The models opened Thursday on eligible commercial enrollments through Azure OpenAI and Direct OpenAI.
Palantir put OpenAI’s GPT-6 Sol and GPT-6 Luna into its Artificial Intelligence Platform on Thursday, Sept. 24. The path is Azure OpenAI and Direct OpenAI. Eligible commercial enrollments only.
The company notice is brief. Two models. Two connectors. Same day.
GPT-6 Sol is designed for complex coding and agentic workflows, the Palantir announcement said. OpenAI’s model page uses nearly the same line: “GPT-6 Sol is built for complex coding and agentic workflows.”
GPT-6 Luna is designed for efficient, focused, high-volume tasks. OpenAI’s wording is sharper: “GPT-6 Luna is our most efficient model for focused, high-volume tasks.”
Both accept text in and out. Image input works. Audio and video do not. Each carries a 1,050,000-token context window and a 128,000-token output cap. Sol’s knowledge cutoff is April 20, 2026. Luna’s is May 18, 2026.
Reasoning effort can be none, low, medium, high, xhigh, or max. Medium is the default. Streaming works. Function calling works. Structured outputs work. Fine-tuning does not.
The built-in tool set — web search, file search, code interpreter, hosted shell, computer use, MCP, and the rest — sits on the Responses API. Chat Completions will still do function calling. Only if reasoning_effort is set to none.
None of that shows up in a Foundry workflow until an enrollment administrator turns the family on.
That switch lives in Control Panel, on the Model enablement tab inside AIP settings. Turning a family on means accepting the provider terms, subprocessors included. States are Enabled, Disabled, or Disallowed. Disallowed usually tracks legal limits, geography, or infrastructure. Palantir Support is the route if a family is blocked. Disable a family and anything wired to it breaks. Organization-level toggles exist as well.
Cost is the next stop, and it is not OpenAI’s public card by itself. AIP turns tokens into compute-seconds. Rates move with the model, the region, and the contract. North America is one band. The EU and the U.K. another. South America, APAC, and the Middle East a third. Palantir tells customers to confirm the terms with their representative. Usage can be pulled from an exported AIP Token Usage dataset, broken out by model and resource.
OpenAI still publishes list prices that matter for raw model math. Sol: $2 per million input tokens, 20 cents cached input, $2.50 cache writes, $10 output. Luna undercuts that by a wide gap: 10 cents input, 1 cent cached input, 12.5 cents cache writes, 50 cents output. Prompts above 272,000 input tokens draw higher multipliers. Regional processing can add 10 percent. Those figures are OpenAI’s. They are not Palantir’s invoice.
Thursday was a crowded model day. GPT-6 Astra opened on the same Azure and Direct OpenAI path for eligible commercial enrollments. Palantir described Astra as OpenAI’s most capable model for complex reasoning, software development, research, and document creation. xAI’s Grok 4.7 landed on U.S. georestricted and non-georestricted commercial enrollments that already have xAI enabled. Palantir called it a reasoning model designed for coding, agentic tasks, and knowledge work. DeepSeek V4.1 Flash arrived through Fireworks on non-georestricted commercial enrollments.
What an enrollment actually sees still hinges on geography, legal acknowledgments, and whether AIP is on for the right user groups. Palantir keeps the live inventory on its supported-LLMs page.
Feedback was requested in one sentence, left as written: “We want to hear about your experiences using language models in the Palantir platform and welcome your feedback.” Support channels are open. So is the Developer Community, under the language-model-service tag.
For executives, the live questions are narrower than the model cards. Who is allowed to call them. What the compute-second burn will look like at volume. Whether Sol’s heavier coding and agent stack is worth the token rate, or whether Luna’s cheaper high-volume pass is the better fit. For developers, the work starts after the admin flip: pick the API surface, set reasoning effort, and watch the usage export.
That is the release. It is live if the enrollment qualifies and the family is on.