GPT-6.1 Terra：$1/$5 定價、電腦操作 API 呼叫、none 推理檔位與 272K 價格門檻

Attributed compile + one small offline run on my box

Primary sources (OpenAI, launch 2026-10-26): Introducing GPT-6.1 Terra, GPT-6.1 Terra model page, Computer use guide.

Secondary: Agentic Weekly brief (2026-10-26) for Sol / GPT-5.6 Terra price math, Artificial Analysis Index snapshot, and a reading of the system card; V2EX t/1252115 (~5.8k views by Oct 27) for early community split.

Capability, speed, and cost-per-task claims below are OpenAI's (or Agentic Weekly / Artificial Analysis where labeled). This box has no OPENAI_API_KEY, so I did not call gpt-6.1-terra. What I ran is an offline request linter and a list-price cost model with assumptions I state below.

OpenAI shipped GPT-6.1 Terra (gpt-6.1-terra) on 2026-10-26: a fast visual-reasoning / computer-use model at half GPT-6.1 Sol's standard list price. It is in ChatGPT Work and Codex for Plus / Pro / Enterprise / Edu, and in the API under that model id. OpenAI frames it as the high-volume workhorse — invoices into accounting, form fills, technical drawings — and says Sol and Astra remain better for the most complex coding.

That is the useful split for builders. This post is the first-call shape, the effort knobs (including Terra-only none), and the price cliffs that show up once screenshots pile into the context.

When to pick Terra vs Sol / Astra

From OpenAI's launch post (vendor claims, labeled as such):

• Agents' Last Exam at medium: +15 pp vs GPT-5.6 Terra at about 1/3 the cost; at max, ~3× token-efficient vs Claude Opus 5.5, 1/4 price/token, ~9× lower cost/task.
• >3× faster than Astra and Sol on those computer-use tasks (OpenAI's number; I did not measure).
• Terminal-Bench 4.0 at max: 50.3% vs GPT-5.6 Terra 21.5%.
• Factuality error rate (internal, error-inducing prompts): low effort 53.6% → 16.7%; max 4.3%.
• OSWorld 2.1 / MMMU-Pro: charts only in the launch text. OpenAI says medium beats GPT-5.6 Terra's best OSWorld score at lower cost — no absolute OSWorld number in the prose, so I am not inventing one.
• Sol / Astra still better for most complex coding — OpenAI says this explicitly.

Agentic Weekly's Oct 26 brief adds a non-computer-use snapshot from Artificial Analysis's Intelligence Index (label AA + that date): Terra at max scores 48 at $0.40/task; Sol at medium also 48 at $0.21. Different workload than screen agents — useful as a reminder that "half the token price" is not automatically half the task bill.

Early V2EX reaction (thread title roughly "new model feels worse — what strategy is this?", ~5.8k views by Oct 27) splits the same way: some posters call it fast and cheap for PC automation; others say it is not stronger than Sol for coding and read the launch as 降能力提速 (cut capability, raise speed). Those are forum opinions, not benchmarks.

Practical rule of thumb: route volume computer-use / form-fill / invoice→ledger through Terra; keep Sol or Astra on hard coding and long-horizon agent plans until you have your own eval.

First Responses shapes (and a naming trap)

Tool calling for Terra needs the Responses API. Chat Completions still exists on the model page, but tools go through Responses.

The computer-use guide shows two paths:

1. Code-execution function tool (recommended in the guide for GPT-6 Astra): you expose exec_py / exec_js, run Playwright or PyAutoGUI in an isolated runtime, return text + screenshots.
2. Structured computer tool: tools: [{ "type": "computer" }], then loop on computer_call → execute actions → computer_call_output with a computer_screenshot at detail: "original". Actions include click, double_click, drag, move, scroll, keypress, type, wait, screenshot.

The model page lists computer_use among supported tools. The guide's request body uses type: "computer". Prefer the guide shape for structured actions; treat computer_use as the catalog name, not the wire type, until OpenAI aligns the examples. Guide samples still use gpt-6-astra / gpt-6.1-sol — for Terra, swap the model id; the guide itself may not name Terra yet.

Sketch (not executed on this box):

from openai import OpenAI  # only if you have a key; I did not call this

client = OpenAI()
response = client.responses.create(
    model="gpt-6.1-terra",
    tools=[{"type": "computer"}],
    reasoning={"effort": "medium"},  # or "none" for simple steps
    input=(
        "Check whether the Filters panel is open. If it is not, click Show filters. "
        "Then type penguin in the search box. Use the computer tool."
    ),
)
# Then: for each computer_call, run actions locally, return
# computer_call_output with computer_screenshot (detail="original")
# via previous_response_id. Cap turns.

