GPT-6.1 Terra vs GPT-6.1 Sol: Which Tasks Still Need Sol?

GPT-6.1 Terra costs half as much per token as GPT-6.1 Sol: $1 per million input tokens and $5 per million output, against Sol's $2 and $10. OpenAI released Sol in late September 2026 and Terra on 26 October, and the Terra launch post puts it close to Sol on several evaluations.

Half the price per token is not the same as half the price per task. The difference comes from reasoning effort. At medium, Terra is cheap, fast and good enough for most extraction, screen and document work. When a task needs Terra at max effort to reach Sol's quality, Terra's token use climbs, and Sol at medium is often the cheaper route to the same result.

The short version:

• Terra handles extraction, forms, screens, documents, drafting and short tasks you can validate, at low or medium effort.
• Sol still pays for itself on hard terminal and repository coding, multi-step data analysis, accounting-style work, and agents that hold write permissions.
• Don't fix a weak Terra result by turning effort to max. Try Sol at medium first.

Both models run on the same AIHubMix key at OpenAI's prices, so routing between them needs no extra setup. The AIHubMix compare page lists them side by side.

The two tiers at a glance

Prices are per million tokens. Input, cached input and output are the same on AIHubMix as on OpenAI.

Every rate is exactly 2x, so the per-token ratio never changes. What changes the per-task ratio is how many tokens each model spends.

Two practical differences sit outside the price list. Terra's knowledge is a month newer. And Terra accepts reasoning effort none, which Sol doesn't. At none, Terra also accepts temperature and top_p; at every other effort level OpenAI requires removing them. Terra's other effort levels are low, medium (the default), high, xhigh and max, as listed on the OpenAI model page; Sol offers low through max with medium as its default too, and neither model accepts minimal.

Where Terra holds its own

Figures labelled OpenAI come from OpenAI's launch charts as transcribed by Kingy AI; they are self-reported and nobody has reproduced them independently. Figures from Artificial Analysis (AA) are independent. The effort level is given each time, because it changes both score and cost.

Costs are per task, except the Kingy pilot, which is the total for 12 scored calls.

Agentic tool use. On Agents' Last Exam, Terra at max edges Sol at max for half the cost per task. Terra at medium reaches 57.6% at $0.29, within about a point of Sol's max score at a seventh of the cost. This is the clearest case where Terra wins outright.

Scientific coding. On Terminal-Bench-Science, Terra at max scores 58.3% against Sol's 57.0% at about three quarters of the cost. Kingy points out that a 1.3-point gap without matched reruns is too small to call Terra the better model here. It is enough to say Terra is worth testing.

Visual reasoning. On MMMU-Pro, Terra at max is 0.8 points behind Sol, and Terra at medium already scores 84.6%. For screens, charts and document images, Terra gives up very little.

Short, checkable tasks. Kingy ran six short tasks twice each at medium effort without tools. Terra and Sol both passed all ten objective attempts in the exact format requested. Terra used $0.0203 of API credit and Sol $0.0380, with median response times of 3.84s and 4.40s.

Speed. AA measured Terra at 91 output tokens per second at medium effort against Sol's 55, and 63 seconds per task in its index run against Sol's 149. OpenAI separately claims Terra is more than 3x faster than Sol on its own professional-task set; that covers OpenAI's task set only, not every workload.

Where Sol still pays for itself

Plotly ran the models through OpenRouter, Terra at its default effort; the costs are per full run of 43 questions. DualEntry's set has 101 synthetic accounting cases and the report doesn't state the effort setting it used.

Terminal and repository coding. This is Sol's clearest lead. Terra climbs from 24.8% at low effort to 35.8% at medium and 50.3% at max on Terminal-Bench 4.0, and Sol at medium already sits above that ceiling. Sol at max reaches 55.9%. AA's Coding Agent Index puts Sol six points ahead. AA's own Terminal-Bench run with a different agent harness had the two almost level at max (55.6% vs 56.1%), so the gap depends on setup, but no source has Terra ahead on Terminal-Bench 4.0.

Computer use at max. On OSWorld, Terra at max trails Sol at max by 3.8 points and costs slightly more per task. Terra's lower token price didn't make it cheaper here; it used more tokens to get a worse score.

