If your team sends routine agent and computer-use work to GPT-6.1 Sol, OpenAI's new GPT-6.1 Terra will probably do most of it for less per task if run at low or medium reasoning effort. Terra is priced at $1 per million input tokens and $5 per million output tokens, half of Sol's $2 and $10. Artificial Analysis scores it 48 on its Intelligence Index against 52 for Sol. The catch is that Terra at maximum effort burns about twice the tokens, so the per-token discount does not carry through to the bill at the top setting.
OpenAI released the model on October 26, 2026. It is rolling out to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and developers can call it in the API as gpt-6.1-terra. It succeeds GPT-5.6 Terra and sits below Sol and Astra but above Luna in OpenAI's four-tier lineup.
What GPT-6.1 Terra Costs Against Sol and GPT-5.6 Terra
GPT-6.1 Terra is the cheapest OpenAI model above Luna, and on list price it undercuts both its predecessor and Sol. Here are the published API rates, per million tokens, as of October 26, 2026:
| Model | Input | Cached input | Cache writes | Output |
|---|---|---|---|---|
| GPT-6.1 Terra | $1.00 | $0.05 | $1.25 | $5.00 |
| GPT-6.1 Sol | $2.00 | $0.10 | $2.50 | $10.00 |
| GPT-5.6 Terra | $2.00 | $0.20 | $2.50 | $12.00 |
Sources: OpenAI's model pages for GPT-6.1 Terra, GPT-6.1 Sol and GPT-5.6 Terra.
Three details on those pages change the math. Terra's cache writes cost $1.25, a quarter more than a normal input token, and that rate buys 30-minute retention; 12-hour retention costs $2.50. A prompt prefix you write once and reuse twice is still a win at $0.05 a cached read, but a prefix that is never reused costs more than sending it uncached. If you are new to that trade-off, our prompt caching cost breakdown walks through where hit rates collapse.
Second, the GPT-5.6 Terra page says prompts over 272K input tokens are billed at 2x input and 1.5x output for the whole request. The GPT-6.1 Terra page carries the same surcharge above 272K input tokens, so long-document jobs still pay the premium.
Third, the context window is 1,050,000 tokens with 128,000 maximum output, the same as Sol's, so moving between the two does not force a change in how you chunk documents.
How Close Terra Gets to Sol on Independent Benchmarks
On Artificial Analysis's numbers, Terra at maximum effort lands 4 points behind Sol and costs $0.40 per index task against Sol's $0.72. Those figures come from the Terra model page (index 48, rank 25 of 221) and the Sol model page (index 52, rank 13 of 221). That makes Terra about 44% cheaper per task, which is less than the 50% the price sheet implies.
The reason is token volume. Artificial Analysis logged 130 million output tokens for Terra to run its index, against 67 million for Sol. At maximum effort Terra thinks for almost twice as long to reach a lower score. Half the price times twice the tokens is close to break-even on output, and the input savings are what leave Terra ahead.
That is why reasoning effort is the setting that decides this migration. According to Trending Topics' summary of the Artificial Analysis results, the benchmarker rates low and medium effort as Pareto-optimal and says Terra at maximum effort is not. The same write-up reports gains of 7 to 13 index points over GPT-5.6 Terra at the lower effort levels, and a per-task cost drop from $1.40 to $0.40 at maximum effort.
On agentic coding the gap to Sol is wider. Trending Topics reports Artificial Analysis's Coding Agent Index v2.0 at 48 for Terra, 54 for Sol and 31 for Luna. That matches OpenAI's own advice: "For in-depth research or complex coding, choose GPT-6.1 Sol or GPT-6 Astra," the launch post says.
Does Terra Hallucinate Less Than GPT-5.6 Terra?
Yes, by a wide margin on two separate tests, one independent and one from OpenAI. On Artificial Analysis's AA-Omniscience benchmark, Terra guesses in 38% of cases where it does not know the answer, down from 88% for its predecessor. OpenAI's own trick-question test shows the error rate at low effort falling from 53.6% to 16.7%, and to 4.3% at maximum effort, which 9to5Mac reports nearly matches Astra and Sol.
The second figure is a vendor claim. The Next Web notes that OpenAI's comparison against Claude Opus 5.5, which put Terra at roughly a ninth of the cost, and its other launch benchmarks "come from OpenAI's own testing and have not been independently checked." Treat the Artificial Analysis numbers as the ones to plan against and the OpenAI numbers as direction.
