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.

Try Apidog today

That claim changes the migration decision. If Terra matches GPT-5.5 on your workload, requests still routed to GPT-5.5 create avoidable spend. Terra is not a budget compromise; it is a candidate for your new production default.

This guide explains how to evaluate Terra, migrate from GPT-5.5, decide when Sol justifies its premium, identify routes that can move to Luna, and regression-test the change before it reaches users.

TL;DR

• gpt-5.6-terra is the balanced GPT-5.6 tier, positioned between flagship Sol and speed-focused Luna.
• Pricing is $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 get Terra by default, signaling OpenAI's confidence in this tier.
• Treat a GPT-5.5 migration as a tuning pass, not just a model-slug change. Test your current reasoning effort and one level lower, and expect shorter outputs.
• All tiers share the feature set: six reasoning effort levels, pro mode, explicit prompt caching, and new Responses API capabilities.
• Before switching production traffic, run representative prompts against gpt-5-5 and gpt-5.6-terra in Apidog and compare quality, latency, and token usage.

What GPT-5.6 Terra is

GPT-5.6 has three tiers:

• Sol: flagship tier with the deepest reasoning.
• Terra: balanced tier for most production workloads.
• Luna: fastest and cheapest tier for high-volume, latency-sensitive calls.

The model ID is gpt-5.6-terra. API access is self-serve for API accounts.

One implementation detail matters: the bare gpt-5.6 alias routes to Sol, not Terra. If you want Terra pricing, set the model ID explicitly.

{
  "model": "gpt-5.6-terra"
}

GPT-5.6 pricing per 1M tokens

Model	Input	Output	Positioning
gpt-5.6-sol	$5.00	$30.00	Flagship, deepest reasoning
gpt-5.6-terra	$2.50	$15.00	Balanced, GPT-5.5-class output
gpt-5.6-luna	$1.00	$6.00	Fastest, high-volume work

Compared with the GPT-5.5 pricing baseline, Terra is positioned as delivering the same class of output for about half the cost.

Per OpenAI's help center, Terra is also the default model for free and Go ChatGPT users.

Terra vs GPT-5.5: treat migration as a tuning pass

Do not only replace this:

- "model": "gpt-5-5"
+ "model": "gpt-5.6-terra"

OpenAI's developer docs recommend treating migration as a tuning pass. Test these changes explicitly.

1. Re-evaluate reasoning effort

Reasoning effort maps differently across generations. Start with your current effort, then test one level lower.

For example, if your GPT-5.5 route uses medium, evaluate Terra at both medium and low:

{
  "model": "gpt-5.6-terra",
  "reasoning": {
    "effort": "low"
  },
  "input": "..."
}

If Terra at lower effort preserves quality, you reduce both latency and output-token usage on top of the lower per-token price.

2. Remove redundant brevity instructions

GPT-5.6 outputs are shorter and less likely to include generic introductions. Prompts such as these may over-constrain Terra:

Be concise.
Skip the preamble.
Do not add context.
Keep the answer extremely short.

Remove or relax those instructions during testing. Otherwise, you may mistake clipped output for a model-quality regression.

3. Track caching behavior

Watch cached-token usage during your evaluation window, especially if you adopt explicit caching.

Early documentation coverage, including Simon Willison's day-one write-up, reports a 1M-token context window, 128K maximum output, and a February 16, 2026 knowledge cutoff. Treat those as reported figures until OpenAI's specification pages settle.

A practical migration workflow

Use 20 to 50 tasks from real production traffic, not synthetic prompts.

1. Export representative requests from logs.
2. Run each request with your existing gpt-5-5 configuration.
3. Run the same requests on gpt-5.6-terra at the current reasoning effort.
4. Run them again one effort level lower.
5. Score output quality against the GPT-5.5 baseline.
6. Record input, output, and cached token counts.
7. Keep routes that regress on a higher effort level instead of abandoning the migration globally.

This lets you route based on measured results rather than model-family assumptions.

Terra vs Sol: when the flagship earns 2x

Sol costs exactly twice Terra's rate:

• Sol: $5 input / $30 output per 1M tokens
• Terra: $2.50 input / $15 output per 1M tokens

GPT-5.6 Sol is the tier represented in OpenAI's launch benchmarks. Per OpenAI, Sol scored roughly:

• 53 on Agents' Last Exam, compared with GPT-5.5's 46.9
• 88.8% on Terminal-Bench 2.1, or 91.9% with the ultra setting
• 73.5 on ExploitBench, compared with 47.9
• 62.6 on OSWorld 2.0, compared with 47.5

Those are launch-day claims, and they are Sol results. OpenAI has not published the same depth of benchmark data for Terra.

Sol also does not lead every benchmark. On SWE-Bench Pro, Claude Fable 5 scores 80.3% to Sol's 64.6%, according to the same launch materials. Benchmark your own workload rather than relying on vendor charts.

