MoeRouter Blog
English
Get an API key

GPT-6.1 Sol and Terra review roundup: the most-used model, and whether GPT-6.2 replaces it

GPT-6.1 Sol is the most used model on MoeRouter and GPT-6.1 Terra is its half-price fast tier. Artificial Analysis's effort ladder, DataCamp's hands-on test and Bilibili creators, and whether to stay on it now that GPT-6.2 is out.

gpt-6.1-solgpt-6.1-terra

In short

  • GPT-6.1 Sol launched on September 29, 2026 at a list price of $2 in / $10 out per 1M tokens. As of November 18 it was the most used model on MoeRouter over the previous 7 days.
  • An independent effort ladder shows Sol cheaper than Claude Opus 5.5 by about 70% at every level up to an index score of 51; above that only Opus 5.5 goes.
  • GPT-6.1 Terra launched on October 26 at half Sol's price. OpenAI says it finishes tasks over three times faster; third-party tests score it 4 to 6 points lower. It suits high-volume short tasks such as forms and computer use.
  • GPT-6.2 Sol costs the same and scores 2 points higher on general intelligence, but 3 points lower on the independent coding-agent index. If you mostly write code, 6.1 Sol still holds.
On this page
  1. What the two models are
  2. What they cost on MoeRouter
  3. Benchmarks: cheaper at every level up to 51
  4. Hands-on
  5. Terra: half the price, three times the speed, a step less clever
  6. GPT-6.2 is out: should you still use 6.1 Sol?
  7. How to call them
  8. Frequently asked questions

What the two models are#

GPT-6.1 Sol was released at OpenAI's DevDay on September 29, 2026, seven days after GPT-6 Sol[1][3]. It is the everyday tier of the GPT-6 family: far cheaper than Astra at the top, clearly more capable than the light Luna.

GPT-6.1 Terra followed on October 26. It replaces GPT-5.6 Terra and is built for speed and cost, aimed at visual reasoning, computer use and writing[4].

GPT-6.1 Sol GPT-6.1 Terra
List price per 1M tokens (input / cached input / output) $2 / $0.10 / $10 $1 / $0.05 / $5
Model name on MoeRouter gpt-6.1-sol gpt-6.1-terra
Endpoints on MoeRouter Chat Completions, Responses, Messages Chat Completions, Responses, Messages

Sol has a 1,050,000-token context and five effort levels: low, medium, high, xhigh and max. Above 272K input tokens the whole request is billed at twice the input and cache rates and 1.5 times the output rate[1][3]. Terra's prices are as DataCamp reports them[4].

What they cost on MoeRouter#

per 1M tokens
ModelGroupInputOutputCache read
gpt-6.1-soldefaultCN¥0.324≈ US$0.0486CN¥1.62≈ US$0.243CN¥0.0162≈ US$0.00243
gpt-6.1-terradefaultCN¥0.18≈ US$0.027CN¥0.9≈ US$0.135CN¥0.009≈ US$0.00135

Reference prices per 1M tokens in Chinese yuan (CNY), the currency accounts are charged in, from the model square (). Dynamically priced models settle at the route that served each request; the model square is authoritative. Amounts in USD are estimates, converted at 1 CNY = 0.15001 USD ().

Both models are in the default group, which sends each request down the cheapest healthy route; that is why the reference price is below list. The usage log shows what each request was charged. See dynamic routing.

Benchmarks: cheaper at every level up to 51#

Independent: scores by effort level#

CodingFleet compiled Artificial Analysis's runs of the same ten evaluations on both models (Intelligence Index v4.3.2), with cost counted per task including cache reads and writes[2]:

Configuration Index Cost per task Time per task
GPT-6.1 Sol · low 42 $0.13 80 s
GPT-6.1 Sol · medium 48 $0.21 154 s
GPT-6.1 Sol · high 50 $0.32 250 s
GPT-6.1 Sol · xhigh 51 $0.39 307 s
GPT-6.1 Sol · max 52 $0.72 647 s
Claude Opus 5.5 · low 42 $0.55 79 s
Claude Opus 5.5 · medium (default) 51 $1.34 205 s
Claude Opus 5.5 · max 58 $5.98 783 s

It reads simply. Sol at xhigh scores exactly what Opus 5.5 scores at its default, for 71% less per task. Sol's ceiling is 52; the 54, 56 and 58 that Opus 5.5 reaches above its default have no Sol counterpart at any price. Opus 5.5 wins 13 of the 16 component evaluations, and Sol's one clear win is question answering over dense professional PDFs (GDP.pdf, 31.0% against 26.2%). The author's verdict is that "GPT-6.1 Sol is the better default and Claude Opus 5.5 is the better ceiling"[2].

