DEV Community

Cover image for Routing GPT-6.1 Terra and GPT-6.1 Sol: The 2x Token Line
AIHubMix
AIHubMix

Posted on Edited on

Routing GPT-6.1 Terra and GPT-6.1 Sol: The 2x Token Line

#ai #api #llm #openai

Terra's price list is exactly half of Sol's: $1 per million input tokens and $5 per million output, against $2 and $10. A router built on that ratio alone will be wrong on a predictable set of tasks. On OpenAI's Terminal-Bench 4.0 chart, Sol at medium effort scores 52.7% for $1.02 a task, while Terra at max scores 50.3% for $1.84 — a better result from the more expensive tier at 45% less per task.

The number that decides a tier comparison is not price per token. It is tokens per task, and the safe boundary is easy to state: because every Terra rate is exactly half of Sol's, Terra stays cheaper only while it uses fewer than twice as many tokens to finish the same job.

The arithmetic behind the routing decision

Reasoning effort moves output tokens, and output is where the money is. Artificial Analysis (AA) measured Terra at roughly 6,000 output tokens per task at medium effort and roughly 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.

Terra max Sol medium
Terminal-Bench 4.0 50.3% 52.7%
Terminal-Bench 4.0 cost per task $1.84 $1.02
AA Intelligence Index 48 48
AA cost per task $0.40 $0.21
AA seconds per task 384 s 149 s

Terminal-Bench costs are OpenAI's estimates; the AA rows are AA's measurements from its index run. The direction agrees across both sources, and AA's own analysis puts Terra's low and medium settings on the cost-efficiency frontier while its higher settings fall behind it.

Two caveats before this becomes a rule. Equal rounded index scores don't mean equal ability on your task. And the index score is one number over a mixed suite — a workload that sits entirely in the part of the suite Terra fails will not behave like a 48-vs-48 tie.

Prerequisites: what the API gives you

Both IDs run on one AIHubMix key at OpenAI's prices, so a routing change is a change of model. Effort is set on the Responses API. In AIHubMix's checks on 27 October, Terra's effort setting took effect on Responses 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) — set effort there and don't trust it until you've confirmed it on your own prompts.

Terra's effort levels are none, low, medium (the default), high, xhigh and max, per the OpenAI model page. Sol offers low through max with medium as its default too, and neither model accepts minimal. Terra accepts temperature and top_p only at none; at every other effort level OpenAI requires removing them.

GPT-6.1 Terra GPT-6.1 Sol
Input $1.00 $2.00
Cached input $0.05 $0.10
Cache write (30 min) $1.25 $2.50
Output $5.00 $10.00
Knowledge cutoff 1 Jun 2026 30 Apr 2026
Effort none Yes No

Both models were released within a month of each other — Sol in late September 2026, Terra on 26 October — and the Terra launch post places it close to Sol on several evaluations. Prices are per million tokens. Input, cached input and output match OpenAI's; AIHubMix's Terra page lists no cache-write line (its Sol page does). OpenAI's Batch and Flex discounts are OpenAI's own, and AIHubMix doesn't serve OpenAI's Batch API, so budget on standard rates.

Where Terra at medium is the right default

Figures below marked OpenAI come from OpenAI's launch charts as transcribed by Kingy AI: self-reported, not independently reproduced. The AA figures are independent. Effort level is stated for every row because it changes both score and cost.

Workload Terra Sol Source Call
Agents' Last Exam, max 60.0% at $1.01 58.6% at $2.00 OpenAI Terra
Agents' Last Exam, Terra medium 57.6% at $0.29 – OpenAI Terra
Terminal-Bench-Science 0.1, max 58.3% at $4.15 57.0% at $5.47 OpenAI Test both
MMMU-Pro, max 85.5% 86.3% OpenAI Either
Terminal-Bench 4.0 50.3% (max) 52.7% (medium) OpenAI Sol
OSWorld 2.1 offline, max 74.2% at $1.19 78.0% at $1.15 OpenAI Sol
Plotly data analysis 40/43 at $1.31 43/43 at $2.61 Plotly Sol if unchecked
DualEntry accounting 78.2% 81.2% DualEntry Sol if unchecked

Agentic tool use is the clearest Terra win: at max it beats Sol by 1.4 points at about half the per-task cost, and at medium it reaches 57.6% for $0.29, within about a point of Sol's max score at a seventh of the price. On MMMU-Pro, Terra at max is 0.8 points behind Sol, and Terra at medium already scores 84.6% — small change for screens, charts and document images. Kingy's own reading of the Terminal-Bench-Science result is that a 1.3-point gap without matched reruns is too small to call Terra the better model there, so treat it as a reason to test rather than a reason to switch.

The throughput numbers point the same way. AA measured Terra at 91 output tokens per second at medium 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, which covers that task set and not every workload.

Kingy's pilot is the lightest test in the set: six short tasks run twice at medium effort without tools, Terra used $0.0203 of credit and Sol $0.0380, both passing all ten objective attempts in the format requested, with median response times of 3.84 s and 4.40 s. Those dollars are the total for 12 scored calls, not per task.

