GPT-6.1 Terra vs GPT-5.6 Terra: API pricing, caching, and upgrading
The new Terra costs 54% less than the old one on the half cached example, so the upgrade comes with a saving to put against the cost of testing it. GPT-6.1 Terra moves the whole request to a higher rate above 272,000 input tokens, so check large prompts separately.
Upgrading is a one line change in your code and a much bigger change in your test results. GPT-6.1 Terra costs 50% less than GPT-5.6 Terra on input and 58% less on output, so moving over cuts the bill before you do anything else. The rates are only half of the decision. Your prompts, your tool definitions and your output checks were tuned against the old model, and a new model can follow the same instructions a little differently. That's why this page treats the move as a cost to estimate and a test to run, not a free swap. It prices both models on the same requests, then shows how many calls it takes for a saving to pay back the work of switching.
Work out what moving from the older Terra to the new one does to the bill, and when the switch pays back.
By TechCompare · Updated
OpenAI: GPT-6.1 Terra
OpenAI's own price
- Input
- $1
- Output
- $5
- Cached
- $0.05
per 1M tokens
Price checked 2026-11-07
OpenAI: GPT-5.6 Terra
OpenAI's own price
- Input
- $2
- Output
- $12
- Cached
- $0.20
per 1M tokens
Price checked 2026-11-07
Cost Comparison
Based on 100,000 input tokens (50% cached), 5,000 output tokens, and 100 requests.Prices come live from OpenRouter at each maker's own price where it's listed. Temporary deals are left out.
Side-by-side specs
| Spec | GPT-6.1 Terra | GPT-5.6 Terra |
|---|---|---|
| Standard input per 1M tokens | $1.00 (better on this spec) | $2.00 |
| Standard output per 1M tokens | $5.00 (better on this spec) | $12.00 |
| Cache reads per 1M tokens | $0.05 (better on this spec) | $0.20 |
| 100K input + 5K output, no cache hits | $0.125 (better on this spec) | $0.26 |
| 100K input + 5K output, half the input cached | $0.0775 (better on this spec) | $0.17 |
| 100K input + 5K output, all input cached | $0.03 (better on this spec) | $0.08 |
| 10,000 requests, half the input cached | $775 (better on this spec) | $1,700 |
| Input / output on long prompts, per 1M | $2.00 / $7.50 above 272,000 input | $4.00 / $18.00 above 272,000 input |
| Batch API input / output per 1M | $0.50 / $2.50 | $1.00 / $6.00 |
How they differ
Input costs $1 per million tokens on GPT-6.1 Terra and $2 on GPT-5.6 Terra. Output costs $5 and $12, and cache reads cost $0.05 and $0.20. A request with 100,000 input and 5,000 output tokens costs $0.125 and $0.26 with no cache hits, $0.0775 and $0.17 with half the input read from cache, and $0.03 and $0.08 with all of it cached. Cache reads cost 75% less on the new model, so a workload that keeps a large prefix warm feels the change most. Output is cheaper on the new model, which helps agents and writing tools that produce long answers. A support chat with 4,000 input tokens and 500 output tokens per call, three quarters of the input cached, costs $1,095 over 300,000 calls on the new model and $2,580 on the old one. The new model also lists Batch API prices here, so a nightly job of 20,000 documents with 6,000 input and 800 output tokens each costs $100 through batch, against $432 on the old model at its standard rates. The old model lists Batch API prices here too. These examples leave out cache writes and tool fees, and the embedded chart follows live OpenRouter rates, which can depend on provider routing.
Verdict
Treat the upgrade as a small project. Pick 50 to 200 real tasks from your logs, run them through both models with the same prompts, and score them with the checks you already trust, such as passing tests or valid output. Then change the prompts and test again, because the new model may need different instructions to do its best. On the half cached request the new model saves $0.0925 per call, or $925 per 10,000 calls. Divide your one time cost, which means engineer hours, eval runs and prompt fixes, by the saving per call to see how many calls it takes to pay back. Don't forget the token counts. Compare the usage fields from a test run on both models, since the same text may not bill the same number of tokens. Roll out in steps, with a share of traffic on the new model and a way back.
Which should you pick?
Choose GPT-6.1 Terra
Move to GPT-6.1 Terra when your evals show it matches or beats the old model on your tasks. The saving then comes free with the move, and it grows with volume. At 100,000 input and 5,000 output tokens with half the input cached, 10,000 calls cost $775 on the new model against $1,700 on the old one. It also makes sense when you're starting a new feature, because you can tune its prompts for one model from day one. A job that can wait for results gets Batch API prices on the new model here.
Official GPT-6.1 Terra referenceChoose GPT-5.6 Terra
Stay on GPT-5.6 Terra while it passes your checks and you have no use for what the newer model adds. A tuned, stable application has a real cost to change. You'd re-run evals, rewrite parts of the prompts and watch a rollout, and that effort is worth it only if the saving or the quality gain is large enough. Waiting does cost something here. Each half cached call costs $0.0925 more on the old model, so a busy product pays for the delay every day. Check OpenAI's pages for how long the older model stays available, and plan the move before you have to make it.