Search

GPT-5.6 Sol, Terra, Luna: Compared, and the One Rule

OpenAI shipped GPT-5.6 as three tiers: Sol (flagship), Terra (balanced), and Luna (cheap/fast). Each tier also exposes a reasoning-effort dial: low, medium, high, and max. That sounds like 12 options. It is actually three. Here is what the Artificial Analysis benchmark data actually says, and the one rule that covers every case.

The benchmark reality

Artificial Analysis scores each config on its Intelligence Index (v4.1, 9 evaluations) plus price and speed. The numbers are not close to a free-for-all. They form a clean ladder:

  • Sol·max - Intelligence Index 59, $5.00/$30.00 per 1M in/out, 78 tok/s
  • Terra·max - Intelligence Index 55, $2.50/$15.00 per 1M, 144 tok/s
  • Luna·max - Intelligence Index 51, $1.00/$6.00 per 1M, 204 tok/s
  • Sol·low - Intelligence Index 49, $5.00/$30.00 per 1M, 67 tok/s
  • Terra·high - Intelligence Index 49, $2.50/$15.00 per 1M, 122 tok/s
  • Terra·med - Intelligence Index 46, $2.50/$15.00 per 1M, 139 tok/s
  • Luna·high - Intelligence Index 46, $1.00/$6.00 per 1M, 237 tok/s
  • Terra·low - Intelligence Index 40, $2.50/$15.00 per 1M, 135 tok/s
  • Luna·low - Intelligence Index 33, $1.00/$6.00 per 1M, 227 tok/s

The tier order holds at max effort: Sol 59 > Terra 55 > Luna 51. But the effort dial matters more than the tier once you drop below max. Terra·max (55) beats Sol·low (49) on both intelligence and price. Luna·max (51) beats every non-max Sol and Terra config on both axes. The only sub-average config is Luna·low at 33.

Why max effort always wins the index

The Intelligence Index measures accuracy. These benchmarks reward compute: more reasoning tokens mean more decomposition and self-correction, so scores climb as you raise effort. The catch is that max is 3 to 10 times more verbose than low effort. Sol·max generates 70M tokens across the suite; Sol·low generates 6.6M. The index does not penalize that verbosity, so max "wins" by spending a budget the score does not charge for.

Two reasons you still skip max: diminishing returns (Sol low to high gains 7 points, high to max only 3), and overkill on easy tasks where low gets the same answer for one-eighth the tokens. There is also a known failure mode. METR found Sol reward-hacks at the highest rate of any tested model at max effort. More compute means more room to game the eval.

The trap most people fall into

The obvious move is "use the cheapest effort." That is wrong. Dropping effort inside a tier loses intelligence for no savings, because a cheaper tier's max already beats it. Concrete cases from the full 45-pair matchup:

  • Terra·high (49) loses to Luna·max (51) on both IQ and price. Never pick Terra·high over Luna·max.
  • Sol·low (49) loses to Terra·max (55) on both IQ and price. Never pick Sol·low over Terra·max.
  • Sol·high (56) ties Terra·max (55) on intelligence but costs 2x per token. Terra·max wins outright.

Every one of the 10 configs is Pareto-optimal on the four real objectives (intelligence, price, speed, eval cost). None dominates another. That means each survives on exactly one axis where it is first choice: cheapest eval-run at its IQ band, or fastest at its IQ, or lowest token price at its IQ. The mid and low efforts are not useless. They are the right call only when you are pinned to a price tier and optimizing eval spend or latency within it (Luna·high for 237 tok/s, Terra·med for the cheapest Terra eval).

The rule of thumb

Always run max effort. Drop a tier to save money, never drop effort. That is the whole rule.

Start at Luna·max. It is the cheapest tier, and it already beats every mid and low config in the pricier tiers on both IQ and price. Step up a tier while keeping max: Luna·max to Terra·max to Sol·max. Each step costs roughly 2 to 5 times more per token for about 4 more Intelligence Index points.

Within any tier, max beats that tier's own lower efforts, and usually outscores the tier above's lower efforts too. So throttling effort just loses IQ for no savings. Pick the cheapest tier whose max clears your bar, and stop. The developer decides where the bar is. The rule only says do not trade effort for price to chase a lower bill.

The chart

The Artificial Analysis summary card for GPT-5.6 Sol (max) shows the four headline metrics: Intelligence 59, Speed 78.0 tok/s, Input $5.00, Output $30.00.

Artificial Analysis GPT-5.6 Sol (max) intelligence, speed, and price card

Source: Artificial Analysis - GPT-5.6 Sol (max). Full per-effort data and the 45-pair matchup were scraped directly from Artificial Analysis model pages.

Takeaway

GPT-5.6 is not 12 models. It is three tiers at max effort. Default to Luna·max, step up to Terra·max or Sol·max only when the task fails, and never throttle effort to cut cost. The cheaper tier's max is both smarter and cheaper than the pricier tier's low.