Where the cheap tier stops being cheap

Terminal and repository coding. Terra climbs from 24.8% at low effort to 35.8% at medium and 50.3% at max on Terminal-Bench 4.0. Sol at medium already sits above that ceiling, and Sol at max reaches 55.9%. AA's Coding Agent Index puts Sol six points ahead, 54 to 48. One counterweight: AA's own Terminal-Bench run with a different agent harness had the two almost level at max, 55.6% against 56.1%. The size of the gap depends on setup, but no source has Terra ahead on Terminal-Bench 4.0.

Computer use at max. On OSWorld 2.1 offline, Terra at max trails Sol by 3.8 points and costs slightly more per task, $1.19 against $1.15. The lower token price did not make it cheaper; it spent more tokens for a worse score.

Analysis where a wrong number is expensive. In Plotly's DataAnalyticsBench, Sol answered all 43 questions and Terra 40. Terra cost $1.31 against $2.61, so the question is what three wrong answers cost. Plotly ran both through OpenRouter, Terra at its default effort, and the costs are per full run of 43 questions. DualEntry's accounting set, 101 synthetic cases, shows the same 3-point pattern, 78.2% against 81.2%. DualEntry doesn't state the effort setting it used. If a human checks every result, Terra's saving is real; if a wrong figure goes straight into a report, the extra $1.30 per run is cheap insurance.

Agents that can act without approval. OpenAI's system card reports two differences that matter once a model holds write permissions. 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%. 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 are controlled simulations, mostly low-stakes and some run without production safeguards, so they don't predict real-world behaviour — but they are the best evidence available for an agent that can delete files, send messages or spend money.

The router

The policy is a small table of task type to model and effort, plus one rule: when a Terra task fails its acceptance check, re-run on Sol at medium instead of raising Terra's effort.

import os
from openai import OpenAI

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

# task type -> (model, effort). Effort is per-model; Sol has no "none".
POLICY = {
    "classify":    ("gpt-6.1-terra", "none"),     # no reasoning needed
    "extract":     ("gpt-6.1-terra", "low"),
    "screen":      ("gpt-6.1-terra", "medium"),
    "draft":       ("gpt-6.1-terra", "medium"),
    "repo_code":   ("gpt-6.1-sol",   "medium"),
    "analysis":    ("gpt-6.1-sol",   "medium"),
    "agent_write": ("gpt-6.1-sol",   "medium"),
}

def run(task_type: str, prompt: str):
    model, effort = POLICY[task_type]
    kwargs = {"model": model, "input": prompt, "reasoning": {"effort": effort}}
    if effort == "none":
        kwargs["temperature"] = 0.2  # accepted only at effort "none"
    return client.responses.create(**kwargs)
Enter fullscreen mode Exit fullscreen mode

Two details worth keeping in the code rather than in a config comment. none is a Terra-only shortcut for classification and formatting, and it is the only setting where temperature and top_p survive. And the escalation path should change model, not effort: on OpenAI's Terminal-Bench chart, Terra at max scores lower than Sol at medium at nearly twice the cost per task.

Start everything on Terra at low or medium. Escalate to Sol at medium when a task fails, or immediately for repo-scale coding, multi-step data analysis, and any agent holding write permissions. For the last category the escalation is human review rather than Sol at a higher effort setting.

Failure modes and fixes

temperature or top_p rejected. OpenAI requires removing both at every effort level except none. Fix: drop them, or pin the call to none on Terra where they're accepted. Sol has no none, so a Sol call always drops them.

Effort appears to be ignored. In AIHubMix's checks on 27 October, the setting took effect on Responses and made no consistent difference on Chat Completions over a few small prompts (none and max both landed around 200-300 reasoning tokens from run to run). Fix: send the request to the Responses API, and confirm the setting is applied before you tune it.

Bill rises after you "upgrade" a weak Terra task. This is the max-effort trap in the table above: 35,000 output tokens against Sol medium's 8,000. Fix: switch model, not effort, then measure cost per correct result.

Per-task cost doesn't match the estimate. The pipeline arithmetic holds only while both models use the same number of tokens and the cache hits. A 10,000-request job with a 2,500-token stable prefix cached after the first call, 1,500 fresh input and 1,500 output tokens per request comes to 15M fresh input, 25M cached input and 15M output — $91.25 on Terra and $182.50 on Sol at list prices. Cache writes are left out; one 2,500-token write per 30-minute lifetime costs under a cent. If the prefix changes per request, the cached line mostly disappears and the estimate breaks.

Budget assumes a Batch discount. AIHubMix doesn't serve OpenAI's Batch API. Plan on standard rates.

What this comparison can't tell you

The OpenAI figures are self-reported launch charts, transcribed by Kingy and not reproduced by anyone independently. AA's numbers are independent measurements, but from a single harness; its own Terminal-Bench run with a different harness put the two models almost level at max, which is a reminder that agent-harness scores move with the harness. None of the published runs includes matched reruns, so a 1.3-point gap is too small to decide on until you repeat it. The system card numbers come from controlled simulations, some without production safeguards.

What is left after those limits is still useful: Terra at medium is the cheaper default for extraction, forms, screens, documents, drafting and short checkable tasks; Sol at medium is the cheaper route to a result that Terra only reaches at max. Measure both on your own inputs, tools and acceptance checks, and compare cost per correct answer rather than cost per token. The AIHubMix compare page has both price sheets side by side, and the Sol model page lists its parameters for the jobs that still need it.

Keep reading: the GPT-6.1 Terra series

Sources

Top comments (0)