GitHub Copilot on X

Most of the talk is praise, and it is practical rather than ecstatic. People credit Copilot with getting work finished: one developer built two tools in parallel in a week that would have taken a month, another says that after initial hesitation about its coding style "there is no going back". They name accurate code suggestions that keep the editor moving, a free plan with roughly two thousand completions a month, the Copilot SDK reachable on that same free plan, and step-by-step help while learning. One person notes that through GitHub Enterprise, Copilot now opens up Claude and Codex models.

The disappointed name specifics: a computer-use release called "very janky", a slide in quality, and pricing. Price is where praise turns conditional. 5 people rate it well and then name the bill, a credit burn they cannot control, or code that still needs checking; one who called it really good has not used it since the price rise. 2 are only trying it out.

Price talk centres on the move to token billing: one post says developer bills went from $29 to $750 a month, another puts it at a $30 base plus a cent a credit, around $400 in a heavy month, and people who liked the product say they dropped it over cost. Right behind, 7 people watch allowances vanish — a Pro quota down to 3% in a day, tokens gone after a few hours on Haiku 4.5, three days of agent work ending when the tokens did.

Below those, the complaints are specific and small. Output misses: a model routed to Sonnet 4.6 after Opus 5.5 was picked, test files that ignore naming conventions, PR reviews called annoying. The app gets in the way — IME input, tabbed review comments, an inbox "barely usable" for open-source work, mobile monitoring gone. Model retirements on 19 October forced pinned settings to be redone and replacements enabled by admins. Long sessions compact and lose context, and one security test says auto-approve raised attack success from 29.2% to 95.6%.

The Cursor comparison runs both ways. One developer calls Cursor "the single worst AI product" for work and says they would rather use Copilot; others sort the two by job, Cursor as the AI-first editor and Copilot as the better fit for GitHub teams and for work next to pull requests and issues.

Gemini is weighed in both directions: some say Copilot beats Gemini at coding, and some treat Gemini as the fallback they moved to or regret buying — one left Copilot over its pricing, bought Gemini Pro and was unhappy with it, and another uses Claude models through Copilot and reaches for Opus on hard work. Codex is the tool people cancel for — one says Copilot can be fully replaced by it and drops the subscription — though another runs Codex and Copilot's free tier side by side. Claude Code draws the quality complaints, and one cancelled their plan after credits ran out.

The task questions are concrete. Two people ask GitHub directly why Copilot cannot write a harness, one asks whether it will bring their site into WCAG compliance using the Copilot app with rough prompts, and another asks the chances of a direct integration for their tools with an Enterprise Copilot subscription.

Model choice is asked in terms of cost and skill: one developer working under a 7,500-credit limit wants to know which models to use for planning and which for coding without losing quality, and a beginner asks whether to pick a flash or pro model for testing inside VS Code. The single subscription question offers Claude Pro, ChatGPT Pro, Cursor Pro and Copilot and asks which one keeps your money.

The docs and harness wish comes as a direct complaint to GitHub — "can't github copilot write harness" and "why can't github copilot write harness" — and as a hope that an agent that has done the thinking could hand over the write-up a person would otherwise have to prepare.

The setup wish is about reach: canvases and the agent harness only work in desktop VS Code in a local Copilot session, not with Claude, Codex, remote or web, and one person on WSL could not open a canvas in a tab. Price is raised by someone whose use is uneven, who wants a free base topped up with paid tokens rather than a flat monthly fee. Others want to watch a running session from GitHub Mobile, to drive sessions remotely the way Codex and Claude Code allow, to see the Japanese IME typing problem fixed in the app, and for the agent to remember more across a long session.