OpenAI now sells two GPT-6.1 models instead of one flagship: GPT-6.1 Sol, released on 29 September 2026, and GPT-6.1 Terra, released on 26 October 2026. They share a family name but not a job description, and the price gap is wide enough to change how you architect a product.
Short version: Terra is the model most teams should test first. Sol earns its price only when a failed task costs more than the tokens.
For the three-vendor picture, see the full AI model comparison for 2026.
GPT-6.1 Sol vs Terra at a glance
| Question | GPT-6.1 Sol | GPT-6.1 Terra |
|---|---|---|
| Released | 29 September 2026 | 26 October 2026 |
| Direct price | $1 input / $5 output | $0.50 input / $2.50 output |
| Position | Highest intelligence in the family | Fast, lower-cost model |
| Computer use | Yes | Yes, precise computer use |
| Best for | Agentic coding, professional work | Support, content, analysis, visual reasoning |
Terra is not a stripped-down toy. What Sol adds is depth on hard, multi-step work, and that depth is worth paying for only when the work demands it.
Pricing: what you actually pay
| Model | OpenAI direct | Amazon Bedrock | Azure AI Foundry |
|---|---|---|---|
| GPT-6.1 Sol | $1 / $5 | $2 / $10 | $2 / $10 |
| GPT-6.1 Terra | $0.50 / $2.50 | $1.10 / $5.50 | $1.20 / $6 |
Prices are list prices per million tokens, checked on 3 November 2026. Cached, batch and priority tiers differ, so confirm the rate for the tier you will use.
A worked example makes the gap concrete. Suppose you process 50 million input tokens and 10 million output tokens a month:
| Setup | Monthly cost |
|---|---|
| Terra, OpenAI direct | 50 × $0.50 + 10 × $2.50 = $50 |
| Sol, OpenAI direct | 50 × $1 + 10 × $5 = $100 |
| Terra, Bedrock | 50 × $1.10 + 10 × $5.50 = $110 |
| Sol, Bedrock | 50 × $2 + 10 × $10 = $200 |
Two things stand out. Sol costs about twice Terra for the same volume. And buying through a cloud marketplace roughly doubles the direct price again, so the same model can cost four times more depending on where you run it.
The long-context tier
On at least one provider, Sol requests above 272,000 prompt tokens move to a higher rate: $4 input and $15 output per million tokens. If your prompts carry large repositories or long transcripts, price that tier rather than the headline rate, and test with your real document sizes.
Where Sol is genuinely better
Agentic coding
Sol is positioned for agentic coding and computer use on professional work. That means tasks with several steps, tool calls and a plan that has to survive errors and retries. If your agent writes code, runs tests and fixes its own failures, this is the tier to test.
High-stakes work
If a wrong answer triggers a refund, a compliance issue or a support escalation, the token difference is small compared with the cost of the mistake. That is the calculation where Sol pays for itself.
Codebase-scale analysis
Large refactors, migrations and repository-wide reviews need a model that can hold a lot of context and keep the thread across many turns. Sol is the better candidate for that class of work.
Where Terra wins
Volume
At half the input price and half the output price, Terra handles the same budget with twice the tokens. For chat, summarisation, extraction, classification and drafting, that is the whole argument.
Computer use
Terra includes precise computer use. In June, computer use was a reason to choose a different vendor entirely. It is now a reason to choose the cheaper OpenAI model.
Latency
Terra is the faster model. For customer-facing features, a marginal reasoning gain is usually worth less than a faster response.
Visual reasoning
Terra is described as a fast visual reasoning model, which covers screenshots, dashboards and document images without paying flagship rates.
Migration from GPT-5.6
The family shape has not changed. Sol is still the flagship and Terra is still the workhorse. The prices have changed substantially.
| Model | June 2026 | November 2026 |
|---|---|---|
| Sol | $5 / $30 | $1 / $5 |
| Terra | $2.50 / $15 | $0.50 / $2.50 |
A straight swap is roughly 80% cheaper at list price for the same tier. That is unusual, and it is worth re-running your cost model even if your usage has not changed.
Do not assume behaviour is identical. Treat the new models as new models: re-run your evaluation set, check refusal behaviour, and confirm that your tool-calling and structured-output formats still work the way your product expects.
Cost per successful task
Token price is not the unit that matters. Cost per completed task is. Here is an illustrative comparison:
| Model | Cost per task | Success rate | Expected cost with a $3 human fallback |
|---|---|---|---|
| Terra | $0.004 | 92% | 0.004 + 0.08 × 3 = $0.244 |
| Sol | $0.008 | 96% | 0.008 + 0.04 × 3 = $0.128 |
Even at double the token price, Sol wins when it halves the failure rate and a failure costs real money. Run this table with your own numbers before you decide that the cheaper model is the cheaper choice.
Common mistakes
- Choosing Sol for everything “to be safe”, then paying twice for work Terra handles.
- Choosing Terra for hard agent work, then blaming the model when it loses the thread.
- Comparing token prices across providers without checking the cloud marketplace markup.
- Ignoring batch and cache discounts, which often matter more than the model choice.
- Assuming a model refresh is a drop-in replacement without re-testing prompts.
How to choose
Start with Terra if:
- Most of your traffic is routine: support replies, summaries, extraction, classification.
- You need computer use at the lowest cost.
- Latency matters for the user experience.
- You are cost-sensitive or running high volume.
Move to Sol if:
- Tasks need long, multi-step execution with tool calls.
- Failures are expensive, slow or visible to customers.
- You are doing codebase-scale analysis or migration work.
- Output quality is the product itself.
Use both. The practical pattern is a router: Terra handles the bulk, and low-confidence or high-value requests escalate to Sol. You get flagship quality where it matters and Terra economics everywhere else.
Questions to ask before you switch
- Which of our tasks actually fail today, and why?
- What does a failure cost in money, time or trust?
- Which provider will we bill through, and at what markup?
- How much of our input is repeated context that could be cached?
- Can any of our work move to batch pricing?
FAQ
Is GPT-6.1 Sol worth the extra cost?
Only when a failed task is expensive. For routine work, Terra handles the same volume at half the price and is the better default.
Does GPT-6.1 Terra support computer use?
Yes. Terra includes precise computer use, so you no longer need the flagship model to automate a browser or desktop workflow.
Which is cheaper on Azure?
Terra, at $1.20 input and $6 output per million tokens, against $2 and $10 for Sol. The gap is smaller than on direct pricing but still substantial.
Is GPT-6.1 a drop-in replacement for GPT-5.6?
Re-test rather than assume. The family structure is the same and the prices are much lower, but prompts, tool schemas and refusal behaviour should be checked on your own data.
Next step
If you are choosing between the two for a product you are building, our AI app development cost guide covers the build side. For a short evaluation on your own data, contact 24Bit System.
Related reading: Claude Fable 5.1: what changed and Gemini 3.6 Flash: pricing and use cases.
