GPT-5.6 Terra: OpenAI's quiet GPT-5.5 replacement at half the cost

When OpenAI shipped GPT-5.6 on July 9, 2026, the headlines went to Sol, the flagship tier with the deepest reasoning and the benchmark charts. Read the GPT-5.6 announcement past the Sol numbers, though, and the model with the biggest production impact sits one row down. GPT-5.6 Terra costs $2.50 per million input tokens and $15 per million output tokens, and OpenAI positions it as competitive with GPT-5.5 at roughly half the price.

That one claim reframes the launch. If Terra matches GPT-5.5 on your workload, every request you still route to GPT-5.5 is spend you no longer need. Terra is not the budget compromise in the family; it is the new default production choice. Its arrival also marks the quiet end of GPT-5.5’s run as OpenAI’s workhorse model.

This guide covers what Terra is, how to treat the GPT-5.5 migration, when Sol earns twice the price, when Luna undercuts both, and how to regression-test the swap before your users ever see it.

TL;DR

• gpt-5.6-terra is the balanced tier in the GPT-5.6 family, between the flagship Sol and the speed-focused Luna.
• Pricing: $2.50 input / $15 output per 1M tokens. OpenAI positions Terra as competitive with GPT-5.5 at roughly half the cost.
• Free ChatGPT users now get Terra by default, a strong signal of OpenAI’s confidence in this tier.
• Migration from GPT-5.5 is “a tuning pass, not only a model-slug change” per OpenAI’s docs: test your current reasoning effort and one level lower, and expect shorter outputs.
• All three tiers share the full feature set: six reasoning effort levels, pro mode, explicit prompt caching, and the new Responses API capabilities.
• Before flipping production, run your saved prompt set against both gpt-5-5 and gpt-5.6-terra in Apidog and compare outputs and token counts side by side.

What GPT-5.6 Terra is

GPT-5.6 is three models in one generation. Sol is the flagship with the deepest reasoning. Luna is the fastest and cheapest, built for high-volume, latency-sensitive calls. Terra sits between them: enough reasoning depth for most production work, priced to undercut the model it replaces. The tier names are durable, and OpenAI says each tier will advance on its own cadence from here. Terra is a lane you can build around, not one-off launch branding.

The model ID is gpt-5.6-terra, and API access is self-serve for any API account. No waitlist, no plan gating. One detail worth flagging: the bare alias gpt-5.6 routes to Sol, not Terra. If you want the cheaper tier, name it explicitly, or a lazy config default will double your bill without changing a line of code.

Here is the family pricing per 1M tokens:

Set those numbers against the GPT-5.5 pricing baseline and the pitch becomes blunt: the same class of output for about half the invoice. OpenAI is not framing Terra as a lightweight spin-off. It is framing Terra as the model that makes its previous flagship redundant on price.

There is one more signal hiding in plain sight. Per OpenAI’s help center, Terra is now the model free and Go ChatGPT users get. OpenAI put this tier in front of its largest and least forgiving audience on day one. Companies do not do that with a model they hedge on.

Terra vs GPT-5.5: a migration decision, not a find-and-replace

The tempting move is to swap the model string and ship. OpenAI’s own developer docs warn against exactly that: “Treat migration as a tuning pass, not only a model-slug change.”

Three behaviors shift underneath you:

• Reasoning effort maps differently. OpenAI recommends testing your current effort level and one level lower. Terra at reduced effort may match your GPT-5.5 output quality while cutting both latency and output tokens, which compounds the price advantage.
• Outputs get shorter. GPT-5.6 writes tighter answers with fewer generic intros. If your prompts carry instructions like “be concise” or “skip the preamble”, remove them; stacked brevity directives on an already-brief model can clip content you wanted to keep.
• Caching is worth re-checking. Watch cached-token usage during your test window, especially if you adopt the new explicit caching mode covered below.

Specs come with a hedge for now. Early documentation coverage, including Simon Willison’s day-one write-up, reports a 1M-token context window, 128K max output, and a knowledge cutoff of February 16, 2026. Treat those as reported figures until OpenAI’s spec pages settle.

The practical migration plan looks like this: pull 20 to 50 representative tasks from real production traffic, not synthetic prompts. Run them against Terra at your current effort setting and again one level lower. Score the outputs against your GPT-5.5 baseline, and log token counts for every run. If quality holds at the lower effort, you bank savings twice, once on the per-token rate and once on shorter outputs. If it degrades on a specific route, keep that route on higher effort rather than abandoning the migration. The point of a half-price model is lost if you burn the savings on regressions, so measure before you move.

Terra vs Sol: when the flagship earns 2x

Sol costs exactly twice Terra’s rate, $5 / $30 against $2.50 / $15. GPT-5.6 Sol is the tier OpenAI’s launch benchmarks describe, and the deltas are real, per OpenAI: roughly 53 on Agents’ Last Exam against GPT-5.5’s 46.9, 88.8% on Terminal-Bench 2.1 (91.9% with the ultra setting), 73.5 on ExploitBench against 47.9, and 62.6 on OSWorld 2.0 against 47.5. Those are launch-day claims, and they are also Sol numbers. OpenAI has not published the same depth of benchmark data for Terra, which tells you where the 2x premium lives: at the hard ceiling of agentic work.

An honest caveat belongs here: even Sol does not lead everywhere. On SWE-Bench Pro, Claude Fable 5 scores 80.3% to Sol’s 64.6%, per the same launch materials. The frontier is contested this cycle, which is one more reason to benchmark your own tasks instead of trusting any vendor’s chart.

