OpenAI released GPT-5.6 on July 9, 2026 — a family of three models named Luna, Terra, and Sol, ordered least to most capable, available immediately across ChatGPT, Codex, and the OpenAI API. The same day, Microsoft made the series generally available in Microsoft Foundry. For teams running production workloads on OpenAI models, the headline is not any single benchmark. It is the tier ladder: every deployment now has to decide, workload by workload, which of three price points it actually needs.
Key takeaways
- Three tiers, one decision. GPT-5.6 ships as Luna, Terra, and Sol across ChatGPT, Codex, and the API, with same-day GA in Microsoft Foundry.
- Vendor numbers need verification. OpenAI says Sol scored 80 on the Artificial Analysis Coding Agent Index using less than half the output tokens of Anthropic's Fable 5. Treat that as a claim to test, not a fact to budget on.
- Pricing spreads 25× in Foundry. Luna input tokens list at $0.20 per million against Sol's $5.00 — tier routing is now a first-order cost lever.
- The Sol discount is temporary. Promotional pricing of $4 input / $20 output runs September 1 through November 30, 2026.
- Prices differ by platform. OpenAI's API and Microsoft Foundry list different rates for the same models. Price the platform you deploy on.
What OpenAI shipped in GPT-5.6
The family splits by capability and cost. Luna is the fastest, cheapest tier, aimed at high-volume, latency-sensitive workloads. Terra is the balanced middle — Microsoft positions it as competitive with GPT-5.5 at lower cost. Sol is the advanced reasoning tier.
The performance story is a vendor story, and we treat it that way. Per OpenAI, Sol scored 80 on the Artificial Analysis Coding Agent Index — 2.8 points above Anthropic's Fable 5 — while using less than half the output tokens, taking half the time, and costing one-third less. CEO Sam Altman put the token-efficiency claim at 54% on AI coding tasks. If those numbers hold on your workloads, they change cost-per-task math for coding and support automation. That is an if worth testing.
Two more pieces round out the launch. OpenAI calls Sol its "strongest cybersecurity model yet" for defensive work — threat modeling, code review, patching, and blue teaming. And ChatGPT Work, a new product for enterprise teams, covers document drafting, spreadsheets, and presentations across desktop, web, and mobile.
Luna, Terra, and Sol pricing compared
Launch pricing on the OpenAI API, per million tokens: Sol at $5 input / $30 output; Terra at $2.50 / $15; Luna at $1 / $6.
Microsoft Foundry lists different rates. Sol matches at $5.00 input / $30.00 output, with cached input at $0.50. But Terra drops to $2.00 / $12.00, and Luna to $0.20 / $1.20. That puts the Foundry spread between the cheapest and most capable tier at 25× on input tokens. It also means the same model carries two prices depending on where you deploy it — a detail that belongs in your cost model, not in a footnote.
There is also a clock on the top tier. Promotional Sol pricing of $4.00 input / $20.00 output runs September 1 through November 30, 2026. Any budget built on the promo rate needs a December line at full price.
GPT-5.6 in Microsoft Foundry
The Foundry release is broader than model availability. Deployment options span Global Standard and Priority Processing across 28 regions, Data Zones Standard, Global Provisioned, and a new Asia-Pacific Data Zone designed to keep data processing regionally anchored — Viva Republica (Toss) is quoted on that point.
Production agents in Foundry Agent Service also reached general availability: hosted agents, VNet integration, the Foundry IQ knowledge layer, Toolboxes, memory and routines, an agent optimizer, and an ROI dashboard. Microsoft cites Adobe, Telefonica, and Tata Consultancy Services as deploying production agents. For regulated and multinational operations, the data-zone and VNet pieces matter as much as the models: they decide where inference runs and what your compliance story looks like.
A benchmark score is a headline. Cost per accepted task is a budget line. Re-run your own numbers before you re-platform.
What operators should do now
- Re-benchmark against acceptance criteria. Run your existing evaluation set against each tier. The question is not which model scores highest — it is the cheapest tier that passes your acceptance criteria.
- Design for routing, not for one model. With a 25× spread, sending every request to the top tier is a design flaw. Route high-volume, low-stakes tasks to Luna and escalate only where Terra fails.
- Verify the token-efficiency claims yourself. If Sol's output-token numbers hold on your coding or support workloads, the effective cost gap to Terra narrows. Measure it; do not assume it.
- Budget past November 30. Model the Sol promo expiry now, and price on the platform you actually deploy on — the API and Foundry differ.
We looked at how reasoning tiers reshape workflow design in What reasoning models change for enterprise workflows, and our note on acceptance criteria, not demos covers how we benchmark before we build. If your team is re-pricing an OpenAI deployment against the new tier ladder, book a consult.