GPT-5.6 Terra: The Middle Tier and Its 272K Pricing Cliff

Most teams pick a model tier once and never revisit it. With GPT-5.6 that habit is costly, because the middle tier, Terra, gets the same context window, the same tool set and the same reasoning controls as the flagship at half the per-token price. The question is not whether Terra is good enough in general. It is which of your workloads actually need more.

This guide covers what Terra is, the one pricing rule that catches people out, and a practical way to decide between Terra and its siblings.

Where Terra sits in the GPT-5.6 lineup

OpenAI released GPT-5.6 on 9 July 2026 as three tiers. Sol is the flagship for complex professional work, Terra balances intelligence and cost, and Luna is for efficient, high-volume workloads. OpenAI's model page says Terra roughly corresponds to the mini tier in earlier GPT-5 families, which is a useful mental anchor: it is the model you are meant to run most of your traffic on, not a cut-down afterthought.

The spec sheet is shared with Sol almost line for line:

• Context: 1,050,000 tokens, with up to 922,000 input and 128,000 output.
• Knowledge cutoff: 16 February 2026.
• Modalities: text and image in, text out.
• Reasoning effort: none, low, medium (the default), high, xhigh and max.
• Responses API tools: web search, file search, code interpreter, hosted shell, apply patch, computer use, MCP and tool search, among others.

The window and tools are not where Terra saves money. The difference is the model itself, which OpenAI positions below Sol on capability and prices accordingly.

What it costs

OpenAI lists Terra at $2 per million input tokens, $0.20 cached and $12 output, after a 20% price cut on 30 July 2026. OpenRouter lists the same $2 and $12 at the time of writing. Sol is currently $4 and $20, so Terra is half the input price and 60% of the output price.

For comparison, GPT-5.6 Luna is listed at $0.20 input and $1.20 output, a tenth of Terra on both. That gap is wide enough that the Terra decision is really two decisions: is Luna too weak for this job, and is Sol worth twice the input price on top of it.

OpenRouter also lists openai/gpt-5.6-terra:batch at $1 input and $6 output. It is the same model on an asynchronous endpoint at half price, worth using for anything that does not need an answer in seconds. Batch API savings math covers when the latency trade is worth it.

The 272K pricing cliff

The rule that surprises people is on the model page: prompts over 272K input tokens are billed at twice the input rate and 1.5 times the output rate, and the surcharge applies to the full request, not just the tokens past the threshold.

Work it through. A 250K-token prompt with a 4K-token answer costs about $0.50 in input and $0.05 in output. Push the same prompt to 280K and the whole request is billed at $4 input and $18 output per million, so input alone rises to about $1.12. Twelve percent more context more than doubles the bill.

The practical consequences:

• Trim before you cross the line. If a retrieval step regularly lands at 260K to 300K, cutting the bottom-ranked chunks is worth far more than it looks.
• Cache the stable prefix. Cached input is $0.20 per million, though cache writes cost 1.25 times the uncached rate, so caching pays off only when the prefix is reused.
• Do not treat 1M as free capacity. Context window cost trade-offs explains why the largest window is rarely the cheapest way to answer a question.

Choosing an effort level

Because Terra has the full effort range, the effort setting is often a better lever than switching tier. A Terra request at high may cost less than a Sol request at medium while closing much of the quality gap on reasoning-heavy work, but that is a claim to measure on your own tasks, not to assume.

A sensible default is medium, dropping to low or none for extraction, classification and formatting, where reasoning tokens add cost and latency without changing the output. OpenAI's own guidance is to reach for xhigh only when evals show a clear benefit.

If max effort still leaves Terra short, there is one more step before Sol: GPT-5.6 Terra Pro, the same model with OpenAI's pro reasoning mode enabled.

Terra, Sol or Luna: a decision rule

Start every new workload on Terra at medium effort and measure the pass rate on twenty real examples. Then move in whichever direction the numbers point:

1. Terra passes nearly everything: try Luna. If it holds up, you have cut the bill by roughly ninety percent.
2. Terra fails a handful in ways that look like reasoning errors: raise effort, then try Terra Pro, before paying for Sol.
3. Terra fails on breadth of knowledge or long multi-step planning: that is where Sol earns its premium.
4. The answer is mixed: route by task type rather than picking one tier for everything.

One more consideration: OpenAI's models page now lists GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna as its flagship models. Terra is still a current, supported model, but if you are starting a new project, compare against the GPT-6 tier at the same price point rather than treating GPT-5.6 as the default.

Takeaway

Terra is the right starting point for most GPT-5.6 work because it shares Sol's window, tools and controls at half the input price. Keep prompts under 272K tokens unless you have a reason not to, use batch for anything asynchronous, and escalate by effort and mode before escalating by tier.