GPT-5.6 Sol vs Terra vs Luna: Which Tier to Use

OpenAI’s GPT-5.6 family is a named tier ladder, not a single model with a slider. Marketing tables blur the gaps; operators feel them in the bill and in long-context recall. This post is the comparison sibling to the Luna complete guide — which tier for which job, with honest numbers.

Quick comparison table

Official list prices after the July 30, 2026 cut (confirm on OpenAI’s pricing page). Terminal-Bench from OpenAI’s GPT-5.6 release (July 9):

• Luna (gpt-5.6-luna) — $0.20 / $1.20 · Terminal-Bench 2.1 84.7% · volume / latency lane
• Terra (gpt-5.6-terra) — $2.00 / $12.00 · Terminal-Bench 2.1 87.4% · everyday balance
• Sol (gpt-5.6-sol) — ~$5 / $30 · Terminal-Bench 2.1 88.8% (91.9% with Ultra / multi-agent) · flagship

All three share the family surface: ~1.05M context, 128k max output, Responses API tools (including programmatic tool calling). The split is quality vs dollars vs long-range recall — not “does it have tools?”

The Terminal-Bench story (accurate version)

Some roundups claim Luna “beats” Terra on Terminal-Bench. OpenAI’s own table does not say that. Terra leads Luna (~87.4% vs ~84.7%); Sol leads both (~88.8%). The interesting fact is how close Luna stays while costing an order of magnitude less after the −80% cut.

For agent operators, a ~3-point Terminal-Bench gap at 10× lower API list rates is often the right trade — especially when tool loops dominate spend. For one-shot hard coding or high-stakes agentic work, pay for Terra or Sol and stop guessing.

Price math that actually matters

Per 1M tokens (input/output), Luna is roughly 10× cheaper than Terra and ~25× cheaper than Sol on list rates. Cached input widens the gap further on repetitive agent prompts ($0.02 Luna vs $0.20 Terra vs ~$0.50 Sol).

Agents multiply those rates by tool rounds. A duty that burns 200k tokens/day is a rounding error on Luna and a line item on Sol. That is why the July 30 Luna cut redraws “default model” choices for scheduled monitors and extraction workers — covered in the price-cut news post.

Long context: where Luna loses

All three advertise ~1M context. Independent reporting and vendor tables show Sol and Terra hold multi-needle recall much better deep in the window; Luna’s usable recall can degrade substantially — often described as starting to summarize instead of cite around ~300k tokens. If your workflow is “stuff half a codebase and ask precise questions,” do not pick Luna because the spec sheet says 1M.

Decision guide: which tier

Use this as a default; measure on your harness:

• Choose Luna — high-volume agent loops, tagging, monitors, first-pass drafts, Free/hard-cap products where current-gen OpenAI must fit the budget
• Choose Terra — everyday assistant quality when a few Terminal-Bench points and stronger recall matter more than 10× savings
• Choose Sol — hardest agentic / coding / long-horizon work; Ultra multi-agent when you will pay for the last few points
• Never choose by alias alone — bare gpt-5.6 is Sol; pin the tier slug

ChatGPT vs API (and CloudAxis)

On ChatGPT Work/Codex, Free and Go users get Terra; selecting Luna needs Plus or above (OpenAI’s July 30 post). The API sells all three without that consumer gate. CloudAxis does not offer a tier dropdown: Auto Mode is powered by GPT-5.6 Luna on every plan including Free — capacity upgrades, not a model catalog.

Related reading

Cluster links:

• GPT-5.6 Luna complete guide (pillar)
• Luna −80% price cut
• gpt-5.6 → Sol alias gotcha
• AI cost calculator

Frequently asked questions

Does GPT-5.6 Luna beat Terra on Terminal-Bench?

No. OpenAI’s GPT-5.6 release lists Terminal-Bench 2.1 at about 88.8% Sol, 87.4% Terra, and 84.7% Luna. Luna is close to Terra — not ahead — at a much lower API price after the July 30 cut.

Which GPT-5.6 tier is cheapest?

Luna at $0.20/$1.20 per 1M tokens (cached input $0.02). Terra is $2/$12; Sol is about $5/$30 on common list quotes.

When should I pick Sol over Luna?

When a wrong answer is expensive: hard coding, long-horizon agents, or deep long-context recall. Luna wins when tool-loop volume dominates the bill.

Is gpt-5.6 the same as Terra?

No. The gpt-5.6 alias routes to Sol. Use gpt-5.6-terra or gpt-5.6-luna explicitly.

Which tier does CloudAxis Auto Mode use?

GPT-5.6 Luna on every plan including Free. There is no manual Sol/Terra/Luna picker — plans differ by credits and capacity.