If GPT-5.6 Terra's guessing rate is the reason you kept a human reviewer on an extraction or form-filling flow, re-test that decision on this release. At 88%, the old model filled most blanks with a plausible guess; at 38%, it still guesses on more than a third of the questions it cannot answer, so keep a reviewer on anything that moves money.
What Terra Is Built For: Screens, Forms and Speed
OpenAI positions Terra as its computer-use model for everyday office work. It can look at a screen, understand what it shows, and act by clicking, typing and moving between apps, with example uses including form filling, drafting proposals and reading technical drawings. OpenAI also says it finishes tasks over 3x faster on average than both Astra and GPT-6.1 Sol. Artificial Analysis lists Terra's output speed as unknown, so that speed figure has no independent check yet.
Two things to weigh before you point it at production screens. Terra reaches the "Critical" level for cybersecurity skills under OpenAI's Preparedness Framework and ships with the same protections as GPT-6 Astra. And computer use on logged-in sessions is a permissions problem before it is a model problem; we covered the admin gaps in the ChatGPT Chrome extension's work mode. For deterministic vendor portals, a Playwright script still beats any model on cost and reliability, as our browser automation comparison found.
Developers get one setting Sol lacks. Terra's model page lists reasoning effort levels of none, low, medium (default), high, xhigh and max, while Sol's page lists low through max with no "none" option. A developer on OpenAI's forum notes that "none" also unlocks temperature and top_p, which matters for classification and extraction jobs that need sampling control.
Rate limits match Sol at each tier: 5,000 requests and 1 million tokens per minute at Build, up to 15,000 requests and 40 million tokens at Grow, so rate limits give you no reason to stay on Sol.
Where Sol Still Earns Its Price
Keep Sol on the hardest agentic coding and multi-step research. A 6-point gap on Artificial Analysis's coding agent index is large enough to show up as failed pull requests and retried runs, and a retried run at Terra prices can cost more than a first-pass success at Sol prices. Our GPT-6.1 Sol DevDay analysis covers how Sol itself trades against Astra on that axis.
The strongest case for staying put is consistency. If you have a working Sol pipeline with tuned prompts and stable eval scores, a 44% per-task saving has to cover the engineering hours to re-tune prompts and re-run evals, plus the risk that a regression slips through. For a workload spending a few hundred dollars a month, that may not pay. For a computer-use fleet running thousands of sessions a day, it almost certainly does.
What to Do About GPT-6.1 Terra
This Week:
- Pull last month's OpenAI usage by model and effort level, and list every workload running on GPT-5.6 Terra or on Sol at low or medium effort. Those are your migration candidates.
- Re-run your existing eval suite on
gpt-6.1-terraat low and medium effort for each candidate, and record cost per successful task alongside the pass rate. Do not test only at max; that is the setting where Terra loses most of its price advantage. - Check which workloads run on Codex or ChatGPT Work seats. OpenAI says your usage allowance goes further with GPT-6.1 Terra than with GPT-5.6 Terra, so routine Codex tasks that hit seat limits are a quick win.
This Month:
- For any flow that reuses a long system prompt or document, confirm your cache hit rate is high enough to justify the $1.25 cache-write rate before you rely on caching for the savings.
- Re-test the human review steps you added because GPT-5.6 Terra guessed on unknowns. With a lower guess rate on AA-Omniscience, some of those checks may be sampling instead of reviewing every item.
- Add Terra as a routing tier between Luna and Sol in your gateway, with Sol kept as the fallback for coding tasks that fail Terra's first pass. If you route through a managed gateway, our look at Cloudflare's Auto Router shows why you should own the routing rule rather than delegate it.
The Bottom Line
This is the same pattern as GPT-6 Sol's launch in September: a halved token price, with the headline benchmark run at an effort level where token volume eats much of the saving. Budget on cost per successful task at the effort level you actually run, since that is what shows up on the invoice. Terra's numbers at low and medium effort are good enough to move most routine agent and computer-use work off Sol, and Artificial Analysis, as Trending Topics reports it, rates those two settings as the Pareto-optimal ones.
Run the eval at medium effort before you change the default model string.
Continue Reading
- GPT-6.1 Sol's $5.47 Task Scores 11 Points Below Astra's $23.80
- GPT-6 Sol Halves the Token Price but Benchmarks It at Top Effort
- OpenAI Pulled GPT-6.1 Astra Over a Flaw GPT-6 Astra Still Has
- Claude Sonnet 5.5 Costs More Than Opus 5.5 at the Same Score
- Google Cloud Lets Codex Users Run OpenAI Models on Its Platform
- Prompt Caching Savings: One Timestamp Turns 80% Into 6%