GPT-6.1 Terra: API, Benchmarks, Pricing & How to Use It

What is GPT-6.1 Terra?

OpenAI now ships named tiers within each model generation — roughly Astra (flagship), Sol, Terra and Luna — where the number (6.1) marks the generation and the name marks the capability-and-cost tier. Terra is the balanced, production tier: not the absolute frontier on raw reasoning, but tuned to be fast and cheap enough to run agentic and computer-use workloads at volume.

The defining capability this release is computer use. Terra pairs visual understanding with the ability to operate browser and desktop applications, surfaced as a hosted tool in the Responses API. In plain terms: it can look at a screen and click, type and navigate through it — the core loop behind practical computer-using agents — at a price that makes those loops economical.

Key specifications

Pricing

Terra’s pricing is the story. At $1/$5 per million tokens it undercuts frontier models several times over, and the batch/flex discount halves that again for asynchronous work.

Two cost traps worth flagging: fast mode doubles the rate, and any request whose input exceeds 272K tokens is billed at 2x input/cache and 1.5x output for the entire request. Long-context agent loops can quietly cross that line, so measure actual token shapes rather than assuming the headline $1/$5.

Benchmarks

OpenAI’s launch-day figures emphasize security, multimodal troubleshooting and health reasoning:

• ExploitBench: 97.10%
• Multimodal Troubleshooting (Virology): 52.54%
• HealthBench Professional: 62.7 (up from 57.7 for GPT-5.6 Terra)
• Standard HealthBench: 55.1 (down slightly from 57.0 for GPT-5.6 Terra)
• Sandbox Bench: 5 of 22 targets

Read these as launch-day vendor numbers, not a full production verdict. Notably, Terra is not uniformly ahead of its predecessor on every benchmark (standard HealthBench dipped), which is a good reminder that tier models trade raw capability for speed and cost. OpenAI also claims Terra is more than 3x faster than GPT-6 Astra and GPT-6.1 Sol — a meaningful latency edge for interactive agents, but a vendor claim from internal testing.

How GPT-6.1 Terra compares

The useful comparison is on price-for-capability against the models you would otherwise route to. All figures are per million tokens and reflect published launch pricing (verify before relying on them).

The takeaway: Terra sits in the same low-cost band as Gemini 3.6 Flash and well below Claude Opus 4.8 and Fable 5.1, while bringing first-class computer use. It is not trying to beat Fable 5.1 or GPT-6.1 Sol on the hardest reasoning — it is trying to make the routine 80% of agent work cheap. That is precisely the split a gateway is built to exploit: Terra for volume, a frontier model for the hard tail.

Limitations

• Tier, not frontier. Terra trades peak reasoning for speed and cost. For the hardest long-horizon tasks, Fable 5.1 or a flagship tier will still win.
• Long-context cost cliff. Requests over 272K tokens are billed at 2x input/1.5x output for the whole request — easy to trigger accidentally in long agent loops.
• No fine-tuning. Terra does not support fine-tuning at launch.
• Access gaps. Not available in regular ChatGPT Chat, Free or Go at launch, and not available for UAE inference residency.
• Text-only output. Inputs are text and image; output is text (no audio or video).

How to access and use GPT-6.1 Terra

There are two practical paths to production.

1. Call the OpenAI API directly

Terra is available on the OpenAI API at the Build tier and above, with computer use exposed as a hosted tool in the Responses API. This is the fastest way to prototype a single app against a single model.

2. Route it through the TrueFoundry AI Gateway (recommended for production)

The moment you run Terra alongside other models — Claude for hard reasoning, Gemini for cheap multimodal, Terra for computer use — you need routing, fallbacks, budgets and one cost ledger. That is what the TrueFoundry AI Gateway provides: a single OpenAI-compatible endpoint in front of every provider.

1. Point your client at the TrueFoundry gateway endpoint and select gpt-6.1-terra as the model.
2. Add routing rules — send computer-use and high-volume calls to Terra, escalate the hard tail to a frontier model, with automatic fallback if a provider errors or rate-limits.

You get per-key, per-team and per-agent budgets, a cross-provider cost and token ledger, and full audit — without rewriting application code when prices or models change. The gateway adds only a few milliseconds of overhead and is OpenTelemetry-compliant.

Put GPT-6.1 Terra behind one governed, multi-provider endpoint with budgets and fallbacks. Talk to the TrueFoundry team →

When to use GPT-6.1 Terra

• Reach for Terra when: the workload is computer use or agentic automation; volume is high and cost matters; you need a large context at a low per-token rate; latency matters for an interactive agent.
• Escalate to a frontier model when: the task is genuinely hard long-horizon reasoning or complex coding where a wrong answer costs more than the tokens — route those to Claude Fable 5.1, Opus 4.8 or GPT-6.1 Sol.
• Consider Gemini 3.6 Flash when: you want the cheapest fast multimodal option and don’t need Terra’s computer-use tool.