Specialists take bounded implementation and checks.
SGT Fleet
Frontier usage benchmark
80% fewer frontier tokens when orchestrated
SGT's first-party internal Gap Closer measurement: 80% fewer frontier tokens when the frontier model is forced into orchestration mode. Results vary; raw counts and method are not yet public.
80%Gap Closer reduction, orchestrator mode
Plans and reviews.
Gap Closer result + optional JEV
100%baseline usage
20%remainder after the 80% Gap Closer result
10%conditional JEV case
Optional JEV is a separate SGT-reported result: 50% less frontier orchestration cost in one repeatable, Frontier-templated routine workflow, not 50% of all tokens. 100 → 20 → 10 is conditional arithmetic, valid only if that 50% also applies to the eligible 20% remainder on a comparable basis. No end-to-end combined run has been measured.
Results vary by task; token savings do not guarantee bill savings.
All measured results and their limitsHow SGT Fleet works with Claude Code and CodexBenchmark definitions and claim boundariesAgent-readable SGT Fleet summary
Optional routine orchestration
Frontier creates the template. JEV routes the routine.
Claude Code or Codex creates the reusable guidance. Frontier hands off matching work, steps in for new situations, and reviews the result while JEV keeps the routine moving on repeatable, Frontier-templated work.
JEV by TypeSafe AI. Official docs.Built for repeatable work. Adapt a proven website, vary the details, and bring Frontier into each new situation.
The 50% is an SGT-reported internal result and has not been independently audited. Frontier still creates the templates and reviews exceptions, and no bill reduction is guaranteed. See what is measured.
Measured results
Usage benchmarks and visual comparison
German village comparison
Fable 5.1 orchestrated both builds from the same brief. The sub-agent differed: SGT Fleet on Flash 4.1 versus Claude Sonnet 5.
Claude Sonnet 5
Pricing
Simple on purpose.
Start free as a verified student or educator. Personal and enterprise plans add hosted Skills and OOP retrieval, updates, and team-ready access.
STUDENTS + EDUCATORS
Free
For enrolled students and teaching staff.
- Full Fleet software
- Both launch CLIs
- Updates included
PERSONAL
ONE PERSON
$5 / month
For individual builders using their own model subscriptions.
- Full Fleet software
- Hosted Skills + OOP retrieval
- Updates included
ENTERPRISE
$10 / seat / month
For organizations that need per-seat licensing.
- Per-seat licence
- Private-server deployment options
- Invoicing
Launch scope: maintained OOP and Skills currently target DeepSeek Flash v4.1. Codex and Fable are conductor integrations. Customers can create and maintain additional model knowledge. Provider services use the customer’s own accounts. SGT’s only hosted service is OOP/Skill querying from our server in Germany, used by the SGT tools inside Claude Code or Codex; no AI model, Claude, OpenAI or other API access is included.
For AI agents
Built to be driven by Claude Code and Codex.
Open Claude Code or Codex in your project and it drives the SGT agents itself through the sgt_* tools: it plans, sends one complete task to the agent that owns the craft, checks the files the agent wrote, and judges the result. An agent can do the whole install for you from the install guide.
Notes
Practical notes.
What is actually tuned?＋
SGT builds and validates procedural Skills and OOP role packages—the instructions, routing, and review structure around frontier orchestration. It does not change provider model weights.
Can education and paid users build domain expertise?＋
Yes. Verified education access and paid plans include the local tools to build and validate OOP/Skills for the domain, workflow, and use case you need.
Who owns locally created intelligence?＋
You do. Customer-created OOP/Skills stay in your local Fleet workspace. They are yours to keep, back up, move, and reuse across future work.
What does the hosted service retain?＋
Hosted OOP/Skills retrieval processes each query in memory. SGT does not store query text or customer-created OOP/Skills, and the query endpoint does not write access logs. We retain the account, academic eligibility, access and security, and billing records needed to provide the service. Model providers handle data sent through their accounts under their own terms.