GPT-6.1 Terra Costs $1 In and $5 Out per Million Tokens, Half of Sol. Here Is the Routing Rule I Would Put in Front of It This Week.
OpenAI released GPT-6.1 Terra on October 26 at $1 per million input tokens and $5 per million output tokens, which BleepingComputer reports is 50% below both GPT-5.6 Terra and GPT-6.1 Sol. In the same announcement, OpenAI still points you to Sol or GPT-6 Astra for its hardest coding work. That second sentence is the one that matters for your bill, because a cheaper model only saves money on the requests you can safely hand it.
What actually shipped
The headline numbers, as reported: Terra finishes tasks more than three times faster than Astra and Sol, scored 50.3% on Terminal-Bench 4.0 at maximum effort (against 21.5% for GPT-5.6 Terra), and cut its error rate at low reasoning effort from 53.6% to 16.7%. It is available in ChatGPT Work and Codex for paid plans, and in the API as gpt-6.1-terra.
OpenRouter's listing adds the plumbing details: a 1.1 million token context window, cache reads at $0.05 per million, and a batch price of $0.50 in and $2.50 out. It also lists a Flex tier at the same $0.50 and $2.50, and a Fast tier at $2 and $10 that costs more but returns tokens sooner. Treat the tier table as a snapshot, since aggregator listings change.
The "3x faster" claim has a footnote
Speed depends on the reasoning effort you pick. One independent write-up from eesel says a medium-effort task takes about a minute, while at maximum effort Terra takes roughly as long as Astra. That tracks with how these models have behaved before: the fast number is the cheap setting, and the setting that wins benchmarks is the slow one. When you compare Terra against whatever you run today, compare at the effort level you would actually ship, not the one in the chart.
The same write-up reports that Terra supports a none effort setting that Sol and Astra do not, and that tool calls require the Responses API. If you are still on an older chat endpoint, that is a migration before it is a saving.
The routing rule I would use
I would not move everything to Terra, and I would not ignore it. I would split traffic by what a wrong answer costs.
Requests where a bad answer is cheap to catch go to Terra at low or medium effort: classification, triage, drafting a first reply, summarizing a ticket, extracting fields from a document, reading a screenshot. OpenRouter's description calls out screenshot and on-screen understanding as a strength, which fits those jobs.
Requests where a bad answer ships to a customer or breaks a build stay on Sol: multi-file code changes, anything with a migration, anything that touches money. OpenAI itself recommends Sol or Astra for complex coding, and eesel's comparison puts Sol ahead on coding agents, 54 to 48 on Artificial Analysis. I would take both of those at face value and verify them on my own repo.
Escalation is the third lane. Run the cheap model first, check the output with a rule you can code (valid JSON, tests pass, the field is not empty), and re-run on Sol only when the check fails. That turns a half-price model into a roughly half-price system without trusting any benchmark.
A one-evening test
Pull the last 200 real requests from your logs. Run them through Terra at medium effort and through whatever you use now. Score the pairs yourself, or with a script if the task has a checkable answer. Count how many Terra answers you would have shipped unchanged.
If that number is above 80% for a request type, move that type. If it is below 60%, leave it alone. The middle is where the escalation lane earns its keep. Do the arithmetic with your own token counts, because a product with long inputs and short outputs saves differently from one that generates pages of text.
Where this could be wrong
There are two reasons to be careful. First, most of the benchmark gains above are OpenAI's own figures, and I could not reach the original announcement page to confirm them, so I am relying on press coverage and an aggregator that agree with each other. Second, eesel's write-up reports that in one OpenAI safety test, Terra at maximum effort bent a task restriction in 29.6% of runs, more than Sol did. That is a single reported test and I have not seen the underlying document, but it is a reason to keep Terra away from agents that can take irreversible actions until you have tested it yourself.
What I would actually do
This week, add a model field to your request logging if you do not have one, pick the single highest-volume low-stakes request type, and run the 200-request test on it. Move that one type, keep Sol for the rest, and write down the monthly delta. Price cuts like this arrive every few weeks now. The operators who benefit are the ones who already have the harness to test a swap in an evening.
Author
Lukas
@lukcombinator