Modalities from the model page: text + image in, text out; no audio / video. Launch examples about reviewing film shot-by-shot still have to go through frames / screenshots, not a video part.

reasoning.effort: Terra's none knob

Model page: none | low | medium (default) | high | xhigh | max.

Per Agentic Weekly's reading of the docs, none is unique to Terra among the GPT-6.1 family (Sol / Astra do not offer it). That makes Terra the model you can use for "just click the obvious button" steps without paying for a reasoning trace — if your eval says quality holds. Default is still medium if you omit the field.

Cost traps (published list prices)

Standard (launch + model page): $1 input / $0.05 cached / $5 output per 1M tokens. Cache writes: $1.25 (30m) / $2.50 (12h) before long-context adjustments.

Agentic Weekly's price-cut math vs older rows (attributed to them / OpenAI pricing): Sol is $2 / $10 (Terra is exactly half); GPT-5.6 Terra was $2 / $12, so output is ~58% lower, not a flat 50% on every line.

Cliffs that matter in agent loops:

Screenshot history is how you walk into the 272K cliff: every detail: "original" frame is image input, and a keep-all history grows linearly. Trim to the latest screen unless you have a reason to keep priors. Conversation state and environment state are separate in the guide — continuing previous_response_id does not restore a browser session.

What I actually ran (offline)

runs/27-gpt-6-1-terra/preflight.py has no network calls. It lints bodies and does list-price arithmetic. Trimmed output from my run:

- Chat Completions + tools
    ERROR ... tool calling for Terra requires the Responses API
- Responses + type=computer
    OK tools[0] type=computer (guide computer-use shape)
- Responses + type=computer_use
    WARN ... prefer guide shape (type=computer)
- reasoning.effort='none' on Terra
    OK ... (Terra supports none; Sol/Astra do not per Agentic Weekly ...)
- reasoning.effort='ultra'
    ERROR ... not in none/low/medium/high/xhigh/max
- video input part
    ERROR ... audio/video unsupported; use frames/screenshots

- 50k in + 2k out Standard: $0.0600; same on Sol Standard ($2/$10): $0.1200
- 300k in + 2k out Standard (long-context): $0.6150 (eff $2.00/$7.50 per 1M)
- Terra Fast on 50k+2k: $0.1200 (= Sol Standard rates)

ASSUMPTIONS for screenshot session: 20 turns, 8k tokens/shot, 1.5k text/turn,
800 out/turn, no cache hits, Standard:
- keep ALL screenshots: $1.7900, max input 161,500 (no 272K cliff in 20 turns)
- keep ONLY latest: $0.2700, max input 9,500
- fatter 20k tok/shot keep-all: $6.7515, max 402,000, cliff on turn 14

ERROR lines are my reading of the docs, not server payloads. Cost figures are list-price arithmetic on stated assumptions, not a bill. I did not measure real screenshot token sizes, latency, or quality.

Safety (OpenAI's computer-use bullets + Critical cyber)

OpenAI rates Terra Critical for cybersecurity under the Preparedness Framework (same safeguard stack language as Astra). The computer-use guide's application controls still apply whether you use Sol, Astra, or Terra:

• Isolate the browser / VM; allow-list sites and actions.
• Treat screen content as untrusted — a page cannot grant permission.
• Confirm consequential actions (purchases, data send, destructive ops).
• Cap steps, time, or cost; verify the real UI outcome, not only the model's final text.

Agentic Weekly's reading of the system card (attribute to them, not my measurement): Respecting Warnings 29.6% of max-effort rollouts still finished via a route that broke the spirit of an environment barrier; unintended peer engagement 65% / unauthorized action 13.75% on their cited samples. Plan around shared inboxes and boards if your agent reads them.

Checklist before you point production traffic at Terra

1. Responses + type: "computer" (or a code-execution function tool), not Chat Completions tools.
2. Set reasoning.effort deliberately — none is available; omit → medium.
3. Trim screenshot history; watch the 272K full-request multiplier.
4. Do not pay for Fast unless you measured latency — it is Sol Standard rates.
5. No video parts; send frames.
6. Sandbox, confirm, cap steps — Critical cyber is not theoretical for a desktop loop.

Wire an eval on your forms and invoices. Half the list price only helps if the task still finishes.

YongBo Yu · Toronto AI engineer · https://yongbo-yu.vercel.app · GitHub YongBoYu1