When does Sol earn double the price?

• Long-horizon agentic coding, where one failed run costs an engineer’s afternoon.
• Multi-step tool orchestration where reasoning depth compounds across turns.
• Workloads where you would enable ultra, the multi-agent setting that runs four agents in parallel and trades extra token spend for faster wall-clock results. Ultra lives in ChatGPT Work on Pro and Enterprise plans, and in Codex from Plus up.

And when does it not? Most chat. Summarization. Extraction. Classification with context. RAG answering, where retrieval quality matters more than reasoning ceiling. For those, Terra at half the cost is the rational default. The arithmetic is hard to argue with: a service pushing 10M input and 2M output tokens a day bills $55 on Terra and $110 on Sol. Over a month, that gap is $1,650, enough to fund a serious eval suite for the routes where you suspect Sol might matter.

Terra vs Luna: when to drop further

Luna costs $1 / $6 per 1M tokens, 60% below Terra. It is the fastest tier, built for high-volume, latency-sensitive work: classification, extraction, routing, first-pass drafting. If the task is narrow and the prompt does the heavy lifting, Luna often gets you there for less than half of Terra’s rate.

A useful mental model: Terra is where you start, and Luna is where individual routes graduate once you have evals proving the cheaper tier holds. Splitting traffic across tiers is normal operation now, not an optimization for later. The Sol vs Terra vs Luna comparison breaks the decision down route by route if you want the full framework.

Every tier gets the full feature set

Downgrading tiers used to mean losing capabilities. Not here. All three GPT-5.6 models share the same surface:

• Six reasoning effort levels: none, low, medium, high, xhigh, max.
• Pro mode (reasoning.mode: "pro"), a quality-first setting available on every tier. It is a parameter, not a separate model.
• Explicit prompt caching: set prompt_cache_options.mode: "explicit" with a ttl field. Cache writes bill at 1.25x the uncached input rate, cache reads keep the 90% discount, and cached content lives for at least 30 minutes. For an agent or chatbot with a long, stable system prompt, this stacks with Terra’s base price to push effective input cost well below the sticker rate.
• The new Responses API capabilities: programmatic tool calling, where the model writes JavaScript that orchestrates tool calls inside an isolated V8 runtime with no network access; multi-agent execution in beta; persisted reasoning across turns via reasoning.context; and vision detail settings (original / auto) that preserve original image dimensions.

The consequence is clean: tier choice is purely a price-versus-depth decision. You never trade away an API feature by picking Terra.

A minimal Terra call through the Responses API looks like this:

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-terra",
    "reasoning": { "effort": "low" },
    "input": "Summarize this support thread and flag any refund request."
  }'

Regression-test the swap before production sees it

The migration checklist above only works if you can compare runs cleanly, and that is a tooling problem before it is a model problem. Download Apidog and the workflow takes an afternoon:

1. Save your Responses API request once, with the model name as an environment variable instead of a hardcoded string.
2. Create two environments: one sets the variable to gpt-5-5, the other to gpt-5.6-terra.
3. Run your saved prompt set against each environment and compare the responses side by side.
4. Read the usage fields on every response: input tokens, output tokens, and cached tokens tell you what the swap costs, not what the pricing page implies.

Expect Terra’s responses to run shorter than GPT-5.5’s on the same prompts. That is by design, but it has downstream consequences: parsers that assume a minimum length, UIs that pad short answers awkwardly, token budgets tuned to the old model. The token-count comparison is the step most teams skip, and it is the step that decides the real economics of the migration.

Once Terra passes on your prompt set, add a third environment for gpt-5.6-luna and re-run the cheap routes. The same saved requests answer the next cost question before anyone asks it.

FAQ

Is GPT-5.6 Terra better than GPT-5.5?

OpenAI positions Terra as competitive with GPT-5.5, not strictly better, at roughly half the price. For most workloads the trade is comparable quality with shorter outputs and a smaller bill. The honest answer for your stack comes from benchmarking your own representative tasks before you switch traffic.

How much does GPT-5.6 Terra cost?

$2.50 per 1M input tokens and $15 per 1M output tokens, with cache reads keeping a 90% discount under the new explicit caching mode. For the full family economics, including Sol, Luna, and the caching math, see the GPT-5.6 pricing breakdown.

Which ChatGPT plans include Terra?

All of them. Free and Go users get Terra as their model. Plus and higher can choose between Sol, Terra, and Luna and set a per-model reasoning effort level, with Sol available at medium effort and up on Plus.

Do I need code changes to use Terra?

The call shape is unchanged; Terra works with your existing Responses API integration, and the model IDs are confirmed in OpenAI’s developer docs. What needs attention is tuning: re-test your reasoning effort level, strip brevity directives from prompts, and confirm that downstream code handles shorter outputs.

The default has moved

GPT-5.5 is not deprecated today, but its economics are. When the same vendor sells comparable output at half the price, staying put is a decision you should be able to defend with data, and most teams will not be able to. Terra is the new starting point: begin there, promote hard agentic routes to Sol when evals demand it, and demote narrow high-volume routes to Luna when they pass.

The next step is small. Pull 20 real prompts from your production logs, run them against gpt-5-5 and gpt-5.6-terra in Apidog, and let the outputs and token counts make the call. Half-price capability is a claim until your own evals confirm it, and confirming it costs you one afternoon.