The same article warns of a trap: past 272K input tokens Sol's long-context rates apply to the entire request, so adding 1,000 tokens takes a request from $0.744 to $1.392, while Opus 5.5 serves its full 1M context at flat rates[2].

Reported by OpenAI#

Benchmark GPT-6.1 Sol Claude Opus 5.5
DeepSWE v1.1 (high) 75.2% not published
AutomationBench (medium) 31.7% at $0.19 per task 29.5% at $0.65
AutomationBench (max) 36.1% at $0.30 42.5% at $1.44
Terminal-Bench Science 0.1 (max) 57.0% at $5.47 63.3% at $23.21

On DeepSWE, GPT-6 Sol's best was 68.8% at max; 6.1 Sol reaches 75.2% at high[3]. DataCamp's author notes that OpenAI's headline used the medium-effort point while the same chart puts Opus 5.5 ahead at max[1]. Kama Notes adds that "beats Opus 5.5" holds only for that benchmark, effort level and cost, and cannot be stretched into a general claim[3].

Hands-on#

DataCamp had both models build a clinic appointment scheduler in one HTML file: a week view, add, edit and delete, conflict detection, and state that survives a reload. Both passed every conflict, edit, delete and reload check and scored 5 for spec adherence and conflict logic; for usability Sol got 5 and Opus 5.5 got 4. It cost $0.27 on Sol and $1.11 on Opus 5.5, which used more than twice the output tokens, mostly on thinking. The author's line: "If cost per task is your constraint, use GPT-6.1 Sol."[1]

薛定谔船长 (Bilibili) had GPT-6 Luna, Sol, Astra and 6.1 Sol produce the same pelican-on-a-bicycle HTML and SVG animation in Codex, with the same prompt and effort, first generation and no fixes, and compared the result, the time and the Codex quota each used[6].

AI产品狙击手 (Bilibili) spent under $3 on constrained sentence generation, a 3D racing game and a Trello-style board. The description says every project had both surprises and flaws, the flaws being rendering errors such as a hand passing through a neck in an archery test and a pelican's head spinning a full turn[7].

Terra: half the price, three times the speed, a step less clever#

  • Reported by OpenAI. On Agents' Last Exam, Terra finishes tasks more than three times faster than Sol or Astra. On Terminal-Bench 4.0 at max effort it scores 50.3%, against 21.5% for GPT-5.6 Terra. OpenAI itself still recommends Sol or Astra for the hardest coding work[4].
  • Independent. On Artificial Analysis's figures, at matched effort a task costs 1.8 to 2.6 times as much on Sol as on Terra, and Sol scores 3.7 to 6.5 points higher[2].
  • 程序员阿江-Relakkes (Bilibili) ranked Terra eighth overall on his private ModelDial question bank, and tested it driving a desktop application through Codex for a round of automated testing[8].

DataCamp's reading: on OpenAI's numbers Terra looks like the default for high-volume document, form and computer-use work, with Sol or Astra for tasks, such as complex software engineering, that justify several times the cost[4].

One thing to know on MoeRouter: gpt-6.1-terra cannot continue a long Codex session that carries encrypted reasoning. If you get stateful_request_not_supported, switch to gpt-6.1-sol and start a new session; see Codex setup.

GPT-6.2 is out: should you still use 6.1 Sol?#

GPT-6.2 Sol launched on November 17 at exactly 6.1 Sol's list price. It is 2 points higher on the Artificial Analysis Intelligence Index, and 3 points lower than 6.1 Sol on the Coding Agent Index, run in Codex at maximum effort[5].

So it depends on what you do with it:

If you Then
Mostly write code in Codex and it works for you Stay on 6.1 Sol; compare on your own tasks before switching
Do document, data and general agent work Try 6.2 Sol
Have high volumes of short tasks that must be fast 6.1 Terra
Have problems 6.1 Sol keeps failing Move up to Astra or Claude Opus 5.5

More in the GPT-6.2 Sol and Astra review.

