AI & DEV NEWS · · 4 min read

GPT-6.1 Terra: half the price, and the claims to test first

OpenAI listed GPT-6.1 Terra in its API changelog this week: $1 in, $5 out per million tokens. The price is checkable. Most of the speed and computer-use claims are still OpenAI's own.

What GPT-6.1 Terra is, and what it costs

The API id is gpt-6.1-terra. theagenticweekly.com dates the release to Oct 26 from the OpenAI changelog. Some roundups say Oct 25, so treat the day as approximate. It has a 1.05M-token context window, 128K max output, and takes text and images in and text out.

Standard pricing is $1 input and $5 output per million tokens. Per the same page, fast mode costs 2x, while Batch and Flex are 50% lower. ai-daily.dev puts the standard price at roughly 50% below both GPT-5.6 Terra and GPT-6.1 Sol. It is in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu, and in the API.

Some limits matter more than the price. Tool calling requires the Responses API. Fine-tuning isn't supported. Audio and video input aren't either. If your pipeline depends on any of those, the discount doesn't help you yet.

The benchmark numbers are OpenAI's, not ours or anyone else's

OpenAI says Terra scores 15 percentage points above GPT-5.6 Terra on Agents' Last Exam at medium effort, at about one-third of the cost. theagenticweekly.com notes it hasn't measured this and that the launch charts don't expose their values in the page text. The one number it could copy is Terminal-Bench 4.0 at 50.3% with max effort, again OpenAI's figure.

ai-daily.dev adds that OpenAI claims computer-use parity with GPT-6 Astra at about a tenth of the per-token cost, and more than 3x the throughput of Astra and Sol. It also reports the low-effort error rate falling from 53.6% to 16.7%. We found no independent latency measurement for any of this. Launch-week speed claims tend to describe a quiet system, so we'd wait for third-party numbers.

One more detail for anyone in security review: the same Agentic Weekly page lists Terra as meeting the Critical threshold in cybersecurity under OpenAI's Preparedness Framework. That's worth reading before you hand it a shell.

Cheaper tokens don't fix brittle computer-use agents

Computer use is the headline pitch, and it's where we'd be most careful. A MindPattern briefing summarises a study of 17 agent systems. When something relevant changes partway through a run, the success rate falls to 24.79%. Only 20.34% of scenarios succeed on both the clean run and the shifted one. That study isn't about Terra, and we haven't verified it ourselves. It does describe the failure mode you'll hit: the UI moves, a session expires, the agent keeps going.

A cheaper model makes retries cheaper. It doesn't make the agent notice that the world changed. Our OpenAI Agents SDK teardown makes a related point: durability, checkpoints and recovery are your job, whatever model sits underneath.

What we'd do this week

Nothing needs to ship on Monday. But a model at half the price is worth a spike, especially for high-volume vision or form work where you're paying premium rates today.

dev.to reports that Vercel's AI Gateway already routes to openai/gpt-6.1-terra with no markup over direct API pricing, which makes a side-by-side test a one-line model swap if you're on it. For the testing method, explainx.ai suggests 100 to 200 real tasks scored on cost per completed task, not per million tokens. We agree.

  • Pick one workflow that uses a premium model today and replay 100 to 200 real tasks through Terra.
  • Score cost per completed task and failure recovery, not just accuracy on the happy path.
  • Test the same task after you change the UI or expire a session, since that's where agents fall over.
  • Keep your current model as a tested fallback, so routing is a config change and not a rewrite.

Where this fits in your routing

If you run n8n, Make or Node.js workflows with model calls inside, the change is small: a new option on the routing table. That's the kind of work we do in AI workflow automation, and the principle holds either way. Route by task, measure, and keep the switch cheap.

Sources

  1. Changelog | OpenAI API · OpenAI
  2. GPT-6.1 Terra: release date, specs, price · Agentic Weekly
  3. AI Daily Dev, October 27, 2026 · AI Daily Dev
  4. GPT-6.1 Terra, Claude Opus 5.5 pricing, and the infrastructure week AI tooling needed · DEV Community
  5. Ramsay Research Agent, October 26, 2026 · MindPattern
  6. OpenAI Discounts Up to 90% vs Anthropic: What Is Verified · explainx.ai

Want this working in your stack, not just in a demo?

We are two senior engineers who build ai workflow automation with a fixed price per written scope.

Keep reading.