
Factory’s coding agent scored 7.6 percentage points higher than Claude Code in Factory’s Terminal-Bench 4 evaluation. Both were configured for Opus 5.5, and moderation-blocked trials were rerun on Opus 5.
Choosing a coding model is only part of choosing an agent. Factory's coding agents, called Droids, combine that model with tools, context handling and an execution loop for implementation, testing and code review.
Better task completion with the same model
When two coding agents use the same model, their results can still differ. The harness around the model controls instructions, available tools, how results enter context and the execution loop. Factory tunes those components for each supported model.
Factory tested Droid and other harnesses on Terminal-Bench 4 in October 2026, with the same model and high reasoning settings in each comparison. With Opus 5.5, Droid passed 65.2% of tasks versus Claude Code's 57.6%. With GPT-6.1 Sol, Droid scored 57.6% versus Codex's 54.0%. On Opus 5.5, Droid also beat Cursor, OpenCode and Devin, and it took 16% to 26% less time per passed task than every other harness tested. These results are published by Factory.
Get more done with fewer model requests
A higher completion rate matters, but so does the cost of work that actually gets done. In Factory's published comparison, Droid cost 12% less per passed task than Codex when both used GPT-6.1 Sol.
Factory's cost and timing comparisons cover matched tasks both agents passed in all three trials, at model providers' list prices, excluding setup and grading. Moderation-blocked Opus 5.5 trials were rerun on Opus 5.
Factory’s Droids reduce repeated model work through programmatic tool calling. The model writes a short program to run tools, independent calls can run together and raw files, search results and logs remain program variables. The model receives selected output instead of reading every raw result.
In Factory's illustrative replay, the same 15 tool calls required five model requests instead of ten and finished 15.2 seconds sooner. Cache-friendly context and on-demand tool loading address another source of repeated processing.
Choose a model mix that fits the work
Not every engineering task needs the same model. Factory Router selects efficient models for work they can handle and stronger models when needed. Organization-specific guidance can reflect a team's codebase, workflow, model preferences and cost requirements. Developers can also choose a supported model directly.
For the week ending September 13, 2026, Factory reported 63% aggregate inference savings for routed production sessions against published frontier-model pricing. From September 5 to 18, 84% of routed production sessions using Factory-managed inference saved at least half versus published frontier-model rates, with median savings of 72.5%.
This measures model selection across a whole workload, so actual savings depend on each team’s tasks and models.
"Factory Router saved us thousands of dollars in inference costs in one week,"
Teams can use efficient models for routine work like documentation updates and keep stronger models for changes to critical services.
Turn vulnerability findings into tested fixes with Security Factory
Security Factory applies the same agent infrastructure to a different bottleneck: unresolved security findings. It brings together signals from your existing security tools, investigates which vulnerabilities are reachable, and carries the evidence into remediation. The result is a tested change your engineers can review.

Factory’s Droids can develop code or dependency changes, test them against the original exploit conditions, check for regressions and prepare patches or pull requests for review. Security teams retain control through policies, audit trails and configured autonomy levels.
In one example Factory shares, a team estimated four weeks to clear its open vulnerabilities and finished in five days, with 50–65% less engineering time.
Teams can request access through Factory's account team.
Fit agents into your engineering environment
For teams with stricter infrastructure requirements, Factory Private supports Droid sandboxes in a customer-controlled VPC through OpenCompute. Code execution stays there; inference and session storage follow the configured model and data settings. Factory hosts the control plane by default.
Blackstone's public case study makes those requirements concrete. The firm chose Factory for deployment flexibility, model choice and fit with its governance and security controls. Its teams use Droids for code review, testing, documentation and larger engineering tasks. Onboarding a new developer went from up to two weeks to minutes, and a proof of concept that would have taken about a month took one engineer roughly a week and a half.
"It was absolutely critical to ensure that agents ran inside of our data center, under our control,"
Evaluate Factory on your own work
Pick a real task your team knows well and compare the result, the checks, the model cost and the time it took.



