fullauto.online

Models

GPT-6.1 Terra is half the price of Sol per token and about a quarter cheaper per task

OpenAI's new mid-tier model costs $1/$5 against Sol's $2/$10, but at max effort it writes 56% more output to finish the same benchmark. On Artificial Analysis figures read on 4 November 2026, the 50% list discount becomes about 24% per task. Where Terra fits, the 272K line that doubles a request's cost, and how to test it in an afternoon.

Published
04 Nov 2026
Reading
6 min
Class
models

half-life 60dfrom 4 Nov 2026

OpenAI's pitch for GPT-6.1 Terra, released on 26 October 2026, fits in one line: half the per-token price of GPT-6.1 Sol. The line is accurate. It is also the wrong number to budget with, because you pay for tasks and Terra spends more tokens to finish one. On the figures Artificial Analysis showed on 4 November, running both at maximum reasoning effort, Terra's price advantage over Sol shrinks from 50% to about 24%.

That still makes it a good deal for the right work. It just changes which work that is, and how you should set it up.

The list prices, and the line that doubles them

Per 1M tokensGPT-6.1 TerraGPT-6.1 Sol
Input$1.00$2.00
Cached input$0.05$0.10
Cache writes$1.25$2.50
Output$5.00$10.00
Context window / max input1,050,000 / 922,0001,050,000 / 922,000
Effort levelsnone, low, medium (default), high, xhigh, maxlow, medium (default), high, xhigh, max
Knowledge cutoff1 June 202630 April 2026

Source: OpenAI model pages for Terra and Sol, read 4 November 2026.

The row that catches people out is not in that table. On both models, a prompt above 272,000 input tokens is billed at twice the input and cache rates and 1.5 times the output rate, for the whole request. A million-token window does not mean a million tokens at the headline price. Take a single uncached call with 10,000 tokens of output:

Input sizeTerraSol
270,000 tokens$0.32$0.64
300,000 tokens$0.675$1.35

Our arithmetic from the published rates. Caching changes the totals but not the step.

So 11% more context costs roughly 110% more. For a long-running agent, that step matters more than the choice between these two models. If your harness can compact below 272K, or split the job among subagents, it is usually worth it. Sol's page adds that usage drawn from a Codex or ChatGPT Work subscription carries no surcharge; the Terra page we read did not include that sentence, so check before assuming the same.

Why half the price is not half the bill

At max effortGPT-6.1 TerraGPT-6.1 SolTerra vs Sol
Intelligence Index51554 points lower
Cost per index task$0.96$1.2724% cheaper
Output tokens across the index99.4M63.7M56% more
Time per index task12.8 min20.4 min37% less

Source: Artificial Analysis pages for Terra and Sol, read 4 November 2026. Percentages are ours.

The most visible explanation is the output-token row. At max effort Terra writes over half as many tokens again to get through the same index, and output is the most expensive token either model sells. You get a real saving, but a good part of it leaks back through verbosity.

Two warnings about that table. These pages show the max-effort variant only, and max is exactly the setting nobody should default to on a budget model. The numbers also move. Launch-day coverage quoted the same source at 48 and 52 on the index and $0.40 and $0.72 per task, a 44% gap. We cannot say from the public pages exactly why the figures changed. If a decision rests on one of these numbers, read it again on the day. That write-up also reports Artificial Analysis rating low and medium effort as Pareto-optimal and max as not, which fits the picture but which we could not confirm on the pages we read.

Read the 3× claim carefully

OpenAI says Terra “finishes tasks over 3× faster on average than both Astra and GPT-6.1 Sol” (announcement, 26 October 2026). On the independent index tasks above, at max effort, the gap is about 1.6 times. The two figures do not contradict each other, because they come from different tests and OpenAI is pitching Terra at computer use: reading screens, clicking and moving between apps. But do not carry 3× over to your coding agent. The same goes for OpenAI's claim, reported by The Next Web, that Terra finishes at roughly a ninth of the cost of Claude Opus 5.5. That is OpenAI's own testing and has not been independently checked.

Where it fits

OpenAI's own guidance is unusually direct: “For in-depth research or complex coding, choose GPT-6.1 Sol or GPT-6 Astra.” Early forum reports point the same way. One developer on the announcement thread says Terra picks the wrong tool more often than Sol and lags after errors. That is one person's experience, but tool selection is cheap to test, and it is where a failed run gets expensive.

  • Good candidates: computer-use automation, form filling, document and office work, and high-volume agent steps where a retry costs little.
  • Keep on Sol or Astra: hard multi-file coding, long research runs, and anything where one bad tool call wastes an hour of downstream work.
  • Worth a look at the lower end: none effort is available on Terra and not on Sol, and a forum user reports it also unlocks temperature and top_p. If you are classifying or extracting at volume, test that before reaching for a dearer tier.

One thing not to miss if you run agents with access to real systems: The Next Web reports that OpenAI places Terra at “Critical” for cybersecurity skills under its Preparedness Framework, with the same protections as GPT-6 Astra. The cheaper tier is not the less capable tier here. Budget models get handed broad automation because they are cheap, so give this one the same containment as anything else you run.

How to test it in an afternoon

  1. Start low. Run your existing evals at low and medium before you touch high. The only change you need is the model and effort:
    {
      "model": "gpt-6.1-terra",
      "reasoning": { "effort": "low" }
    }
  2. Score cost per accepted task. Total spend, retries included, divided by results you would actually ship. A cheap model that needs two attempts has lost its discount before you count anyone's review time.
  3. Log tool-call errors separately. That is the failure mode being reported, and an overall pass rate can hide it.
  4. Plot context size per request. Count how many calls cross 272K. If it is more than a few, fix that first. It will save more than a model swap.
  5. Route, do not replace. Send the routine work to Terra, and escalate to Sol when the first pass fails. Tuned Sol pipelines with stable evals may not repay migration unless the volume is large.

None of this makes Terra less interesting. A model four points below Sol at a quarter less per task, with a newer knowledge cutoff, is a strong default for a lot of agent work. Just model the bill on tasks rather than tokens. For how these models compare with the rest of the field, see the model leaderboard and the Coding Agent Index reset, and for the cache economics behind long sessions, Claude Code's 54-minute rule.

Sources