Summary
OpenAI released GPT-6.1 Terra, a midrange model for visual reasoning, computer use and technical writing. API pricing is $1 per million input tokens and $5 per million output tokens, which OpenAI says is 50% below GPT-5.6 Terra and GPT-6.1 Sol. The same day it reached ChatGPT Work, Codex, the API, GitHub Copilot and Netlify's AI Gateway and Agent Runners. The model may well be good. Our point is narrower: many teams will first get it through intermediaries whose defaults, billing and limits they don't control, and the only performance numbers in our evidence come from the vendors.
What happened
OpenAI pitches GPT-6.1 Terra as a fast, efficient model for workday tasks: drafting proposals, reading technical drawings, filling in forms and editing video. Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. OpenAI reports a 15-percentage-point gain over GPT-5.6 Terra on Agents' Last Exam at medium reasoning effort, at about one-third of the cost. It also reports 50.3% on Terminal-Bench 4.0 at maximum effort, against 21.5% published for GPT-5.6 Terra. Simon Willison notes the price is about ten times that of GPT-6 Luna. He quotes OpenAI's Ted Sanders saying on Hacker News that Terra is faster than Luna for vision and computer-use work, stronger there, and cheaper than Sol. OpenAI says the same ChatGPT Work and Codex allowance now covers more tasks than with GPT-5.6 Terra, but it doesn't publish per-plan numbers.
GitHub's October 26, 2026 changelog makes Terra generally available to Copilot Pro, Pro+, Max, Business and Enterprise users. It covers VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the CLI, the cloud agent, the Copilot app, github.com and Mobile, and the rollout is gradual. In the changelog's terms as summarized in our pack, usage-based billing is at "provider list pricing", with no amounts given. Business and Enterprise admins control access through Copilot model policy, and new models are enabled by default unless the global default is disabled or the model is explicitly disabled. GitHub says early testing solved more terminal and C# tasks than GPT-5.6 Terra, but gives no method or counts. Netlify's changelog, also dated October 26, lists gpt-6.1-terra in its AI Gateway and Agent Runners and says the gateway removes the need for separate API-key management. The raw Netlify excerpt we hold is cut off before the Terra entry, so everything we say about it here relies on the pack's summary.
What the archive adds
The archive mostly repeats the launch materials. A fragment of OpenAI's announcement page, cut off at the start, says Terra completes some task set "over three times faster" than GPT-6 Astra and GPT-6.1 Sol. The benchmark name is lost in the cut. The page plots Terra against GPT-5.6 Terra, GPT-6.1 Sol, GPT-6 Astra, GPT-6 Luna and Claude Opus 5.5 on a cost-versus-score chart for Agents' Last Exam v1. It says Terra beats GPT-5.6 Terra on MMMU-Pro at every reasoning setting, and that sentence is also truncated. All of these are vendor-run evaluations.
The dossier helps on pricing. According to OpenAI's GPT-5.6 efficiency post, cutting tokens per task and improving inference serving (load balancing, speculative decoding, caching, kernel optimization) are ways to lower cost. According to OpenAI's "The Work Now Within Reach", inference-system improvements raised token-generation efficiency by more than 15%. Both are OpenAI's own claims. They make a lower price plausible but don't prove the cut is cost-driven. The dossier also points to allowance volatility. According to The Next Web's DevDay coverage, OpenAI cut the included Work capacity of its $200 Pro tier and launched a $500 plan. That suggests "more tasks per allowance" describes terms OpenAI has recently been willing to reset. The editor asked us to cite an independent Artificial Analysis score and a third-party finding that OpenAI's claims weren't reproduced. Neither appears in the dossier we received, so we don't state those figures here.
Trend signal
This launch put one model onto the lab's own surfaces and two partner platforms on the same day. Coding assistants, deploy platforms and AI gateways are becoming the place where developers first meet a new model. That changes what a launch means in practice. It used to be a new model ID you chose to adopt. On hosts with default-on policies, it can now arrive inside tools you already pay for, on the host's terms. Terra is a sharp case because it targets high-volume agent work: terminal sessions, form filling, desktop control. In that kind of work, small differences in behaviour or price add up across many runs.
Why now
1) Serving efficiency. According to OpenAI's GPT-5.6 efficiency post (d1) and "The Work Now Within Reach" (d2), fewer tokens per task and inference optimizations, including a reported 15%+ gain in token-generation efficiency, lower OpenAI's costs. That makes a 50% list-price cut plausible, though not proven to be cost-driven; competitive pricing could also be part of it. 2) Product focus on agent workloads. The launch is built around computer use, terminal tasks and form filling, where cost per task decides whether automation pays off (primary). 3) Day-one partner availability. GitHub and Netlify listed Terra on launch day, and on Copilot it is default-on unless an admin opted out (primary). Why the partners shipped the same day, and on what commercial terms, isn't in our evidence.
