TL;DR
- Pick GPT-6.1 Terra for volume and cost: it's half Sol's price ($0.25 / $1.25 vs $0.50 / $2.50), and short structured tasks cost exactly half as much per call.
- Pick GPT-6.1 Sol for speed: across the same six tasks, Sol finished first every time.
- Both are live on HiAPI with the same key and the same endpoint; switching is one
modelfield. Open the Terra model page to try it.
At a glance
| GPT-6.1 Terra | GPT-6.1 Sol | |
|---|---|---|
| Model ID | gpt-6.1-terra | gpt-6.1-sol |
| Input / output (per million tokens) | $0.25 / $1.25 | $0.50 / $2.50 |
| Cached read | $0.0125 | $0.025 |
| Reasoning efforts | none, low, medium, high, xhigh, max | low, medium, high, xhigh, max |
| Can reasoning be turned off? | Yes (none) | No, lowest is low |
| Context / max output | 1.05M / 128K | 1.05M / 128K |
| Endpoints | Responses (default), Chat Completions | Responses (default), Chat Completions |
HiAPI prices, checked October 29, 2026.
Same prompt, one run each
Three tasks at medium effort, in English and Chinese, once on each model: 12 runs.
| Task | Language | Terra cost / time | Sol cost / time |
|---|---|---|---|
| Extract an order from an email | English | $0.0003 / 23 s | $0.0006 / 7 s |
| Extract an order from an email | Chinese | $0.0003 / 11 s | $0.0006 / 9 s |
| Find bugs in code | English | $0.0019 / 70 s | $0.0031 / 48 s |
| Find bugs in code | Chinese | $0.0033 / 56 s | $0.0035 / 42 s |
| Write product listing copy | English | $0.0009 / 20 s | $0.0008 / 13 s |
| Write product listing copy | Chinese | $0.0019 / 59 s | $0.0010 / 21 s |
Cost didn't always halve. On the Chinese listing copy, Terra spent 1,411 reasoning tokens to Sol's 293, so Terra's run cost more. Half the rate doesn't guarantee half the bill; it depends on how much the model reasons on that call.
Round 1: extract an order from an email (strict JSON)
Extract the order details from this customer email.
Email: Hi, I placed an order last Tuesday (Oct 20), order number HX-20481: two dark grey hoodies in size L and one pair of black joggers in size M, $86.40 in total. The package arrived today and the hoodies are light grey; the joggers are fine. I'd like to exchange both hoodies for dark grey, or get a refund if you're out of stock. You can reach me at +1 415 555 0123, afternoons work best. - Sarah
Both requests used the same JSON schema (order ID, date, items, total, issue, requested action, phone, contact time), and both returned valid JSON.
| GPT-6.1 Terra | GPT-6.1 Sol |
|---|---|
"issue":"Both hoodies arrived light grey instead of dark grey; the black joggers are correct." | "issue":"Both hoodies arrived in light grey instead of dark grey. The joggers are fine." |
"requested_action":"Exchange both hoodies for dark grey ones, or provide a refund if dark grey hoodies are out of stock." | "requested_action":"Exchange both hoodies for dark grey, or refund them if out of stock." |
| Other fields: HX-20481, Oct 20, hoodie / dark grey / L / 2, joggers / black / M / 1, $86.40, +1 415 555 0123, Afternoons | Identical |
What you can see: every extracted value matches, word for word. Only the free-text fields differ in wording, and Sol's are shorter. Terra returned compact single-line JSON; Sol returned it indented. On the Chinese email, Terra wrote "dark grey (received light grey)" into the color field, which would need cleanup before going into a database.
Round 2: write product listing copy
Write Amazon product listing copy from these specs: a title under 80 characters and 3 bullet points, each under 25 words. Avoid words like "ultimate" or "premium".
Specs: portable blender, 380 ml (13 oz), USB-C charging, 15 blends per charge, 6 stainless steel blades, Tritan cup, 420 g, IPX7 waterproof so the whole unit can be rinsed.
| GPT-6.1 Terra | GPT-6.1 Sol |
|---|---|
| Title: Portable Blender, 380 ml/13 oz, USB-C Rechargeable, 6-Blade, IPX7 Waterproof | Portable Blender, 13 oz (380 ml), USB-C Rechargeable, 6 Stainless Steel Blades |
| Blend smoothies, shakes, and drinks anywhere with a compact 380 ml Tritan cup and lightweight 420 g design. | Take blending on the go with a 13 oz (380 ml) Tritan cup and lightweight 420 g design. |
| USB-C charging delivers up to 15 blends per charge for convenient use at home, work, or on the go. | Enjoy 15 blends per charge with convenient USB-C charging and six stainless steel blades. |
| Six stainless steel blades support smooth blending, while IPX7 waterproofing lets you rinse the whole unit after use. | Rinse the whole unit for easy cleanup thanks to its IPX7 waterproof rating. |
What you can see: both titles fit under 80 characters (76 and 78), every bullet is under 25 words, and neither used the banned words. Terra put IPX7 in the title and wrote longer bullets (18–19 words); Sol led with metric-plus-imperial sizing and kept bullets to 13–18 words.
Round 3: find bugs in code
The input was a short Python snippet with three planted bugs: a mutable default argument, a pagination slice that drops one item, and division by zero on an empty list. Both models found all three, and both also flagged that the tags list gets shared with the caller. They differed in ranking and framing:
- Terra ranked the average-price denominator as the top bug and stated it should divide by total quantity.
- Sol first noted that the right denominator depends on whether you want average order value or average unit price, gave both fixes, then ranked by impact.
If you want a direct verdict, Terra's answer is quicker to act on. When the requirement itself is ambiguous, Sol lays out the ambiguity first.
Which one for which job
- Bulk extraction, classification, tagging: Terra, at half the cost per call.
- Batch jobs with reasoning turned off: Terra, the only one with a none effort.
- Live chat and support where a user is waiting: Sol, which finished first in all six tasks.
- Code review and requirements analysis with long answers: costs are close, so pick based on whether you want a direct verdict or stated assumptions.
- Not sure: run a batch on Terra, spot-check, and send the misses to Sol.
Switching cost
Both models use the same HiAPI key and the same POST /v1/responses endpoint; only the model field changes. The one catch is reasoning effort: Sol has no none, so change reasoning.effort from none to low when moving a Terra request over.
{"model": "gpt-6.1-sol", "input": [...], "stream": true, "store": false, "reasoning": {"effort": "low"}}
FAQ
Is Terra a cheaper Sol?
It's half the price with largely the same feature set: the same 1.05M context, strict JSON and tool calls. Terra adds a none effort; Sol answered faster in these tests.
Does every task cost half as much?
No. Short structured tasks came in at about half. For long answers, it depends on how many reasoning tokens the model spends, and on the Chinese copy task above, Terra cost more.
Can I use both?
Yes. One key calls both, so route by task.
Pricing details are in GPT-6.1 Terra API pricing, and integration code is in the GPT-6.1 Terra API guide. Copy the order email prompt above and paste it into the Terra model page to try it.
