GitHub adds GPT-6.1 Sol to Copilot on an unquantified claim of fewer tokens and steps

GitHub's efficiency claim has a direction and a comparison group. According to the announcement, as reported by Superpower Daily, early tests found Sol reliably completed multistep coding tasks while using noticeably fewer tokens and steps than GPT-6 and GPT-5.6 ^[2]. GitHub did not publish the size of the reduction, benchmark details or test-by-test results ^[3]. That leaves "reliably" without a completion rate and "noticeably" without a percentage, and a reader cannot tell how many tasks the early tests covered.

The two measures answer different questions. Tokens count the text the model reads and writes. Steps count the moves it makes through a task ^[10]. Under Copilot's usage-based billing at provider list pricing ^[7], tokens are what the invoice counts. Steps are closer to what a developer waiting on an agent experiences.

A per-job bill is tokens used multiplied by the price per token ^[1]. GitHub has described the first factor only by its direction. For the second, the announcement points customers to its Copilot models and pricing documentation ^[7]. If Sol's per-token rate exceeds the older model's by a larger proportion than its token saving, the job costs more. If the rate is the same or lower, the saving is real and of unknown size ^[1].

In my view the early-test claim justifies running a trial and does not yet justify anything else. The useful trial uses a team's own backlog: the same tasks on Sol and on the model the team runs now, completion checked by a reviewer, token counts recorded for each run. GitHub aimed Sol at multistep jobs and work directed through a terminal ^[9]. Such a job may start in an editor and continue through the command line or the coding agent, as Superpower Daily notes ^[11]. Sol is selectable in the listed editors and in Copilot CLI and the coding agent ^[4][5], so one job can be followed end to end on one model. A test confined to the editor measures part of the work GitHub says Sol is for.

The default decides who acts first. In Business and Enterprise organizations, new models are enabled automatically unless an administrator has switched off the global default or disabled Sol specifically ^[6]. Developers there can start using Sol as the rollout reaches them, before anyone has run that comparison. GitHub called the model generally available when it announced it on September 29, 2026, but the rollout is gradual ^[1]. An admin who wants the trial first therefore has a window of unknown length to disable it ^[8]. Meanwhile a developer who cannot see Sol may be waiting on the rollout or may be blocked by an organization's policy ^[12].