Outlier vs GitHub Copilot: local coding AI vs cloud coding AI

I built Outlier, so treat this as a comparison written by an interested party. Every claim about Copilot comes from GitHub’s own docs, cited at the bottom, and I’ve been blunt about where Copilot is the better tool. Copilot is a cloud service that reaches into your editor; Outlier is a local app that reaches into nothing.

Copilot is everywhere; Outlier is one window

Copilot’s surface area is the part no local Mac app matches. GitHub’s feature matrix lists six editors — VS Code, Visual Studio, JetBrains, Eclipse, Xcode, NeoVim — with uneven coverage: VS Code and Visual Studio get completions, chat, agent mode and next edit suggestions, while NeoVim gets code completion only, no chat and no agent mode. Beyond editors there’s a terminal CLI, a native desktop app for macOS, Linux and Windows, plus GitHub.com, GitHub Mobile and Windows Terminal — one subscription.

Outlier is one app: native macOS, Apple Silicon, macOS 12 or later. No Windows, no Linux, no mobile, no editor extension. You don’t get ghost text as you type, and you move code between the app and your editor yourself — a real ergonomic cost. What you get back is one process, no network.

Which model is doing the work

Copilot routes to other companies’ models: OpenAI’s GPT-5 family, Anthropic’s Claude Haiku 4.5 through Opus 4.8, Google’s Gemini 3.x, GitHub’s Raptor mini, Microsoft’s MAI-Code-1-Flash, Moonshot’s Kimi K2.7 Code — a lot of frontier capability behind one login. You can’t download those weights, pin a version, or archive one, and the free tier offers no model choice at all: auto selection only.

Outlier ships seven tiers you download once, from Nano 4B up to Plus 397B-a17b, including Code 27B (15.13 GB, 24 GB RAM). Six are Apache 2.0, Quick is under the Gemma Terms of Use, and the weights are published openly — what you’ve downloaded can’t be deprecated out from under you.

What Copilot sends to GitHub, and what it trains on

With Copilot, yes, and GitHub documents where: OpenAI models on OpenAI’s and GitHub’s Azure infrastructure, Anthropic models on AWS, Anthropic and Google Cloud, Google’s on Google Cloud. GitHub holds zero-data-retention agreements with OpenAI and Anthropic, and says Google commits not to train on GitHub data — better than a default API relationship.

The training policy needs a careful read. Since April 24, 2026, on Free, Pro, Pro+ and Max, GitHub may use your inputs, outputs, code snippets and surrounding context — code around the cursor, file names, repository structure — to train its models, on by default. The opt-out is real but manual. Business and Enterprise data is excluded by contract, so the strongest protection goes to buyers with procurement, not the individual on a $100 Max plan.

VS Code does support local models through Ollama or LM Studio for chat, which its docs say need no account, no plan and no internet. But those same docs state you can’t use a local model for inline suggestions, and semantic search still needs the cloud. With Outlier the question doesn’t arise.

Cost, and how predictable it is

Copilot Free is $0 for up to 2,000 completions and 50 chat requests a month. Pro is $10/user/month with $15 in credits, Pro+ $39 with $70, Max $100 with $249. One credit is $0.01, billing is per model, and published input prices run $0.20 to $10.00 per million tokens. Paid plans include unlimited completions, which don’t consume credits.

Copilot Pro at $10 and Outlier Pro at $9 are a dollar apart, which is not a reason to choose either one. What changed is predictability: on June 1, 2026 GitHub replaced request-based billing with usage-based billing, so spend now tracks which model you pick. Annual billing is winding down too — GitHub’s docs say an account is automatically downgraded to Copilot Free when its annual plan ends. Business is $19 and Enterprise $39 per granted seat per month — GitHub’s marketing page routes you to a sales contact, but the docs list both.

Outlier’s free tier is Nano and Lite, no account, no caps. Pro covers all seven tiers, Founders Lifetime at $249. That number doesn’t move with usage — heavy days cost electricity, not credits.

Agentic ability and repo-scale context

This is Copilot’s strongest ground. Agent mode runs in the major editors, cloud agents run on GitHub’s infrastructure and open pull requests, and the desktop app runs several isolated sessions at once, each with its own git worktree and branch, so parallel agents don’t collide. Cloud sandboxes, in public preview, keep long runs off your battery.

Now my numbers, unflattering part first. Code 27B scores 0.866 on HumanEval. On a blind slice of SWE-bench Verified, the local 27B resolved 18 of 40 issues, about 45% — measured in my own harness, not a leaderboard submission, and below what the strongest cloud coding agents report. Core 27B separately matched Claude Opus on 98.9% of rubric checks across 54 prompts. Strong on well-specified work, behind on long autonomous repo work, and there’s no GitHub integration at all.

Copilot and Outlier, line by line

Where GitHub Copilot genuinely wins

• Raw capability ceiling. The largest models Copilot serves are far beyond anything that fits in 64 GB of unified memory. On hard reasoning that’s a durable advantage.
• It’s in your editor. Inline completions while you type are a different workflow from switching to an app. If that’s how you work, nothing local replaces it today.
• Multi-vendor access on one bill. Five vendors’ models, no separate accounts or API keys.
• The GitHub side. Pull requests, issues, automated review, agents that open PRs for you. A local app has no equivalent — I’m not going to build a code host.
• Cloud runs don’t cost you hardware. A long agent session in a GitHub-hosted sandbox doesn’t heat your laptop or eat your RAM. On a MacBook Air, that matters.
• Governance, and a free tier that isn’t a demo. Content exclusion, org policy, audit logs, a Data Protection Agreement, a duplication filter for public-code matches of 65 lexemes or more, and 2,000 completions a month at $0.
• Auditable in part. Copilot for Xcode is published under MIT.

GitHub documents the limits too: Copilot Chat may produce code that looks valid but isn’t, and quality drops on obscure languages. The headline 55%-faster figure and the claim that 60–75% of developers felt more fulfilled and less frustrated are GitHub’s own research, not independent studies — and the second is a share of survey respondents, not a size of improvement.

Which one fits your workflow

Pick GitHub Copilot if you want completions in your editor, you live inside GitHub, you need a frontier model on hard problems, you’re on Windows or Linux, or you want agents opening pull requests while you do something else. Business and Enterprise also get the training exclusion in writing.

Pick Outlier if your code can’t leave the machine — client work under NDA, regulated codebases, security research. Also if you want a flat bill instead of metered credits, or you work offline.

Use both, which is what I do: Outlier for the sensitive and the routine, Copilot for what you’re stuck on. There’s no integration between them — one is an app, the other is an editor extension, and you move the code across yourself.