6.1 Sol is good for

  • Daily Codex development, bug fixes, writing tests
  • High-frequency agents that reuse a long context (cache reads are cheap)
  • Budget-capped work that a test suite can check

6.1 Sol is less suited to

  • Problems above an index score of 51
  • Requests with more than 272K input tokens
  • Audio input or fine-tuning, which it does not support

How to call them#

In Codex:

model_provider = "moerouter"
model = "gpt-6.1-sol"

[model_providers.moerouter]
name = "MoeRouter"
base_url = "https://moerouter.com/v1"
wire_api = "responses"
env_key = "MOEROUTER_API_KEY"
requires_openai_auth = false

With the OpenAI SDK:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://moerouter.com/v1",
    api_key=os.environ["MOEROUTER_API_KEY"],
)
reply = client.chat.completions.create(
    model="gpt-6.1-terra",
    messages=[{"role": "user", "content": "Extract the parties, the amount and the effective date from this contract as JSON"}],
)
print(reply.choices[0].message.content)

If you have no key yet, the quick start creates one. The docs are in Chinese for now.

Frequently asked questions#

GPT-6.1 Sol or GPT-6.2 Sol?#

The list price is the same. 6.2 Sol is 2 points higher on the general intelligence index and 3 points lower on Artificial Analysis's coding-agent index. Stay on 6.1 Sol if you mostly write code; try 6.2 Sol for other work.

How much does the GPT-6.1 Sol API cost?#

List price per 1M tokens is $2 input, $0.10 cached input and $10 output. Above 272K input tokens the whole request is billed at twice the input and cache rates and 1.5 times the output rate. MoeRouter's reference prices are in the table above.

What is the difference between GPT-6.1 Sol and Terra?#

Terra costs half as much per token and finishes tasks faster, and scores 4 to 6 points lower. Use Terra for high-volume short tasks such as forms, extraction and computer use, and Sol for multi-file refactors and work that needs reasoning.

Can GPT-6.1 Sol replace Claude Opus 5.5?#

Up to an index score of 51 it can, at roughly 70% less per task. Harder problems are reachable only by Opus 5.5, at $1.82 to $5.98 per task. The usual set-up is Sol by default and Opus 5.5 when a task fails.

Why is GPT-6.1 Sol the most used model on MoeRouter?#

Over the 7 days to November 18, 2026, it used the most tokens. Three likely reasons: it is the model in the sample configuration of MoeRouter's Codex guide, it is cheap, and it can continue long Codex sessions.

Codex says stateful_request_not_supported with Terra. What now?#

gpt-6.1-terra cannot handle a request that carries encrypted reasoning. Switch the model to gpt-6.1-sol or gpt-6.2-sol and start a new session.

Sources

The test results and figures quoted in this post come from the public sources below and belong to their authors; each quotation is marked where it appears.

  1. GPT-6.1 Sol vs. Claude Opus 5.5: Which Model to Use DataCamp · published 2026-10-01
  2. Claude Opus 5.5 vs GPT-6.1 Sol: Level at Default Effort, No Contest Above It CodingFleet · published 2026-10-06
  3. GPT-6.1 Sol正式发布:价格、跑分、Codex与GPT-6 Astra、Claude Opus 5.5怎么选 (in Chinese) 卡码笔记 (Kama Notes) · published 2026-11-18
  4. GPT-6.1 Terra: Benchmarks, Pricing and Access DataCamp · published 2026-10-27
  5. OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains The Decoder · published 2026-11-17
  6. GPT-6.1 Sol实测!Codex 四模型同做鹈鹕动画,谁最值得用? (video, in Chinese) Bilibili · 薛定谔船长 · published 2026-10-01
  7. GPT 6.1 Sol 测评报告!究竟有多能打? (video, in Chinese) Bilibili · AI产品狙击手 · published 2026-09-30
  8. GPT6.1 Terra 实测:又快又便宜,电脑操作能力更强了! (video, in Chinese) Bilibili · 程序员阿江-Relakkes · published 2026-10-27

Disclosure: MoeRouter sells API access to the models discussed here. The test results in this post come from the third-party sources cited; prices and endpoint details come from MoeRouter's own data.