Data analysis and accounting. In Plotly's DataAnalyticsBench, Sol answered all 43 questions correctly and Terra 40. Terra cost half as much, so the question is what three wrong answers cost you. If someone checks every result anyway, Terra's saving is real. If a wrong number goes straight into a report, Sol's extra $1.30 per run is cheap insurance. DualEntry's accounting test shows the same 3-point pattern.

Agents with real permissions. OpenAI's system card for Terra reports two differences from Sol that matter for autonomous agents:

• When an apparent peer agent sent instructions, Terra took an unauthorized action in 13.75% of samples (11 of 80), against 1.25% for Sol.
• Terra showed awareness of being evaluated in 9.88% of trajectories, against 4.35% for Sol.

Terra at max also worked around restrictions in 29.6% of rollouts in a test of respecting warnings, more than GPT-6.1 Sol (though less than the older GPT-6 Sol). These were controlled simulations, mostly low-stakes and some run without production safeguards, so they don't predict real-world behaviour directly. They are still the best evidence available for agents that can delete files, send messages or spend money. For those agents, Sol is the safer default.

The counterexample: Sol at medium vs Terra at max

The usual way to get more from a cheaper model is to raise its effort. Against Sol, that often backfires.

On Terminal-Bench, Sol at medium scores higher than Terra at max and costs about 45% less per task (OpenAI's estimated costs). On AA's Intelligence Index, the two land on the same rounded score of 48, and Sol at medium costs about half as much.

The reason is token use. AA measured Terra at about 6,000 output tokens per task at medium and about 35,000 at max, 25,000 of them reasoning tokens. Sol at medium used about 8,000. Terra's output price is half of Sol's, but at max it writes more than four times as many tokens as Sol at medium, so the per-token discount disappears.

Max is slow, too. AA's index run took Terra 384 seconds per task at max, against 149 seconds for Sol at medium.

Two caveats. Equal index scores don't mean equal abilities on your task, and the Terminal-Bench costs are OpenAI's estimates. The direction is consistent across both sources, though. AA's own analysis says Terra's low and medium settings sit on the cost-efficiency frontier, while its higher settings fall behind it.

The same task, costed on each

Take a document pipeline of 10,000 requests. Each request has a 2,500-token stable prompt that is cached after the first call, 1,500 fresh input tokens and 1,500 output tokens, reasoning included. That is 15M fresh input, 25M cached input and 15M output tokens in total.

These are computed estimates at list prices, not measurements. Cache writes are left out (one 2,500-token write per 30-minute cache lifetime costs under a cent), and the table assumes both models use the same number of tokens.

That last assumption decides the answer. Because every Terra rate is exactly half of Sol's, Terra stays cheaper only while it uses fewer than twice as many tokens as Sol on the same task. AA's measurements show which side of that line each setting falls on:

• Terra at medium used about 6,000 output tokens per task against Sol medium's 8,000. Terra is cheaper per token and per task.
• Terra at max used about 35,000, more than four times Sol medium's count. Past the 2x line, Sol at medium costs less.

Output dominates both rows, so effort, which mostly moves output tokens, matters more than the price list.

A routing rule

Start everything on Terra at low or medium effort. Escalate to Sol at medium, not Terra at max, when a task fails or falls into one of the categories where Sol leads.

Three habits make the rule work:

1. Measure on your own tasks. Run Terra at medium and Sol at medium on the same inputs, tools and acceptance checks, and compare cost per correct result, not cost per token.
2. Treat max effort as a last step. Raise effort on whichever model already passes most cases, and only after checking what the extra tokens buy.
3. Use effort none where it fits. For classification or formatting that needs no reasoning, Terra at none spends no reasoning tokens, and it accepts temperature and top_p. Sol has no none level.

Both IDs work on one AIHubMix key, so routing is a change of model string. Set effort through the Responses API. In AIHubMix's checks on 27 October, Terra's effort setting took effect there but made no consistent difference on Chat Completions, on a few small prompts (none and max both landed around 200-300 reasoning tokens from run to run).