Use Sol when failed runs are expensive:

• Long-horizon agentic coding where a failure costs an engineer's afternoon.
• Multi-step tool orchestration where reasoning quality compounds across turns.
• Workloads that justify ultra, the multi-agent setting that runs four agents in parallel and trades more token spend for faster wall-clock results.

For chat, summarization, extraction, classification with context, and many RAG answer-generation routes, Terra is often the better starting point.

For example, at 10M input tokens and 2M output tokens per day:

Tier	Daily cost	30-day cost
Terra	$55	$1,650
Sol	$110	$3,300

That $1,650 monthly difference can fund evaluation work for the routes where Sol may actually matter.

Terra vs Luna: when to drop further

Luna costs $1 input and $6 output per 1M tokens, 60% below Terra.

It is designed for high-volume and latency-sensitive tasks such as:

• Classification
• Extraction
• Routing
• First-pass drafting

Use Terra as the default evaluation target, then move individual routes to Luna once your evals show that quality holds.

This route-level split is a normal production architecture:

High-risk agentic workflows  -> Sol
General assistant and RAG     -> Terra
Narrow, high-volume tasks     -> Luna

The Sol vs Terra vs Luna comparison provides a route-by-route decision framework.

Every tier gets the full feature set

Choosing Terra does not mean giving up API capabilities. All GPT-5.6 tiers support:

• Six reasoning effort levels: none, low, medium, high, xhigh, and max.
• Pro mode: reasoning.mode: "pro", a quality-first parameter available on every tier.
• Explicit prompt caching: set prompt_cache_options.mode to explicit with a ttl.
• Responses API capabilities: programmatic tool calling, multi-agent execution in beta, persisted reasoning through reasoning.context, and vision detail settings such as original and auto.

For explicit caching:

• Cache writes bill at 1.25x the uncached input rate.
• Cache reads retain a 90% discount.
• Cached content lives for at least 30 minutes.

For agents and chatbots with long stable system prompts, caching can reduce effective input cost further.

Minimal Terra Responses API request

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

A clean model comparison requires repeatable requests, shared variables, and token-level observability. Download Apidog and set up the comparison workflow.

1. Parameterize the model ID

Save one Responses API request and use an environment variable instead of hardcoding the model.

{
  "model": "{{model}}",
  "reasoning": {
    "effort": "{{reasoning_effort}}"
  },
  "input": "{{prompt}}"
}

2. Create comparison environments

Create at least two environments:

Environment	model
GPT-5.5 baseline	gpt-5-5
Terra candidate	gpt-5.6-terra

Optionally add a third environment for Luna:

Environment	model
Luna candidate	gpt-5.6-luna

3. Run the same prompt set

Execute the saved prompt collection against each environment. Compare responses side by side and use the same inputs, tools, and reasoning settings for every run.

4. Record actual usage

Inspect the usage object in every response. Track:

• Input tokens
• Output tokens
• Cached tokens
• Latency, if available in your request logs
• Task-specific quality scores

The pricing page gives you rates. Response usage tells you what your application actually costs.

Expect Terra responses to be shorter than GPT-5.5 outputs. Check downstream assumptions that may depend on output length:

• Parsers expecting a minimum number of sections
• UI layouts that assume long-form answers
• Token budgets calibrated for GPT-5.5
• Validation rules that reject concise but valid output

After Terra passes, repeat the same test set for routes that may be suitable for Luna.

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 expected trade is comparable quality, shorter outputs, and lower cost. Benchmark representative production tasks before moving traffic.

How much does GPT-5.6 Terra cost?

Terra costs $2.50 per 1M input tokens and $15 per 1M output tokens. Cache reads retain a 90% discount under explicit caching. See the GPT-5.6 pricing breakdown for Sol, Luna, and caching economics.

Which ChatGPT plans include Terra?

All of them. Free and Go users get Terra as their model. Plus and higher users can choose Sol, Terra, or Luna and set per-model reasoning effort. Sol is available at medium effort and above on Plus.

Do I need code changes to use Terra?

The request shape is unchanged if you already use the Responses API. Update the model ID to gpt-5.6-terra, then re-test your reasoning effort, prompt instructions, and downstream handling of shorter outputs.

The default has moved

GPT-5.5 is not deprecated today, but Terra changes its economics. If comparable output costs roughly half as much, keeping GPT-5.5 should be a measured decision.

Start with Terra. Promote hard agentic routes to Sol when evals justify the premium. Demote narrow, high-volume routes to Luna when they pass the same tests.

The next step is straightforward: select 20 real prompts from production logs, run them against gpt-5-5 and gpt-5.6-terra in Apidog, and compare output quality and token counts. The migration claim becomes useful only when your own evals confirm it.