Side by side
| Dimension | Trajectory | Orchestra |
|---|---|---|
| Primary job | TrajectorySpecialized models for specific use cases. Its field reports include post-training an open model on Harvey's legal benchmark. Its platform turns user edits, retries, and accepts into training signal.2,1 | OrchestraEvaluates production work, improves the prompts and routes behind it, and trains specialist models your team owns.5 |
| Who it serves | TrajectoryTeams with one hard, specialized problem who want a research partner working beside them. | OrchestraEvery department, from technical and non-technical product teams to engineering and machine learning teams. |
| How the work gets done | TrajectoryHand in hand with each customer, led by researchers. Trajectory connects to a product through an SDK that captures traces, corrections, and edits.1 | OrchestraAutomated across workflows, with little hands-on work, in under seven days. Your existing OpenAI or Anthropic SDK points at Orchestra's gateway, and approved changes roll out gradually with the prior route kept as a fallback.5 |
| Evaluation approach | TrajectoryTreats user behavior as a better measure of success than benchmarks. Its field reports grade against expert rubrics.1,2 | OrchestraBuilds task-specific benchmarks, such as whether the model updated the right record or followed your policy, and compares models on the same held-out work.5 |
| Model training | TrajectoryPost-trains open models such as NVIDIA Nemotron and updates the model, prompts, and harness from real usage.2,1,3 | OrchestraRepairs prompts, schemas, parsers, and tool contracts first. For a recurring task, trains a specialist from accepted traces and expert corrections.5 |
Trajectory builds specialized models.Orchestra measures, improves, and trains for every team.
Trajectory is the high-touch choice when one hard problem needs a specialized model. Orchestra gives every company specialized models, along with the measurement and improvement that decide when one is worth training.
What Trajectory offers
- Researchers who work closely with your team on a specialized use case.
- Post-training of open models, as in its field report on Harvey's legal benchmark.2
- An SDK that captures traces, corrections, re-prompts, and edits from your product.1
- Training that updates the model, prompts, and harness from real usage.1
What Orchestra offers
- Support for every department, technical and non-technical.
- Builds task-specific benchmarks, such as whether the model updated the right record or followed your policy, and compares models on the same held-out work.5
- Repairs prompts, schemas, parsers, and tool contracts first. For a recurring task, trains a specialist from accepted traces and expert corrections.5
- Approved changes roll out gradually, with the prior route kept as a fallback.5
Where Orchestra is different
Built for scale
Orchestra measures, improves, and trains for every workflow automatically, with little hands-on work from people, in under seven days. Specialized models do not depend on a dedicated research team for each project.5
A platform for every team
Technical and non-technical product teams, engineering, and machine learning teams all work in Orchestra. Trajectory fits best as a partner for one specialized problem.
Success defined before training
Orchestra writes the examples, rubrics, and checks that define success for each workflow, then compares models on the same held-out work.5
When each approach makes sense
Trajectory makes sense when
- You have one hard, specialized problem and want researchers working on it with you.
- You want an open model post-trained for a narrow domain, such as legal work.2
Orchestra makes sense when
- You want specialized models across many workflows, not one project at a time.
- Teams across the company, technical and non-technical, rely on AI in production.
- You know what you want from the work, and every change is checked against that goal on your own examples.5
- You want to improve prompts and routes, then train a specialist you own.5
Orchestra case studies
- Orchestra case study
Teaching a CRM agent to finish the job
75.5%lower model cost than Sonnet 4.6
Model spend per seven-task AutomationBench CRM run: $0.27 for Qwen 3.6 Plus with a revised prompt, $1.12 for Sonnet 4.6. The revised prompt also scored 13.1% higher on mean task score.
- Orchestra case study
Cutting response times for everyday operations
5.2×faster median model response than Sonnet 4.6
The small model passed every task check on 92.2% of attempts; Sonnet passed 100%. Both ran 30 AutomationBench workflows, with 90 attempts per model.
- Orchestra case study
Finding purchase intent in 40,000 comments
$2.82model cost for 39,962 comments
An Orchestra post-trained 30B open model labeled 39,962 YouTube comments from a Snowflake table and returned 39,949 valid outputs. Sonnet cost $12.48 on the same table.
These are Orchestra experiments that compare models on specific workloads. None of them compare Orchestra with Trajectory.
Sources (8)ShowHide
Sources reviewed October 2026. Trajectory statements come from its own website and documentation. If something is not described there, we say so instead of assuming it is missing.
- 1.Trajectory homepage: The platform for continual learning trajectory.ai (opens in a new tab)
- 2.Trajectory field report: Post-training NVIDIA Nemotron 3.5 Super VL on Harvey LAB trajectory.ai/field-notes/harvey-nemotron-3-5-super (opens in a new tab)
- 3.Trajectory field report: Continual learning cheap enough to meter trajectory.ai/field-notes/continual-learning-cheap-enough-to-meter (opens in a new tab)
- 4.Orchestra homepage
- 5.Orchestra platform
- 6.Case study: Teaching a CRM agent to finish the job
- 7.Case study: Cutting response times for everyday operations
- 8.Case study: Finding purchase intent in 40,000 comments
Bring one workflow. See how it scores.
Pick a recurring task. We'll define what good looks like and compare models on your own examples.