OPENAI MODEL GUIDE
GPT-5.6 Terra API
The balanced GPT-5.6 tier for production workloads that need strong reasoning and tools at a lower price than Sol.
General coding, review, tool-using assistants, long documents, and structured business workflows.
The hardest repository tasks, or simple high-volume work Luna already handles reliably.
Read by decision
Why Terra should be the first baseline
Terra sits between Sol and Luna and is the useful first test for most production work. It is not a placeholder compromise: it reveals which requests really need escalation and which can be routed down for savings.
- Production coding and code review.
- Tool-using assistants and structured business workflows.
- Long-document analysis where flagship pricing is difficult to justify.
- A default GPT-5.6 evaluation baseline before escalating to Sol.
Escalate to Sol or route down to Luna?
| Observed signal | Action | Why |
|---|---|---|
| Complex tasks miss the quality bar | Escalate to Sol | Test whether the higher ceiling adds passes |
| Task is stable and narrow | Evaluate Luna | Reduce high-volume cost |
| Tool retries are frequent | Fix prompt and schema first | A pricier model may not fix protocol errors |
| Long-context cost is high | Retrieve, compact, cache | Remove wasted input first |
Model the real cost of Terra
Vendor list pricing is $2 input and $12 output per million tokens, with $0.20 cached input. Production budgeting must also include reasoning output, tool loops, retries, and long-context tiers.
- Measure cost per accepted result, not per API call.
- Separate the normal path from retry and recovery paths.
- Recalculate the entire request above 272K input tokens.
- Compare Sol's incremental pass rate and Luna's retry rate.
Tune latency and quality with reasoning effort
Lock a medium baseline first. Try none or low for simpler work; move to high, xhigh, or max only when hard-sample pass rates improve clearly.
- The same 1.05M context and 128K maximum output published for the GPT-5.6 family.
- Streaming, function calling, structured outputs, and Responses API tools.
- Six configurable reasoning levels for latency and quality control.
- Image understanding for visual documents, screenshots, and mixed text-image prompts.
Minimal Terra request
Confirm the model ID in Model Plaza, then test streaming, structured outputs, and tool parameters in the same way the application will use them.
curl https://app.llmfly.ai/v1/chat/completions \
-H "Authorization: Bearer $LLMFLY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-5.6-terra",
"messages": [{"role": "user", "content": "Summarize the tradeoffs in three bullets."}],
"stream": false
}'Run a two-direction production evaluation
- Compare the hardest Terra failures with Sol.
- Compare high-frequency narrow tasks with Luna.
- Record quality, latency, retries, and cost per accepted result.
- Keep escalation and downgrade rules in routing configuration, not scattered business logic.
Terra boundaries and risks
- Sol may still be a better candidate for the hardest reasoning and repository tasks.
- Luna can be materially cheaper for simple, high-volume workloads.
- Long-context requests above 272K input tokens use a higher vendor pricing tier.
- Do not assume every provider-specific tool is exposed by the gateway route.
GPT-5.6 Terra specification sources and usage notes
Specifications and vendor prices were checked against OpenAI GPT-5.6 Terra documentation: https://developers.openai.com/api/docs/models/gpt-5.6-terra. Confirm the live LLMFly AI route and rate in Model Plaza.
Frequently asked questions
Is Terra a good default GPT-5.6 model?
It is a strong production baseline, subject to real-task evaluation. Escalate hard failures to Sol and test Luna for simple high-volume work.
How do I decide whether Sol's premium is worth it?
Compare incremental accepted results on the same evaluation set, not token price alone.
When should Terra route down to Luna?
When the task is narrow, the schema is stable, and Luna's retry rate does not erase its lower token rate.
What context window does Terra support?
The vendor publishes 1.05M context and 128K maximum output; account for long-context pricing above 272K input.
How do I confirm the LLMFly AI Terra model ID?
Copy the model ID available to the current API key in Model Plaza and verify it with a small request.
Test GPT-5.6 Terra with your own task
Create an API key, confirm the model ID, and start with a small representative request.