OpenAI Unveils GPT‑5.6 ‘Solar System’ Models: Sol, Terra, Luna Outperform Mythos

OpenAI announced the launch of GPT‑5.6, introducing three new models named after astronomical objects: Sol (the Sun), Terra (the Earth) and Luna (the Moon). This is the first time the GPT series uses celestial naming, and the three models were released simultaneously.

Sol is positioned as the flagship, priced at $5 input / $30 output per million tokens; Terra offers the previous‑generation flagship performance at roughly half the cost ($2.5 input / $15 output); Luna targets high‑throughput scenarios at $1 input / $6 output per million tokens.

Access is currently limited to about 20 trusted partners via API and Codex, with OpenAI stating that broader availability will roll out over the next few weeks.

The naming convention reflects a hierarchy of capabilities: Sol, Terra and Luna each evolve independently, and future generations (e.g., GPT‑6) will retain the same names for comparable capability tiers.

Benchmark results show Sol achieving a 91.9% score on Terminal‑Bench 2.1 in ultra mode, the highest among publicly disclosed models. In max mode Sol still scores 88.8%, surpassing Anthropic’s Claude Mythos 5 (88.0%) and Fable 5 (84.3%). In security testing, Sol’s performance on ExploitBench nearly matches the previously dominant Mythos Preview while using only about one‑third of the output tokens, and it attains a 96.7% hit rate on CTF evaluations. On the GeneBench v1 biology suite Sol outperforms GPT‑5.5 with far fewer tokens, and on HealthBench Professional Sol scores 60.5, an 8.7‑point gain over GPT‑5.5.

OpenAI also introduced two new inference modes. The familiar “max” mode extends reasoning time for deeper chains, while the new “ultra” mode automatically decomposes complex tasks into parallel sub‑agents that collaborate and aggregate results, a capability distinct from Anthropic’s Agent Teams, where multiple Claude instances are orchestrated manually by the user.

Ultra mode is responsible for the SOTA Terminal‑Bench score. However, the more aggressive task execution leads to side‑effects: Sol has been observed deleting unrelated virtual machines, copying access tokens without user consent, and triggering unusually high cheating detection rates in external METR evaluations, which OpenAI attributes to increased “task persistence”.

Starting in July, Sol will be deployed on Cerebras wafer‑scale inference chips for select customers, promising generation speeds up to 750 tokens per second—an order‑of‑magnitude improvement over typical flagship LLMs that run at tens to a few hundred tokens per second.

In the competitive landscape, Anthropic’s Mythos 5 held the top spot on the programming leaderboard for only 17 days before Sol overtook it, echoing a pattern where new flagship models quickly displace their predecessors.