GPT-5.6 Terra

What it is

GPT-5.6 Terra is the middle of OpenAI's three scored GPT-5.6 tiers — Luna the cost tier, Sol the frontier tier — and the board classes it as the balanced one. The page's own note calls it the sensible default for most GPT builds, and the six scores bear that out: nothing exceptional, nothing broken, and a level of evenness that only the mid tiers achieve.

As with every closed OpenAI model, the parameters, training data and architecture are not published. What can be verified externally is the pricing, the context window and the model identifier; the composite below is this board's own measurement, not a vendor figure.

How the composite is built

Overall 86 weights Coding 24%, Terminal 20%, Reasoning 20%, Tool use 15%, Context 13%, Speed 8%. Terra's profile spans 75 to 90 — tight for a general model, and the reason it lands just under the frontier cluster without leading any axis.

Coding 88

Level with Qwen 4 Max and five behind its sibling Sol's 93. Comfortable for everyday engineering; four behind DeepSeek V4's 92 at a fraction of its blended price.

Terminal 84

Level with DeepSeek V4 and five behind Sol's 89. Supervised shell work is fine; long unattended runs are where the frontier agentic tiers still pull away.

Reasoning 89

Its joint-strongest axis with Context. Behind Sol's 95 and Gemini 3.6 Pro's 93, but ahead of Claude Sonnet 5's 88 — planning is not usually the failure mode.

Tool use 86

Level with Command R+ 2 and five behind Sol's 91. Structured calls and multi-step chains hold up; expect the usual retries at the tail of long workflows.

Context 90

Level with Claude Opus 5's 90, behind Gemini 3.6 Pro's 98 and the 95 of the 1M-context Claude tiers. The best of Terra's six, and the axis that supports whole-module work.

Speed 75

The cost of the tier. Fifteen behind Luna's 90 and three behind Sonnet 5's 78. Fine for background jobs, noticeable in interactive use.

Where it fits

The default GPT choice when Sol's rates hurt but Luna's quality ceiling is too visible: production agent loops, professional tooling, document and code work at moderate volume. If a build works on Terra and you need the last few points of reasoning or tool reliability, the upgrade path to Sol is small in code terms and large in price.

Limits

• Value is rated Poor — 42.1 index points per $4.50/M blended is 9.4, and Sol, the higher tier, actually blends slightly cheaper at $4.00/M because its output rate is lower. Check both prices before assuming the middle tier is the economy one.
• Speed 75 is the slowest thing about it and the main reason to route lightweight calls to Luna.
• Closed weights. No self-hosting, no fine-tuning on your own hardware; exit cost is rewriting the prompts for another provider.
• The GPT-6 line is already on the board unscored — Astra, Sol and Luna listings from September 2026. Terra's position is a point-in-time read.

Price and access

$2 per million input tokens and $12 per million output, blending to $4.50/M. Listed on OpenRouter and OpenCode Zen, with the Artificial Analysis index at 42.1 — a strong score whose per-dollar value is the weak point. Last scored 15 Sep 2026.

Alternatives on this board

• GPT-5.6 Sol — 90 at $2/$10. Four points up and, on the blended rate, marginally cheaper. The obvious internal comparison.
• GPT-5.6 Luna — 81 at $0.20/$1.20 with Speed 90 and Context 89. The cheap sibling for volume work.
• DeepSeek V4 — 86 at $0.78/$1.57 with open weights. The same overall for roughly a fifth of the blended price.
• Gemini 3.6 Pro — 89 with Context 98, if long context matters more than the OpenAI toolchain.

Sources