OpenAI launches GPT-6.1 Terra: Astra-grade computer use at a tenth of the price

OpenAI shipped GPT-6.1 Terra on October 26, a mid-range model built for computer use and visual reasoning. The API prices it at $1 per million input tokens and $5 per million output tokens (cached input $0.05) — about a tenth of flagship GPT-6 Astra’s standard rates of $10/$50. OpenAI says Terra matches Astra on the Agents’ Last Exam benchmark and finishes tasks over three times faster on average than Astra and GPT-6.1 Sol. It went live the same day in ChatGPT Work, Codex and the API, and the aggregator OpenRouter listed it within hours, reporting a 1.05M-token context window.

Key points

• A tenth of the flagship price: Terra costs $1/$5 per million tokens (cached input $0.05) against Astra’s $10/$50 — and half of predecessor GPT-5.6 Terra’s $2/$12.
• The vendor’s benchmark claim: matching Astra on Agents’ Last Exam, 15 percentage points above GPT-5.6 Terra at about a third of the per-task cost; 50.3% on Terminal-Bench 4.0 (max reasoning) versus the predecessor’s 21.5%.
• Availability: same-day for Plus/Pro/Enterprise/Edu users in ChatGPT Work and Codex, plus the API; OpenRouter followed the same day with a 1.05M-token context window.

Background

Terra is the third step in OpenAI’s push to sell last month’s frontier capability at a discount. At DevDay, GPT-6.1 Sol arrived claiming near-Astra performance at a fifth of the price; in early October Sol gained an Ultrafast tier that costs six times as much for speed. The flagship itself has been rocky — GPT-6.1 Astra was pulled days before its planned release over poor alignment scores and eventually shipped as the priciest tier in the lineup. Terra fills the middle of that ladder, and The Decoder reads it as a direct answer to Anthropic’s mid-tier Sonnet line: Claude Sonnet 5.5 runs $2/$10, twice Terra’s standard price.

The facts

Terra’s pricing and benchmark numbers below come from OpenAI’s October 26 release materials; the independent cross-checks get their own section.

	GPT-6.1 Terra	GPT-5.6 Terra (previous)	GPT-6 Astra (flagship)
Input / output (per 1M tokens)	$1 / $5	$2 / $12	$10 / $50
Cached input (per 1M tokens)	$0.05	$0.20	—
Agents’ Last Exam	matches Astra	15 points lower	benchmark
Terminal-Bench 4.0 (max reasoning)	50.3%	21.5%	—

• Two discount baselines: “about a tenth of Astra” is measured against the flagship’s rates; The Decoder’s “halves API prices” headline is measured against GPT-5.6 Terra’s $2/$12. Both hold; don’t blend them.
• What it’s for: OpenAI’s own scenarios are browser forms, spreadsheets, technical drawings and video editing — screen-understanding work — plus a “more natural” writing style. On the MMMU-Pro visual-reasoning benchmark it comes close to Astra and Sol.
• The vendor’s own caveat: for the most complex coding tasks, OpenAI says GPT-6.1 Sol and GPT-6 Astra remain the better choice — coding is the only workload the release page names.
• Safety posture: OpenAI says Terra crosses its Preparedness Framework’s “Critical” cybersecurity threshold and inherits Astra’s safeguards.

The independent numbers

The sharpest third-party figures come via The Decoder citing Artificial Analysis, which maintains an intelligence-per-dollar index: Terra scores 48 (GPT-5.6 Terra: 42; Sol: 52) at $0.40 per task versus the predecessor’s $1.40 — sitting on the cost-performance Pareto frontier. Elsewhere on their indexes: 48 on Coding Agent Index v2.0, six points below Sol’s 54 and seventeen above Luna’s 31; the model now guesses when it doesn’t know 38% of the time, down from 88%; indirect prompt-injection defense rose to 98.8% from 96.2%; and on OSWorld 2.1 it beats the predecessor’s best score at medium reasoning effort for a fraction of the cost. The direction agrees with OpenAI’s own numbers: Terra doesn’t top the frontier leaderboard — it resets the price on the “capable enough” line.

What others say

Practitioners spent launch day on price and speed. OpenAI president Greg Brockman wrote on X that computer use with Terra “has crossed a qualitative threshold where it feels faster at using a computer than what i can do unassisted” — to which a popular reply asked how often the model clicks the wrong button. OpenRouter listed Terra the same day at $1/$5 with a stated 1.05M context; the 95% cache discount is not an official line, but it follows directly from the $0.05/$1 rates — and for agents, which re-read large chunks of context at every step, that line item often outweighs the sticker price. The Decoder compressed the whole story into its headline: API prices halved, and the model still gains ground on OpenAI’s own larger models.

Our take

For teams building browser automation and RPA-style agents, Terra moves the cost line. Until now the computer-use shelf offered a flagship price ($10/$50) and a legacy price ($2/$12); $1/$5 with $0.05 cached input makes the read-the-whole-screen-every-step loop affordable at volume for the first time. It also clarifies the product ladder: Terra for bulk computer use, Sol and Astra for the hardest coding — OpenAI’s own recommended split. The caveat worth keeping: these are launch-day numbers, vendor-reported or run on vendor-set benchmark configurations; the miss-and-retry rate on real tasks is yours to measure. As of October 27 we could not find any Chinese-language coverage of the launch, so the figures here are cross-checked against OpenAI’s announcement, Digital Trends and The Decoder. One pricing lever is worth watching: that $0.05 cached-input rate is where the real money is in a 1.05M-token context — OpenAI’s 12-hour prompt cache already showed how cache pricing rewrites agent cost models, and this is the same lever applied to a cheaper model.