--- type: analysis slug: terra-s-lower-list-price-doesn-t-tell-you-what-copilot-will-cost-you title: "Terra's lower list price doesn't tell you what Copilot will cost you" description: "GPT-6.1 Terra is now on by default for Copilot Business and Enterprise, billed per token at a list price GitHub doesn't state. That price is almost half of Sol's. Our read is that what a team actually saves will depend on how the model performs on its own tasks, and nobody has published that yet." published_at: 2026-10-28T02:00:00.000Z topics: ["github","openai"] source_urls: ["https://github.blog/changelog/2026-10-26-gpt-6-1-terra-in-github-copilot","https://simonwillison.net/2026/Oct/26/gpt-61-terra","https://openai.com/index/introducing-gpt-6-1-terra","https://github.blog/changelog/2026-10-14-hydrafusion-updates-and-new-max-option","https://www.testingcatalog.com/openai-prepares-gpt-6-1-terra-microsoft-code"] author: "Andrey Reshetnikov" author_profile: https://newruntime.com/owner-profile.md --- # Terra's lower list price doesn't tell you what Copilot will cost you GPT-6.1 Terra is now on by default for Copilot Business and Enterprise, billed per token at a list price GitHub doesn't state. That price is almost half of Sol's. Our read is that what a team actually saves will depend on how the model performs on its own tasks, and nobody has published that yet. ## Summary On October 26, 2026, GitHub made OpenAI's GPT-6.1 Terra generally available across Copilot. It's billed by usage at provider list pricing and enabled by default for Business and Enterprise unless an admin turns it off. OpenAI's API price is $1 per million input tokens and $5 per million output tokens, which Simon Willison describes as almost half the price of GPT-5.6 Terra and GPT-6.1 Sol. GitHub's changelog doesn't give the per-model rate inside Copilot or explain how it tested its claim that Terra solves more tasks. That leaves admins deciding with a price signal and no cost data. ## What happened Start with policy. GitHub's changelog says new models in Copilot are enabled by default unless the global default is disabled or the specific model is explicitly disabled. Usage is billed at provider list pricing. So unless an admin acts, GPT-6.1 Terra appears in every Business and Enterprise model picker, and every use of it is charged by usage. GitHub pitches Terra as a lower-cost option than GPT-6.1 Sol for everyday agentic coding and terminal work. Its evidence is that early testing found Terra solved more terminal and C# tasks than GPT-5.6 Terra. The changelog gives no methodology, task counts, absolute results or list-price amounts. We could not find Terra's Copilot rate anywhere in the evidence we have. OpenAI's launch post supplies the API numbers and makes two separate benchmark claims. On Agents' Last Exam, at medium reasoning effort, OpenAI reports a 15-point gain over GPT-5.6 Terra at about one-third the cost. On Terminal-Bench 4.0, at maximum reasoning effort, it reports 50.3% against GPT-5.6 Terra's 21.5%. Neither result has been independently replicated, and the cost comparison is attached only to the Agents' Last Exam figure. OpenAI also positions Terra a little differently from GitHub. Its post leads with visual reasoning, precise computer use and technical writing. Willison adds the useful comparison: Terra is about ten times the price of GPT-6 Luna. So "cheaper" means cheaper than Sol, not cheap. ## What the archive adds The archive shows Terra was signposted before launch. TestingCatalog reported that GPT-6.1 Terra had surfaced in a Microsoft pull request, around the time The Information reported a new Terra model was coming. That suggests Microsoft's tooling was being prepared for it before GA. On billing, Zed's blog has a post on what GitHub Copilot's usage-based billing means for Zed users (d2). It confirms that per-usage charging is now part of how Copilot is consumed outside GitHub's own editors too. As an analogy only, a Caprock guide to OpenCode pricing (d1) describes agentic usage exceeding a $15 monthly model allowance more than six times over in under four weeks. That's a different product, but it shows how fast agent sessions can consume tokens. OpenAI's API pricing page (d3) also lists charges beyond tokens, such as tool calls billed per thousand. Those are API prices. We can't confirm how, or whether, they map to Copilot billing. GitHub's HydraFusion update from October 14 matters here too. That research-preview orchestrator works from a pool of Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. It uses medium reasoning effort for drafts and high effort for review, and its Cascade workflow asks for a repair only when a judging step decides one is needed. ## Trend signal Coding assistants now offer a menu of tiers: Luna at the bottom, Terra in the middle, Sol and Opus-class models at the top, plus orchestrators like HydraFusion that combine several models. Once billing is per usage, every model choice is also a spending choice. Vendors report cost against a benchmark at a fixed reasoning effort. Teams pay for whole sessions, including retries and repeated reads of the same context. The difference between those two numbers is where Copilot budgets are actually decided, and right now only the vendors' half is published. ## Why now 1) Usage-based billing. GitHub bills Terra at provider list pricing, and Zed's post (d2) covers how Copilot's usage-based billing reaches third-party editors. When customers see per-use cost, the price of each model tier becomes a selling point. 2) Agentic workloads use a lot of tokens. In a different product, Caprock's OpenCode guide (d1) shows agent usage outrunning flat allowances within weeks, which gives vendors a reason to offer a cheaper mid-tier model. 3) OpenAI's release cadence. The launch is explicitly framed against GPT-5.6 Terra, at almost half the price with better self-reported scores, which suggests Terra is refreshed on a regular cycle. 