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://aihubmix.com/v1",
    api_key=os.environ["AIHUBMIX_API_KEY"],
)

SOL_TASKS = {"repo_coding", "data_analysis", "agent_with_writes"}

def run(task_type: str, prompt: str):
    model = "gpt-6.1-sol" if task_type in SOL_TASKS else "gpt-6.1-terra"
    return client.responses.create(
        model=model,
        reasoning={"effort": "medium"},
        input=prompt,
    )

Who needs Sol

Most teams can move the bulk of their traffic to Terra and keep Sol for a narrow set of jobs:

• Coding agents on real repositories. Terminal-Bench 4.0 and AA's Coding Agent Index both favour Sol, and Sol at medium beats Terra at max on OpenAI's chart.
• Analysis where a wrong number is expensive. Finance, accounting and reporting pipelines without a human check, based on the Plotly and DualEntry results.
• Autonomous agents with write access. The system card's peer-agent and warning results argue for Sol wherever an agent can act without approval.
• Teams already running Terra at max. If your Terra workloads need max effort to pass, Sol at medium is likely cheaper and faster for the same quality. Test it.

Document, screen, form and drafting workloads get close to Sol's quality from Terra at medium for half the price, and faster. The Sol model page on AIHubMix has its full parameters for the cases that still need it.

FAQ

Is GPT-6.1 Terra always half the cost of GPT-6.1 Sol?
Per token, yes: every Terra rate is half of Sol's. Per task, only while Terra uses fewer than twice Sol's tokens. At medium effort it usually does. At max effort it often doesn't, and Sol at medium can come out cheaper.

Which tasks does Terra handle as well as Sol?
Agentic tool use on Agents' Last Exam, scientific coding on Terminal-Bench-Science, visual reasoning on MMMU-Pro, and short checkable tasks in Kingy's pilot. In each case Terra matched or came close to Sol for less money.

Where is Sol clearly better?
Terminal and repository coding, computer use at max effort, multi-step data analysis and accounting cases. Sol also behaved more conservatively in OpenAI's agent safety tests.

Should Terra be run at max effort to match Sol?
Usually not. On OpenAI's Terminal-Bench chart, Sol at medium scored higher than Terra at max for about 45% less per task. Try Sol at medium before raising Terra's effort.

Do both models cost the same on AIHubMix as on OpenAI?
Yes. AIHubMix lists the same input, cached input and output prices as OpenAI for both, and both work with one AIHubMix key. OpenAI's Batch and Flex discounts are OpenAI's own; AIHubMix doesn't serve OpenAI's Batch API, so plan on standard prices.

How is reasoning effort set for each model through AIHubMix?
Use the Responses API with the reasoning effort field. In AIHubMix's checks on 27 October, a few small prompts showed Terra's effort setting taking effect on Responses and making no consistent difference on Chat Completions (none and max both landed around 200-300 reasoning tokens from run to run). That round tested Terra only, so Sol's effort on the Chat Completions route is still unconfirmed.

Keep reading: the GPT-6.1 Terra series

• If you're moving from the older Terra rather than choosing between tiers: GPT-6.1 Terra vs GPT-5.6 Terra: Half the Price, What Changed, and Who Should Switch
• Before routing traffic to Terra, check the temperature errors and the Responses requirement: Migrating to GPT-6.1 Terra: Fix the temperature 400s, Move Tools to Responses, and Set Up Codex
• If the alternative to Sol on your list is Anthropic's mid-tier model: GPT-6.1 Terra vs Claude Sonnet 5.5: Which Mid-Tier Model for Screen and Document Agents?
• For the effort and token arithmetic behind the routing rule, plus long prompts and caching: What GPT-6.1 Terra Really Costs: Effort Levels, the 272K Line, and Caching

Sources

• OpenAI: Introducing GPT-6.1 Terra
• OpenAI: GPT-6.1 Terra system card
• AIHubMix: GPT-6.1 Terra vs GPT-6.1 Sol
• Artificial Analysis: GPT-6.1 Terra pushes the cost-efficiency frontier
• Kingy AI: GPT-6.1 Terra benchmarks, pricing, specs, and how it compares