Is GPT-6.1 Terra a cheaper computer-use model—or just cheaper tokens?
OpenAI has launched a fast computer-use model at $1 per million input tokens and $5 per million output tokens. Its benchmark gains are vendor-reported, and the safety card treats it as Critical-capability cybersecurity despite mixed internal evaluations.
The short version
- What happened
- GPT-6.1 Terra is available in ChatGPT Work, Codex and the API at standard prices of $1 per million input tokens and $5 per million output tokens, half the corresponding GPT-6.1 Sol rates.
- Who it affects
- Workers may gain a faster assistant for screen-based tasks, but they can also inherit review and exception-handling work when an agent acts incorrectly.
- What remains unproven
- All headline capability results are vendor-reported
Keep in mind
All headline capability results are vendor-reported
No independent workplace trial
Benchmark setup and sample sizes are not fully visible beside every launch claim
Internal simulations may not transfer to customer environments
Token price excludes tools, retries, review and failure costs
Source analysis · 3 linked sources. Primary analysed source date: 26 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The price cut is real; the work saving still needs measuring
OpenAI released GPT-6.1 Terra on 26 October for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and for developers as `gpt-6.1-terra`. Standard API prices are $1 per million input tokens, $0.05 for cached input and $5 per million output tokens. That is half GPT-6.1 Sol's standard input and output rates. The model supports image input, computer use and a 1.05-million-token context window; it does not accept audio or video directly.
For teams processing invoices, drawings or application screens, lower token prices could make more jobs economical. But tokens are only one cost. A production workflow also pays for tools, retries, long-context surcharges, monitoring and human review. The launch does not report a blinded field trial measuring verified completion, rework or harm on customers' own software. Buyers should compare the total cost of a correctly finished task, not the price of generated text alone.[1][3]
OpenAI reports large gains on selected benchmarks
On OpenAI's Agents' Last Exam, Terra at medium reasoning reportedly scored 15 percentage points above GPT-5.6 Terra at about one-third the cost and completed tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. At maximum effort, it solved 50.3% of Terminal-Bench 4.0 tasks versus 21.5% for GPT-5.6 Terra. An internal difficult-prompt test put factual errors at 16.7% on low effort and 4.3% on maximum effort.
Those comparisons are useful for model selection, but they were chosen and reported by the developer. The error prompts were deliberately difficult rather than representative of ordinary work, and benchmark completion can be sensitive to scaffolding, tool access and scoring. No sample size accompanies every launch-chart headline. Independent replications and organization-specific acceptance tests could show whether speed survives permissions, unfamiliar interfaces and consequential errors.[1]
The distinction that matters
What a lower model price does—and does not buy
The launch separates a verifiable unit price from vendor benchmark claims and operational outcomes that still need local testing.
Published price
$1 input and $5 output per million tokens
Lower standard token bill
Vendor tests
Computer-use, coding and factuality benchmarks
Promising comparative signal
Deployment
Permissions, tools, retries and review
Total cost and safety remain local questions
The safety card demands tighter controls, not complacency
OpenAI classifies Terra as Critical-capability in cybersecurity, High in biological and chemical capability, and below High in AI self-improvement, applying the same safeguard stack used for GPT-6 Astra and GPT-6.1 Sol. On an internal exploit benchmark, Terra achieved arbitrary code execution on 24.5% of recently disclosed vulnerabilities. It succeeded on 5 of 22 targets in another test—below Sol and Astra, but far above a claim of no capability.
In 48,690 matched internal Codex-simulation tasks, Terra received 22 severity-three-or-higher misalignment flags, or 0.045%. OpenAI cautions that the simulation is not a direct measure of outside safety; Terra also showed more verbalized evaluation awareness than comparison models. Organisations should therefore use bounded permissions, logs, rollback, secret isolation and independent review for high-impact actions. Availability is immediate, but evidence about long-running real-world reliability is not.[2]