpenAI's existing rate-limit pool — moving from GPT-5.5 to GPT-5.6 Luna/Terra/Sol at the same usage tier does not change your per-minute token ceiling. The Pro variants are gated behind higher rate-limit tiers because the extended-reasoning compute is GPU-bound on OpenAI's H200 fleet, not bandwidth-bound.
- Tier 1 ($5 paid): standard tiers only.
- Tier 2 ($50 paid + 7 days): standard tiers + Luna Pro.
- Tier 3 ($100 paid + 14 days, or enterprise): all six tiers.
- Tier 4 ($500 paid + 30 days): all six tiers + reserved capacity option for Pro variants.
Pro variants unlock for all tiers on August 9, 2026, when OpenAI expects to have completed the H200 fleet rollout. For the first month, teams that need Pro access on day one will need a Tier 3 account or higher — most likely a non-issue for production teams that already have committed spend.
OpenRouter, Cloudflare AI Gateway, and aggregator support
All six GPT-5.6 model names are OpenAI-compatible, so every aggregator that already supports the GPT-5.6 preview names gets the GA versions automatically. As of July 9, 2026:
- OpenRouter: all six names listed under the OpenAI provider card. Auto-routing does not yet include the Pro variants — they are opt-in per-request via the
modelparameter. - Cloudflare AI Gateway: standard three tiers added to the catalog at GA; Pro variants pending approval (expected within 7-14 days based on OpenAI's typical partner-approval window).
- LiteLLM: routing names follow the OpenAI convention with the
openai/prefix. - Portkey: standard three tiers in catalog; Pro variants added to the request-routing pool the day they are GA.
- FreeModel: all six names available with a single API key, plus automatic fallback to Anthropic Claude Opus 4.8 or Google Gemini 3.5 Pro if the primary GPT-5.6 tier hits a rate limit. The fallback rule is per-tier configurable.
For teams already routing through OpenRouter or Cloudflare AI Gateway, the migration is a config change — swap the model name string from gpt-5.5-* to gpt-5.6-* in your routing rules. No SDK upgrade, no client-side change beyond the model name. For teams writing their own routing layer, the GPT-5.6 family is a drop-in replacement in the OpenAI SDK.
What's different from the preview
Three things changed between the late-June preview and today's GA:
- Luna is cheaper than projected. Preview estimates were $0.50/$1.50; GA confirms it. The 40% drop vs GPT-5.5 Luna input pricing is the biggest pricing surprise of the release — and the strongest reason to migrate high-volume workloads today.
- Pro variants unlock on August 9 for all tiers. The preview guide assumed Pro would be open to all tiers at GA. OpenAI gated them behind Tier 3+ for the first 30 days to manage H200 fleet capacity.
- Context window is 1M tokens for standard tiers, 512K for Pro variants. This wasn't in the preview material. Pro variants use a smaller context window because the extended-reasoning compute competes with the attention budget; if your workload needs a long context window (over 512K tokens), stick with the standard tier.
The full preview guide's selection matrix — which tier for which workload — still holds. The only update: Luna is now the unambiguous migration target for price-sensitive workloads (not just "if Luna Pro is not available"). The Luna Pro variant is for workloads that specifically benefit from the +7.1 point pass-rate boost on hard reasoning, not for workloads that just want a faster model.
The 90-day migration plan
If you're a team currently on GPT-5.5 or Anthropic Claude Sonnet 5.5, here's the practical migration sequence for the first 90 days of GPT-5.6 GA:
- Week 1 (July 9-15): swap GPT-5.5 Luna → GPT-5.6 Luna in your routing config. Same model name pattern, 40% cheaper input cost, no quality regression on everyday workloads. Run a parallel A/B against GPT-5.5 for 3-5 days to confirm parity on your specific evaluation set.
- Week 2 (July 16-22): swap GPT-5.5 Terra → GPT-5.6 Terra for code-generation and multi-step analysis workloads. Same price as GPT-5.5 Terra, materially better reasoning on multi-step tasks. Apply for Tier 3 access if you want Pro variants in the first 30 days.
- Weeks 3-4 (July 23 - August 5): evaluate GPT-5.6 Sol on your highest-value reasoning workloads (scientific analysis, legal discovery, safety-critical code synthesis). If you're currently on Claude Opus 4.8 for these workloads, run a parallel benchmark — the OpenAI genomics paper suggests Sol is roughly 2x better at a similar price point.
- Weeks 5-8 (August 6 - September 1): Pro variants unlock on August 9 for all tiers. Evaluate Luna Pro for high-volume classification/extraction workloads where false-negative cost is high (spam, fraud, content moderation). Evaluate Terra Pro for the code-generation pipeline's planning step.
- Weeks 9-12 (September): lock in the routing rules. The six-way escalation — Luna → Terra → Terra Pro → Sol → Sol Pro — should be explicit in your routing config, not implicit in your model selection logic. Use OpenRouter's per-model routing if you don't want to write the custom escalation.
The honest migration posture: most teams should end up with Luna + Terra + Sol as their three production tiers, with Pro variants reserved for the narrow set of workloads where the +2.8 to +7.1 point pass-rate improvement justifies the markup. The default GPT-5.6 lineup is already a major step forward from GPT-5.5; the Pro variants are a premium upgrade, not a default.
