GPT-6.1 Terra Pricing and Benchmarks for Developers

OpenAI released GPT-6.1 Terra on October 26, 2026, and the number that matters for builders is the price: $1 per million input tokens and $5 per million output tokens, exactly half of GPT-6.1 Sol on every line of the rate card. It landed the same day in the OpenAI API, ChatGPT Work and Codex, GitHub Copilot and Amazon Bedrock. If you run high-volume agents, Terra is now the default model to test before you reach for Sol or Astra.

What shipped on October 26

Terra is the mid-tier class of OpenAI's GPT-6 family, sitting between the cheap Luna models and the more capable Sol and Astra. GPT-6.1 Terra replaces GPT-5.6 Terra, which had been the Terra-class option since the GPT-6 line launched without one. OpenAI positions it for visual reasoning, computer use and clear writing, and GitHub pitches it for "everyday agentic coding and terminal work."

The availability footprint on day one:

• OpenAI API as gpt-6.1-terra, on Chat Completions, Responses and Batch. See the model page.
• ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users.
• GitHub Copilot, generally available for Pro, Pro+, Max, Business and Enterprise, with a gradual rollout across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI and the cloud agent.
• Amazon Bedrock, generally available, with the option to point ChatGPT Work and Codex at the Bedrock-hosted model.
• Microsoft Foundry, which Microsoft also announced as generally available.

The spec sheet, minus the marketing

From OpenAI's model documentation:

• Context: 1,050,000 tokens total, with a maximum of 922,000 input tokens and 128,000 output tokens.
• Modalities: text and image in, text out. No audio or video.
• Knowledge cutoff: June 1, 2026.
• Reasoning effort: none, low, medium (default), high, xhigh and max.
• Hosted tools (Responses API): web search, file search, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search, among others.
• Not supported: fine-tuning and predicted outputs.

Two details are easy to miss. First, tool calling requires the Responses API. Chat Completions works for plain prompts, but if your agent loop depends on function calls and still runs on /v1/chat/completions, switching the model string alone will not be enough. Second, the only snapshot listed is the gpt-6.1-terra alias. There is no dated snapshot to pin yet, so behaviour can shift under you; keep your eval suite running against it.

Pricing: what half of Sol actually means

Prices as of Oct 27, 2026, from OpenAI's API pricing page, per million tokens at Standard tier and short context:

• GPT-6.1 Terra: $1.00 input, $0.05 cached input, $5.00 output.
• GPT-6.1 Sol: $2.00 input, $0.10 cached input, $10.00 output.
• GPT-5.6 Terra (previous Terra): $2.00 input, $0.20 cached input, $12.00 output.
• GPT-6 Astra: $10.00 input, $1.00 cached input, $50.00 output.
• GPT-6 Luna: $0.10 input, $0.01 cached input, $0.50 output.

The modifiers matter as much as the headline. Prompts over 272K input tokens are billed at 2x input and cache rates and 1.5x output for the whole request, so a long-context Terra call costs $2 input and $7.50 output per million. Batch and Flex are 50% off Standard. Fast mode doubles the price, regional processing adds 10%, and cache writes cost $1.25 per million for 30-minute retention or $2.50 for 12-hour retention.

A worked cost example

Here is an illustrative calculation, not a measurement. Take a coding-agent task that sends 150K input tokens, 120K of which hit the prompt cache, and produces 10K output tokens (reasoning tokens count as output). Ignoring cache-write charges, the per-task cost at Standard rates works out to:

• GPT-6.1 Terra: about $0.086
• GPT-6.1 Sol: about $0.172
• GPT-5.6 Terra: about $0.204
• GPT-6 Astra: about $0.92
• GPT-6 Luna: about $0.009

At 1,000 such tasks a day, moving from GPT-5.6 Terra to GPT-6.1 Terra cuts the bill from roughly $204 to $86. The bigger lever is output: Terra's $5 output rate is less than half of GPT-5.6 Terra's $12, and agents that think at high or max effort spend most of their budget there. The catch is that the token count per task is not fixed. A model that reasons longer or takes more turns can erase a per-token discount, which is why OpenAI frames its own claims in cost per task.

What OpenAI's benchmarks claim

Every number in this section comes from OpenAI's launch post (October 26, 2026). These are vendor-run evaluations, several on internal or new benchmarks, and they are not directly comparable to third-party leaderboards or to scores run with different harnesses.

• Terminal-Bench 4.0: 50.3% at max reasoning effort, versus 21.5% for GPT-5.6 Terra. OpenAI adds that Sol and Astra remain the better choice for the most complex coding tasks.
• Agents' Last Exam (computer use): 15 percentage points above GPT-5.6 Terra at medium effort, at about a third of the cost, and roughly on par with GPT-6 Astra.
• Speed: more than three times faster at completing tasks than GPT-6 Astra and GPT-6.1 Sol.
• Factuality (internal benchmark): error rate at low effort fell from 53.6% with GPT-5.6 Terra to 16.7%; at max effort it is 4.3%.
• Versus Anthropic: OpenAI claims a roughly 9x lower cost per task than Claude Opus 5.5 at max effort on Agents' Last Exam.

