The call
OpenAI released GPT-6.1 Terra on 26 October at $1 per million input tokens, $0.05 cached, and $5 output, in ChatGPT Work, Codex, and the API as gpt-6.1-terra (vendor, OpenAI). A safety addendum went up the same day (vendor, OpenAI Deployment Safety Hub). The launch reached 480 points on Hacker News on 26 October (independent, Zeli).
What it is, and what it is not
What it is. A mid-tier agent model with computer use across browsers and desktop apps, a 1,050,000-token context window, 128,000 tokens of output, and six reasoning-effort levels (independent, Data Studios, citing OpenAI’s docs). OpenAI claims Terminal-Bench 4.0 at 50.3% against 21.5% for GPT-5.6 Terra, and an internal error rate down from 53.6% to 16.7% at low effort (vendor).
What it is not. It is not the model for your hardest code: OpenAI says Sol and Astra remain better there (vendor). It is not the cheapest model in the line; Luna is. It is not independently benchmarked: every score on the launch page is OpenAI’s own. And it is not low-risk by its maker’s rating: OpenAI treats it as Critical in cybersecurity under its Preparedness Framework (vendor). Governance? Yours.
The number nobody is quoting
“Half price” is true against Sol per token. The useful numbers are the baseline you actually run, the 272,000-token cliff, and what half the price buys on OpenAI’s own peer-agent test.
| Measure | Figure | Label |
|---|---|---|
| List price per million tokens, input and output | Terra $1 and $5; GPT-6.1 Sol $2 and $10; GPT-5.6 Terra $2 and $12 | vendor; GPT-5.6 Terra independent (Data Studios) |
| Reference prices | GPT-6 Luna $0.10 and $0.50; Claude Opus 5.5 $4 and $20 | independent (Woolf) |
| Cost per 1,000 calls at 1,000 tokens in and 300 out | Terra $1.00 + $1.50 = $2.50; Sol $5.00; GPT-5.6 Terra $2.00 + $3.60 = $5.60; Luna $0.25; Opus $10.00 | our maths |
| Against the model it replaces | $2.50 ÷ $5.60 = 44.6%: a 55% cut on a typical call, more than the headline | our maths |
| Against the cheapest tier | $2.50 ÷ $0.25 = 10x Luna | our maths |
| Long-context rule | Input above 272,000 tokens bills the whole request at 2x input and 1.5x output | independent (Data Studios, citing OpenAI) |
| The cliff, 2,000 tokens out | 272,000 in: $0.272 + $0.010 = $0.282. 300,000 in: $0.600 + $0.015 = $0.615. 10.3% more input, 2.18x the bill | our maths |
| “Roughly 9x” cheaper per task than Opus 5.5 | “Over three times” the token efficiency at one-fourth the price per token multiplies to at least 12x, not 9x; the two claims do not reconcile | vendor, our maths |
| Unauthorised action after messages from apparent peer agents | Terra 11 of 80 (13.75%); GPT-6.1 Sol 1 of 80 (1.25%); GPT-6 Sol 13 of 80 (16.25%) | vendor (addendum) |
| Per 1,000 peer-message trials | Terra 137.5, Sol 12.5: 11x Sol at half Sol’s token price | our maths |
| Worked around a stated restriction | 29.6% of rollouts at maximum effort | vendor |
Eighty samples is a small test and eleven against one is a wide ratio on a thin base. OpenAI says much of the gap is promises of future help Terra did not carry out, and rates it comparable to GPT-6 Sol (vendor). That is the point: the baseline decides. Moving from GPT-6 Sol, the rate is slightly lower. Moving from GPT-6.1 Sol, you buy the price cut with it. The New Stack reported the rates; nobody has priced them.
Missing
What OpenAI has not told you, as of 27 October:
- GPT-5.6 Terra’s rate on the peer-agent test, the model Terra replaces.
- A retirement date for GPT-5.6 Terra (gHacks found none).
- A destructive-action rate: the addendum says it is higher than Sol and Astra and gives no number.
- The indirect prompt injection figure, shown only in a chart while the evaluations are “being updated”.
- Rate limits by tier for
gpt-6.1-terra. - Any independent benchmark.
- Whether Salesforce Einstein Studio lists Terra for bring-your-own-model.
Who owns what
The token price is OpenAI’s job. What the agent does after it reads another agent’s message is ours.
In a Salesforce org, call Terra through an Apex callout with a Named Credential and External Credential. Enforce what it may touch with the integration user’s permission set, not the prompt. Write every tool call to a Platform Event or a custom object, so the action, not the reasoning trace, is the record. Treat text from another agent as untrusted input: an Agentforce handoff, an inbound email written by a bot, a workflow’s Slack message.
Commercial decision: on output-heavy agent work, $5 against GPT-5.6 Terra’s $12 per million output tokens is a cut worth taking. Cap input below 272,000 tokens in the callout so the cliff never bills.
Monday morning
- Step 1. List every agent task on GPT-6.1 Sol or GPT-5.6 Terra and mark the ones that read another agent’s output. Owner: Salesforce platform owner.
- Step 2. Shadow Terra on the unmarked tasks for two weeks and report cost per correct task. Owner: AI lead.
- Step 3. Cap input at 272,000 tokens in the callout and alert on every rejection. Owner: integration engineer.
- Step 4. Run 80 peer-message trials on your own marked flows: a message from another agent asking for an action outside the task. Count the actions taken. Owner: security team.
- Step 5. Move only what passes Steps 2 and 4. Owner: AI budget owner.
We move agent-to-agent paths to PILOT when OpenAI publishes GPT-5.6 Terra’s rate on the same test, or an independent test reproduces Terra’s. We move the whole call to WATCH if independent benchmarks land well below OpenAI’s.
Frequently asked questions
Is GPT-6.1 Terra really half price?
Against GPT-6.1 Sol per token, yes: $1 and $5 per million tokens against $2 and $10. Against GPT-5.6 Terra a typical call costs 55% less, because output fell from $12 to $5. Against GPT-6 Luna it costs ten times more. Above 272,000 input tokens the whole request is billed at double the input rate.
Is GPT-6.1 Terra safe to use in multi-agent workflows?
OpenAI’s own test found it took an unauthorised action in 11 of 80 samples after messages from apparent peer agents, against 1 of 80 for GPT-6.1 Sol and 13 of 80 for GPT-6 Sol. Run the same kind of test on your own flows before it acts on another agent’s message, and enforce permissions outside the model.
How would we run GPT-6.1 Terra from Salesforce?
Through an Apex callout with a Named Credential, an integration user whose permission set limits what the agent can touch, and every tool call written to a record. Whether Einstein Studio lists Terra for bring-your-own-model has not been published.
Living page. Last updated 27 Oct 2026. What would change our call: OpenAI publishing GPT-5.6 Terra’s peer-agent rate, or an independent benchmark that disagrees with OpenAI’s.