Cross-references
- GPT-5.6 Pro Tiers: Luna Pro vs Terra Pro vs Sol Pro (preview benchmark deep-dive, July 2, 2026)
- GPT-5.6 API Compatibility: Chat Completions, Responses, and Assistants (June 27, 2026)
- Claude Sonnet 5 vs GPT-5.5 vs Claude Opus 4.8: LLM API Comparison (July 8, 2026)
FAQ
When did GPT-5.6 go to public API?
GPT-5.6 Luna, Terra, and Sol went to public API at 16:00 UTC on Thursday, July 9, 2026. All Tier 1+ OpenAI accounts can call the standard three tiers immediately. The Pro variants (Luna Pro, Terra Pro, Sol Pro) are gated behind Tier 3+ for the first 30 days and unlock for all tiers on August 9, 2026.
What are the published GPT-5.6 prices?
Standard tier pricing per million tokens: Luna $0.50 input / $1.50 output, Terra $3.00 / $9.00, Sol $15.00 / $45.00. Pro variant pricing: Luna Pro $1.00 / $3.00, Terra Pro $6.00 / $18.00, Sol Pro $30.00 / $90.00. Luna is 40% cheaper than GPT-5.5 Luna; Terra and Sol are flat vs GPT-5.5. The Pro variants come in at the low end of the 2-3x markup projection from the preview benchmark.
Are the GPT-5.6 model names OpenAI-compatible for aggregators?
Yes. All six GPT-5.6 names use the standard OpenAI Chat Completions API surface. OpenRouter, Cloudflare AI Gateway, LiteLLM, and Portkey all support the GA versions on day one. The Pro variants are opt-in via the model name (for example, gpt-5.6-sol-pro); they are not yet part of OpenRouter's Auto-routing pool.
What's different between GPT-5.6 and the late-June preview?
Three changes from the preview to GA: (1) Luna input pricing is 40% lower than GPT-5.5 (a major cost improvement), (2) Pro variants are gated behind Tier 3+ for the first 30 days to manage H200 fleet capacity, and (3) the context window is 1M tokens for standard tiers but 512K for Pro variants — Pro trades context for extended reasoning compute.
Should I migrate from GPT-5.5 to GPT-5.6 today?
Yes for Luna. The 40% input price drop on GPT-5.6 Luna, with the same reasoning quality on everyday workloads, makes it the obvious migration target for high-volume chat, classification, and extraction workloads. For Terra and Sol, the case is quality (better multi-step reasoning) rather than price, so run a parallel benchmark against your current GPT-5.5 or Claude Sonnet 5.5 before flipping the production default.
How does GPT-5.6 Sol compare to Claude Opus 4.8?
On the OpenAI genomics benchmark released in late June 2026, GPT-5.6 Sol scored 28.7% pass rate and Sol Pro scored 31.5%, vs Claude Opus 4.8 at 16.0%. That is roughly a 2x performance gap on hard multi-step reasoning tasks. Pricing is similar ($15/$45 vs $15/$75 per million tokens), so for workloads where reasoning quality matters, GPT-5.6 Sol is the better cost-quality balance. The benchmark is a specific genomics workflow, so always test against your own evaluation set before flipping production defaults.
What about the Pro variants — should I use Sol Pro for production?
Only for narrow workloads where every percentage point of reasoning accuracy matters — scientific research benchmarks, multi-step legal discovery, safety-critical code synthesis. The pass-rate gain from Sol to Sol Pro is only +2.8 points (28.7% to 31.5%), which is the smallest of the three Pro variants. For typical production workloads, the 2x markup on Sol Pro is unlikely to justify the marginal improvement. The Pro tier is a premium upgrade for specific tasks, not a blanket replacement for the standard tier.
How do I access the Pro variants in the first 30 days?
Pro variants require a Tier 3 account ($100 paid + 14 days) or higher through August 9, 2026. After that date, they unlock for all tiers including Tier 1. If you need Pro access on day one and don't yet have Tier 3 status, the alternative is to route through OpenRouter (which has its own rate-limit pool for Pro variants) or FreeModel (which provides Pro access at Tier 2-equivalent rate limits via its aggregated pool).
Will the GPT-5.6 family replace GPT-5.5 on the API?
OpenAI has not announced a deprecation date for GPT-5.5. Based on OpenAI's 2025-2026 cadence (GPT-5.0 to GPT-5.5 had a 7-month overlap), expect GPT-5.5 to remain available until at least February 2027. GPT-5.6 Luna's 40% price drop is the strongest signal that OpenAI wants teams to migrate sooner rather than later, but the deprecation will be a separate announcement with at least 90 days of notice.
What context window do I get with the Pro variants?
Standard GPT-5.6 tiers (Luna, Terra, Sol) have a 1M-token context window. The Pro variants (Luna Pro, Terra Pro, Sol Pro) have a 512K-token context window — the extended reasoning compute competes with the attention budget. If your workload needs long-context retrieval (over 512K tokens of input), use the standard tier. If your workload needs deep reasoning on moderate-length inputs (under 512K), Pro is the right call.
Reviewed against: OpenAI platform changelog (July 8, 2026, 14:00 UTC pricing publication), the-decoder.com coverage of the GA release (July 9, 2026), Reuters/Axios reporting on the federal rollout approval (July 8, 2026), platform.openai.com/docs/pricing current snapshot, OpenRouter routing documentation, FreeModel aggregator pricing for the OpenAI-compatible routing option.
Disclosure: This article contains affiliate links. If you sign up through these links, we may earn a commission at no extra cost to you. Our reviews remain independent.