Pooled criteria met
Legal task completion, contract provisions, case citations and model judging.
Cortega Legal met 91.4% of criteria in the October 2026 public-suite run. 19.7% of tasks met every criterion, at $0.24 per task in tokens at list price.
Pooled criteria met
Tasks meeting every criterion
Token cost per task
Cortega Legal used GLM-5.3 Flash as its primary model and completed every criterion on 19.7% of 1,743 public Harvey LAB tasks. For that model, Vals AI reports 6.67% task pass rate, Mercor reports 11.1% ± 3.4% (Max), and Artificial Analysis reports 0.8% on its hallucination-gated v1.1 metric.
These are different evaluations, not a controlled comparison of Cortega against direct model access. Task sets, settings and grading differ; Artificial Analysis adds a material-hallucination gate. Sources checked 19 October 2026.
Cortega graded 1,743 public tasks. Vals uses held-out tasks. Mercor lists 1,749 public tasks. Artificial Analysis uses 120 private tasks and a material-hallucination gate. These differences prevent attributing score differences to the workflow alone. The Cortega result reports rubric completion before the separate citation check.
Cortega Legal's results were graded by Cortega with an automated marker. Run-to-run variation is about three points on a sample of 100 tasks. Cost is the tokens used at the model's list price and does not include verification.
The benchmark configuration used GLM Flash as its primary model to examine quality at a low token cost. Cortega’s inline AI quality boosters sit within its governance layers: model routing, configurable independent drafting, review for missing work, and source checks. Different model roles collaborate on the work rather than treating the first generated answer as the final result.
Use the supplied record and connected research sources. Independent drafts and a merge can bring different points into the work.
Review omissions, examine quotations and check available citation records. Source-check status remains distinct from rubric completion.
Assign models to the work through governed routes. Evaluate quality, full workflow cost and deployment requirements together.
In Cortega’s separate citation study, GLM-5.3 Flash’s wrong-citation count fell from 140 before verification and revision to 1 afterward. Supported citation occurrences also fell from 336 to 181, and the initial run returned 24 empty answers out of 100 questions. This measures a reduction in citation errors alongside a loss of coverage; it does not establish complete research or the contribution of every quality booster. Read the study.
The published GLM Flash scores above are lower than Cortega’s recorded complete-task rate. They show what those specific evaluations achieved; they do not establish a ceiling on the model or prove that Cortega’s boosters caused the difference. Measuring that contribution requires the same tasks, model settings and grader with the boosters enabled and disabled. That controlled comparison is not reported here.
In Cortega’s separate 100-contract extraction study, GLM-5.3 Flash and Claude Opus 5.5 each received 74 correct-contract results. Recorded token costs were $0.204 and $8.739 per 100 contracts, respectively. This is a task-specific result, not evidence of overall model parity or an isolated measurement of the boosters. Read the study and limitations.
Those are alternative model choices to evaluate, not guaranteed upgrades. A more capable model may improve some tasks while changing cost and latency; a matched Cortega run is needed to measure the tradeoff. The reported GLM Flash run does not establish the best possible cost-benefit ratio across models.
A suitably licensed open-weight model can run on your own infrastructure; GLM-5.3 Flash’s model card provides its licence and local serving options. A proprietary model accessed through an external provider cannot be self-hosted in the same way; its network access and data terms must be assessed separately. An air-gapped deployment requires all model, research and verification dependencies to operate within that boundary. Review deployment choices.
Explore the results by practice area.