| Date | 24 Sep 2026 |
|---|
Reasoning model is live on U.S. commercial enrollments that already have xAI turned on
Palantir Technologies said Thursday that Grok 4.7 from xAI is available in AIP.
The model is on U.S. georestricted and non-georestricted commercial enrollments with xAI enabled, according to a Foundry announcement dated Sept. 24.
xAI put Grok 4.7 out on Sept. 21. Palantir’s Language Model Service followed three days later. Same week, the company also listed GPT-6 Sol, GPT-6 Luna and GPT-6 Astra through Azure OpenAI and Direct OpenAI for eligible commercial enrollments, and DeepSeek V4.1 Flash through Fireworks on non-georestricted commercial accounts. Grok 4.7 is the item that still depends on an existing xAI enablement.
Palantir calls Grok 4.7 a reasoning model built for coding, agentic tasks and knowledge work. That is the whole product line in one sentence. Developers who already run agents, AIP Logic, Pipeline Builder tools or raw Language Model Service calls can point those workloads at the new ID once an administrator opens the family.
xAI’s own note is denser. The company said Grok 4.7 uses a larger base model than Grok 4.6 and was trained with a longer reinforcement learning run on a harder mix of tasks, weighted toward problems that take many hours to finish. It is better at checking its own work and holding longer context, xAI said. The company also trained the model to work natively with the Grok Bot harness.
List price on xAI’s API is $2 per million input tokens and $6 per million output tokens. That matches Grok 4.6. Cached input is listed at 50 cents per million tokens. The published context window is 500,000 tokens. Text and images go in. Text comes out. Reasoning effort can be set to low, medium, high or xhigh. High is the default.
Palantir does not bill those list rates raw. AIP converts tokens into compute-seconds, then into contract currency. Conversion factors differ by model and region. Anyone who signs invoices should pull the AIP Token Usage dataset and check the current Grok-family rates rather than assume the public API card applies one-for-one.
Access is not automatic. An enrollment administrator has to enable the model family in Control Panel, under AIP settings, on the Model enablement tab. That step includes accepting terms for the subprocessor that serves the model. If the family shows as disallowed because of legal or geographic limits, Palantir Support is the path. Grok models in AIP have been U.S.-scoped in Palantir’s availability tables.
Turn the family off later and anything pinned to it breaks. That is not a warning unique to Grok. It is how AIP model families work.
xAI posted its own scorecard. On CursorBench 4.0, which the company said stresses longer-running coding work, Grok 4.7 xHigh recorded 46.3%, against 40.4% for Grok 4.6 High. Terminal-Bench 4.0 moved from 20.3% to 37.6%. The Harvey Legal Agent Benchmark went from 15.8% to 19.6%. EEBench was listed at 64.0% for Grok 4.7, up from 53.0% on Grok 4.6. Those numbers are xAI’s. Palantir did not repeat them in the Foundry note.
The same xAI post said Grok 4.7 ships with a new safeguard stack and cited LatchBio’s biosafety benchmark at 62.4% and HackerBench v0.3 at a 3.3% pass-through rate on risky dual-use prompts. Again, that language is xAI’s, not Palantir’s.
For people who actually ship on Foundry, the practical sequence is short. Confirm xAI is enabled on the enrollment. Confirm the Grok family is switched on. Check token burn against the live compute table. Run the model against your own evals, not a vendor chart. Then decide whether it belongs in an agent loop, a document job, or a coding path.
Palantir asked customers to send feedback through Support or the Developer Community, using the language-model-service tag.
That is the release. A new model ID on a gated commercial surface. Administrators hold the switch. Developers hold the evals.