4) Integration had started early. According to TestingCatalog, GPT-6.1 Terra surfaced in a Microsoft pull request before launch, which fits with GitHub making it available on OpenAI's announcement day. ## Connections to other events The obvious reading is that a model at almost half the list price makes Copilot cheaper. Read together, the stories give three reasons to hold off on that conclusion. First, the price signal is incomplete. GitHub bills at provider list pricing but doesn't state the amount, and its "more tasks solved" claim has no methodology or task counts. A buyer can't calculate cost per completed task from what has been published. Second, the benchmarks don't support one tidy story. The one-third-cost figure applies to Agents' Last Exam at medium effort. The big jump in coding performance, 50.3% versus 21.5%, applies to Terminal-Bench 4.0 at maximum effort, and OpenAI attaches no cost figure to it. You can't combine the two to conclude that Terra gets the Terminal-Bench result at a third of the cost. Reasoning effort is a setting that changes both quality and spend. Third, GitHub describes Terra as suited to well-scoped tasks. Our inference is that on loosely specified work, the per-token saving may be partly offset by extra turns, and under usage billing extra turns are billed. That's a hypothesis, not a measurement. HydraFusion's design points the same way: GitHub's own answer to cost there is assigning different effort to different stages and repairing only on demand, not simply swapping in a cheaper model. Because Terra is on by default, developers will likely pick it for its "lower-cost" label before any team has checked whether that holds on its own codebase. ## Why it matters For builders: Terra is a reasonable first choice for tasks with a clear finish line, such as a failing test, a shell script or a contained C# change. For open-ended debugging or design, compare it against Sol before switching. Measure cost per accepted result, not cost per token. For admins and buyers: default-on plus usage billing means Terra can shift your spend before anyone has made a decision. Review Copilot model policy now. Decide whether Terra is on for everyone, limited to particular users, or off until it passes an internal comparison. If you can't find Terra's Copilot rate in GitHub's docs or billing view, ask GitHub for it before you roll it out widely. ## What happens next - **GPT-6.1 Terra is added to the HydraFusion or HydraFusion Max model pool.** (6 months, 45%). Watch for: A GitHub changelog entry listing GPT-6.1 Terra among HydraFusion's pool models. - **An independently run Terminal-Bench 4.0 result for GPT-6.1 Terra is published and comes in at least 8 points below OpenAI's reported 50.3%.** (3 months, 35%). Watch for: A third-party leaderboard or harness-maintainer run reporting Terra's score along with its reasoning-effort setting. - **GitHub publicly documents a Copilot per-model rate for GPT-6.1 Terra that a buyer can compare directly with GPT-6.1 Sol.** (3 months, 60%). Watch for: Copilot billing or models docs listing Terra's rate next to Sol's. ## Limits and uncertainty We have no independent benchmarks, no Copilot task-level cost data, and no Copilot credit rate for Terra. The idea that loose task scoping eats into the savings is our inference from GitHub's wording and from how per-usage billing works. It isn't measured. Terra may well be efficient enough per turn to stay cheaper than Sol across real sessions, in which case the price cut does reach users largely intact. HydraFusion has no published pricing or benchmarks, so we can't say whether stage-based routing beats choosing a model by hand. The OpenCode data is from a different product and serves only as an analogy. ## Skill to build now Build a small per-task eval for your own repository that measures cost and quality together. First exercise: take 10 closed issues, five terminal or scripting tasks and five C# or core-language tasks. Write a one-line scope for each, then rerun each in Copilot CLI with Terra and with Sol. Record the usage charged, the number of turns, and whether the diff passed tests without edits. One table of results will tell you more than either vendor's benchmark claims. ## Reader takeaway Before Terra becomes your team's everyday model, run 20 real backlog tasks through both Terra and Sol in Copilot CLI. For each, log the usage charged, the number of agent turns, and whether the change merged without manual edits. Leave Terra on by default only for the task types where it wins on cost per merged result. For the rest, set Sol as the starting choice in your team guidance. ## Open question Will cheaper mid-tier models lower total Copilot spend, or will lower per-token prices lead teams to give agents longer, looser tasks, so that total usage rises anyway? ## Sources - [GPT-6.1 Terra is generally available in GitHub Copilot across IDEs, CLI and cloud agent](https://github.blog/changelog/2026-10-26-gpt-6-1-terra-in-github-copilot) - [GPT-6.1 Terra](https://simonwillison.net/2026/Oct/26/gpt-61-terra) - [OpenAI releases GPT-6.1 Terra, a cheaper fast model for computer use, in ChatGPT Work, Codex and the API](https://openai.com/index/introducing-gpt-6-1-terra) - [HydraFusion updates and new Max option](https://github.blog/changelog/2026-10-14-hydrafusion-updates-and-new-max-option) - [OpenAI prepares GPT-6.1 Terra for the upcoming release](https://www.testingcatalog.com/openai-prepares-gpt-6-1-terra-microsoft-code) - [GitHub Copilot is moving to usage-based billing](https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/) - [How GPT-5.6 fuses frontier intelligence with frontier efficiency](https://openai.com/index/gpt-5-6-frontier-intelligence-efficiency/) - [Models and pricing for GitHub Copilot](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing) - [TERMINAL-BENCH 4.0 leaderboard](https://www.tbench.ai/?efforts=all&version=4.0) - [Claude Opus 5.5 model overview](https://platform.claude.com/docs/en/models/opus-5-5/overview) - [Terminal-Bench 4.0 leaderboard](https://www.tbench.ai/) - [Agents’ Last Exam leaderboard](https://agents-last-exam.org/leaderboard) Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json ## Continue in the graph Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`. - topic hub: [Github](https://newruntime.com/topics/github.md): Explore the github topic hub. - topic hub: [Openai](https://newruntime.com/topics/openai.md): Explore the openai topic hub. - collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis. - collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph. - whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)