Connections to other events
Read together, the four sources describe one model under four different sets of terms:
- OpenAI API. The price is visible ($1/$5, $0.05 cached), and you choose the model ID yourself.
- ChatGPT Work and Codex. You get "more tasks" per allowance, with no per-plan numbers. The DevDay change to the Pro tier shows these allowances can shrink.
- GitHub Copilot. For Business and Enterprise orgs, the changelog says new models are enabled by default unless the global default or the model is disabled. Billing is "provider list pricing", with no amounts given. Rollout is gradual, so engineers on the same team may see different model pickers this week.
- Netlify. The changelog summary says the gateway handles keys and authentication and applies its own caching, rate limiting and auth. It gives no price, limits or plan terms, and our raw excerpt is truncated.
The platform risk here isn't lock-in through quality. It's the host setting your defaults. In an org that never set a Copilot model policy, Terra becomes selectable in engineers' agent tooling as the rollout reaches them. That is true whether or not anyone has run it on the org's code. The evidence available at that point is OpenAI's benchmarks and GitHub's unquantified "more terminal and C# tasks solved". The headline 50.3% Terminal-Bench figure is measured at maximum reasoning effort, while the cost-efficiency claim is made at medium effort. A team picking Terra from a menu doesn't necessarily know which effort setting the host uses, so the published numbers may not describe the configuration they actually run. Our read is that for cheap computer-use models, the governance risk (who switched it on, at what setting, under what limits) currently weighs more than the model risk.
Why it matters
For builders: the per-token price matters less than knowing which layer decides what model runs, and at what effort. Pin model IDs and reasoning-effort settings in your own code. Treat a gateway as an extra party in your dependency chain, and read its documented limits before scheduling agent jobs through it. For computer-use or form-filling agents, Terra's price makes high-volume automation tempting. But no source here reports reliability for computer-control tasks, so keep human review on irreversible actions.
For buyers and admins: check the Copilot model policy now. Per the changelog, the default-on behaviour applies to new models in general, not only to Terra. Ask GitHub and Netlify what Terra actually costs on their surface, because neither changelog gives a number. If you budget on ChatGPT Work or Codex allowances, plan as if they can change, because OpenAI has recently cut one tier's included capacity.
What happens next
- GitHub makes GPT-6.1 Terra the default or auto-selected model on at least one Copilot surface (cloud agent, CLI or auto model selection). (6 months, 35%). Watch for: A GitHub changelog entry naming Terra as the default or auto-mode model for a Copilot surface.
- An independent evaluation publishes a Terminal-Bench 4.0 result for GPT-6.1 Terra at maximum effort that is more than 5 percentage points below OpenAI's reported 50.3%. (6 months, 35%). Watch for: A public leaderboard or third-party eval post listing Terra on Terminal-Bench 4.0 with harness details and a score under 45.3%.
- OpenAI announces a shutdown date for GPT-5.6 Terra in the API. (12 months, 50%). Watch for: GPT-5.6 Terra appearing on OpenAI's API deprecations page with a shutdown date.
Limits and uncertainty
Every performance claim in our evidence comes from OpenAI or GitHub. That doesn't mean independent evaluations don't exist. Third-party leaderboards may already list Terra, but none are in the material we were given, so we can't say whether OpenAI's numbers hold up. If independent results confirm the gains, default-on is mostly a convenience and our concern shrinks. If they show Terra weaker at the effort settings hosts actually use, the concern grows. We don't know context-window size, task-completion reliability, or which reasoning effort Copilot and Netlify apply by default. The Netlify and GitHub excerpts we hold are truncated, so our quoted terms come from summaries of those changelogs. We don't know how many enterprises have already disabled Copilot's global default. Finally, OpenAI's efficiency posts suggest its serving costs are falling, which argues against sudden price hikes. The risk we describe is about defaults and terms, not a predicted price increase.
Skill to build now
Build a model-change regression check for your agents. First exercise: pick 20 real tasks from your backlog, such as terminal fixes, C# edits or form fills. Run them through your current model and through gpt-6.1-terra, with the model ID pinned and reasoning effort set to both medium and maximum. Record pass/fail, tokens used, cached-input share and cost for each run. Rerun the set whenever Copilot or your gateway announces a model change.
Reader takeaway
This week, open your GitHub Copilot model policy and decide deliberately whether new models stay on by default. Then pin gpt-6.1-terra, or whichever model you choose, together with its reasoning effort, in every agent, function and gateway call you own. Don't move computer-use or terminal automation to Terra until it passes a small regression set built from your own tasks, at the effort level you will actually pay for.
Open question
When a coding assistant enables a new third-party model by default and it performs worse on a team's codebase, or takes a costly agent action, who is responsible: the model vendor, the platform that switched it on, or the admin who left the default unchanged?