Terra 5.6 high cuts Clawsweeper GitHub review time by ~40%, Steipete says

• Terra high replaced the ClawSweeper review bot because it ran about 40% faster overall with negligible quality loss, according to steipete's cutover note.
• The tier ordering got weird in real work: Terra high beat Sol low on 10 real review items, and steipete's follow-up said Sol would have falsely closed one live bug.
• Extra-high reasoning did not pay off in this review harness, as steipete's cutover note said xhigh erased the speed win without a noticeable quality gain.
• The bigger workflow idea was persistent repo memory, not one-off PR comments: in the Molty screenshot, the bot argues for an agent that remembers recurring repo patterns.
• The cutover happened under load: steipete's backlog replies said live operations caused a massive backlog, tokens were not unlimited, and servers were hammered.

OpenAI's GPT-5.6 launch post framed Terra as the balanced model between Sol and Luna. The OpenAI model docs put Terra at a 1,050,000-token context window, 128,000 max output tokens, and $2.50 input/$15 output per million tokens. The ClawSweeper README describes the bot as a conservative maintenance reviewer for issues and PRs, while CodeRabbit's Sol and Terra benchmark still treats Terra as the cheaper lane for scoped work. Steipete's result is the fun kind of eval mess: the cheaper lane won the actual job.

Steipete said he moved ClawSweeper, a GitHub review bot, to 5.6 Terra high and saw about 40% faster reviews with negligible quality loss.

The price curve explains why that change matters at bot scale. OpenAI's pricing page lists gpt-5.6-terra at $2.50 input and $15 output per million tokens, while gpt-5.6-sol is $5 and $30.

ClawSweeper is not a toy demo in the repo description. The project README says it reviews open issues and PRs on schedules and exact GitHub events, writes durable reports, syncs public review comments, and only closes unchanged high-confidence proposals.

The follow-up test compared Terra high with Sol low on real review items:

• Terra was faster on 10 of 10 items.
• Overall speed improved by roughly 40%.
• Decisions matched on 8 of 10 items.
• On PR #110103, Sol would have closed a live bug as “implemented on main.”

Steipete's short version was “don’t trust benchmarks” for issue and code review. That lands because the mismatch was not Terra beating an older model, it was Terra high beating Sol low inside the job ClawSweeper actually runs.

Steipete said the useful work was finding the right model setting instead of “yoloing everything on max.” His cutover note said Terra xhigh removed the performance win and did not make a noticeable difference in review evals steipete's cutover note.

That matches a broader developer friction point around GPT-5.6 reasoning controls. In an OpenAI Developer Community thread, one Codex user described an eight-hour Terra Medium run where cost attribution was hard to separate across reasoning, repo context, images, tool calls, repeated reads, and output.

Molty called ClawSweeper advisory: a bad comment can be ignored, humans still review everything, and rollback is one config change.

That made the Terra move unusually easy to justify. The bot's job is to comment, classify, and propose guarded cleanup, not silently ship code.

The same Discord exchange reframed the speedup as fuel for a persistent repo agent. In the Molty screenshot, the “real unlock” is an agent that remembers:

• which file breaks every third PR,
• which contributor forgets the changelog,
• which subsystem is mid-migration,
• which old pattern it should stop suggesting.

The line worth stealing: faster and cheaper Terra makes it viable to run that memory loop continuously per repo instead of per PR.

The operational reason was blunt. Steipete said live operations had created a massive backlog, tokens were not really unlimited, servers were hammered, and the move to Terra was part of making the bot faster.

ClawSweeper's scheduler docs split work into exact event review, hot intake, and normal backfill lanes, with scheduled OpenClaw review running across active shards. A later reply from steipete on Terra priority added that Terra had become much faster because it was running in a priority queue.