Read the Terminal-Bench figure carefully. It compares Terra at max effort with a published score for its predecessor, and max effort is the most expensive way to run the model. If you want to know what Terminal-Bench actually measures and why it diverges from SWE-bench, see our SWE-bench vs Terminal-Bench explainer. The competitive claim against Claude Opus 5.5 is OpenAI's own cost model; treat it as a hypothesis to test on your workload, not a result.

Where to run it: API, Copilot or Bedrock

OpenAI API

The direct API gives you the full hosted-tool set, Batch and Flex discounts, and the rate card above. Rate limits start at 5,000 requests and 1 million tokens per minute on the Build tier, rising to 40 million tokens per minute on Grow. An illustrative Responses API request, with low effort for a routine task:

curl https://api.openai.com/v1/responses -H "Authorization: Bearer $OPENAI_API_KEY" -H "Content-Type: application/json" -d '{"model": "gpt-6.1-terra", "reasoning": {"effort": "low"}, "input": "Summarize the failing test output and propose a fix."}'

Start at low or medium for triage, extraction and routine edits, and only escalate to high or max for tasks that fail at lower settings. That keeps the output-token bill, where most of the cost sits, under control.

GitHub Copilot

In GitHub Copilot, GitHub says Terra is billed at provider list pricing under usage-based billing; check GitHub's models and pricing reference for how that maps to your plan. One operational note for Copilot Business and Enterprise admins: under default model enablement, new models are switched on automatically unless you have turned off the global default or disabled this model in the Copilot model policy. If you control model spend centrally, review that policy today.

Amazon Bedrock and Foundry

On Amazon Bedrock, Terra is generally available through the standard Bedrock APIs, with IAM, logging and invocation auditing you already use. AWS points to the Bedrock documentation for Regions, inference profiles and pricing rather than restating rates in the announcement, so confirm the Bedrock price for your Region before assuming parity with OpenAI's direct pricing. The interesting part is that AWS lets you configure ChatGPT Work and Codex to use the Bedrock-hosted model, which matters for teams whose data rules require inference inside their AWS account. Microsoft Foundry is the equivalent path for Azure shops. If you are routing across providers through a gateway, our AI gateways comparison covers the trade-offs.

Pros and cons for builders

• Pro: half the per-token price of GPT-6.1 Sol, and less than half of GPT-5.6 Terra's output rate.
• Pro: same-day availability across the API, Copilot, Bedrock and Foundry, so you can adopt it without changing platforms.
• Pro: a full hosted-tool set, including computer use, MCP and apply patch, in the Responses API.
• Con: tool calling needs the Responses API, which forces a migration for agents still on Chat Completions.
• Con: no dated snapshot yet, so output can drift while you depend on it.
• Con: the headline benchmark gains are vendor-reported and often quoted at max effort, the costliest setting.
• Con: long prompts above 272K tokens double the input rate, which narrows the gap for whole-repo context dumps.

Our take

GPT-6.1 Terra is the most important OpenAI release this month for anyone paying per token. It does not beat Sol or Astra on the hardest problems, and OpenAI says so. What it does is make the cheap-enough tier much better at the work agents spend most of their time on: reading screens, running commands, editing code and writing. For CI bots, triage agents, document extraction and Copilot sessions on well-scoped tasks, make Terra the default and keep Sol or Astra as an escalation path. Before you switch production traffic, migrate tool-calling code to the Responses API, run your own evals at low and medium effort, and measure cost per completed task rather than cost per token.

Sources

• https://openai.com/index/introducing-gpt-6-1-terra/ (OpenAI, Introducing GPT-6.1 Terra, Oct 26, 2026)
• https://developers.openai.com/api/docs/models/gpt-6.1-terra (OpenAI, GPT-6.1 Terra model documentation)
• https://developers.openai.com/api/docs/pricing (OpenAI, API pricing)
• https://github.blog/changelog/2026-10-26-gpt-6-1-terra-in-github-copilot/ (GitHub Changelog, Oct 26, 2026)
• https://aws.amazon.com/about-aws/whats-new/2026/10/openai-gpt-6-1-terra-on-amazon-bedrock/ (AWS What's New, Oct 26, 2026)
• https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/gpt-6-1-terra-now-generally-available-in-microsoft-foundry-build-agents-that-see/4567769 (Microsoft Foundry blog)
• https://www.datastudios.org/post/openai-gpt-6-1-terra-computer-use-chatgpt-codex-api-pricing (Data Studios writeup)