Table of Contents
🌏 中文版
Sail Research is an AI infrastructure company founded by Neil Movva and Samir Menon. Its formal products cover open-model inference APIs, persistent Linux VMs called Sailboxes, and agent telemetry called Voyages. This is not the Stanford AI Lab or one of several papers named SAIL; the founders, investors, documentation, and sailresearch.com and sail.industries domains all identify the same company.
Sail belongs in “agent-oriented inference” because it reverses the usual serving objective, not because it puts an agent label on an API. Interactive chat prioritizes time to first token. Research, code review, evals, and RL rollouts running for hours often care more about tokens completed per dollar. Sail lets requests declare how long they can wait, allowing its scheduler to exchange latency flexibility for higher fleet utilization and lower prices.
In June 2026, the company announced $80 million across its seed and Series A rounds, led by Sequoia Capital and Kleiner Perkins respectively. Sail also says it has processed trillions of tokens for customers including Parallel, Detail.dev, Jack & Jill, and Quadrillion. This is company-reported adoption, not an independent audit.
Core architecture: let the workload state its patience
Traditional serverless inference usually chases low latency and high throughput at once. Sail's key abstraction is the completion window: the same model and API can assign different scheduling urgency per request.
| Window | Scheduling goal | Best fit |
|---|---|---|
asap | Prioritize low latency | A person is waiting, or the next agent turn is blocked |
balanced | Give the scheduler a wider window | Background agents, subagents, and parallel pipelines |
flex | Best-effort scheduling | Evals, batch jobs, and offline enrichment |
This is not a smaller-model or lower-quality mode. It makes time a resource. Patient background work lets the platform aggregate requests, place them on heterogeneous hardware, consume spot capacity, and fail over between unreliable workers. Sail's manifesto says the team writes CUDA kernels, modifies engines such as SGLang, and runs a global controller to raise fleet utilization. The complete implementation is not public, so treat this as vendor architecture, not a reproducible systems paper.
The second layer is Sailboxes: kernel-isolated Linux VMs with root access, independent disks, Docker, and local NVMe. A Sailbox can sleep while an inference request waits, then wake when the result arrives. This connects “waiting cheaply for the model” with “retaining the agent's computer cheaply,” which is the meaningful difference between long-horizon workloads and ordinary chat endpoints.
API: compatibility is not equivalence
The smallest migration uses the OpenAI SDK with a different base URL and key:
from openai import OpenAI
client = OpenAI(
base_url="https://api.sailresearch.com/v1",
api_key="YOUR_SAIL_API_KEY",
)
response = client.responses.create(
model="zai-org/GLM-5.2-FP8",
input="Review this repository and propose three testable fixes.",
max_output_tokens=2000,
background=True,
metadata={"completion_window": "balanced"},
)
print(response.id) # Poll later, or use a webhookAs of publication, the API support matrix marks Responses, Chat Completions, and Batch as stable, with Anthropic Messages in beta. Responses supports function tools, structured output, reasoning, background mode, and foreground SSE. A background request immediately returns an ID and later completes through polling or a webhook; it cannot stream at the same time.
“OpenAI-compatible” does not mean every field is bit-for-bit equivalent. Tool calling, thinking blocks, image inputs, stop reasons, and token accounting differ across surfaces, and the Anthropic client has metadata typing differences. Before migrating, run existing production payloads as contract tests, especially tool-call round trips, JSON Schema, cancellation, retries, usage, and partial output.
Security and data boundaries
Keep API keys in a backend secrets manager. Long-running jobs also need persisted request IDs, idempotency keys, and verified webhook signatures. Validate function arguments, permissions, and timeouts in your execution layer. An inference provider returning a tool call does not authorize the action.
Data retention deserves particular attention. Sail's terms say requests and responses may be stored and that disabling storage is not currently available. The same terms say customer data is not used for model training unless the customer explicitly opts in or requests a service that requires it. They classify payment-card data, PHI, and government identifiers as prohibited content. An open-weight model does not make a hosted SaaS endpoint appropriate for regulated data.
Sailboxes are not a backup service either. The terms say a host failure may restore only the latest checkpoint, without replaying later commands. Users are responsible for backups. Long jobs should write irreplaceable artifacts to external object storage and workflow state to a replayable event log rather than relying only on VM disk.
Choosing among inference platforms
| Platform | Main optimization | Consider it first when |
|---|---|---|
| Sail | Patient open-model inference plus persistent agent VMs | Token-heavy asynchronous jobs can exchange latency for cost |
| Baseten | Hosted APIs, custom deployment, training, multi-cloud, and production operations | You need private checkpoints, custom containers, environment promotion, or single tenancy |
| Fireworks | Serverless and dedicated inference, broad models and modalities, and fine-tuning | You need mature real-time serving, model breadth, and integrated training |
| Together AI | Serverless, dedicated inference, fine-tuning, and GPU clusters | Your path extends from API calls to training or managing large clusters |
| Cerebras | Extremely low latency and high token generation speed on wafer-scale hardware | Agent turns are sequential and per-turn speed determines task completion time |
Sail does not establish a universal rule that slower inference is always cheaper. If every agent step blocks on the previous one, scheduling delays multiply by the turn count. Cerebras or another low-latency endpoint may lower total task cost because sandboxes, workers, and users spend less time waiting. Evals or deep-research systems that fan out thousands of independent requests are much better candidates for balanced and flex.
Baseten differs through model and deployment control, including custom Docker, dedicated GPUs, autoscaling, and environment promotion. Fireworks and Together expose broader surfaces spanning modalities, fine-tuning, or clusters. Sail is currently more focused on selected open models and agent economics. If a specific model, data region, zero-data-retention policy, or SLA is mandatory, verify it directly instead of comparing only token prices.
Performance claims and limitations
In its own BrowseComp-Plus experiment, Sail reports 90.72% accuracy for its research agent at 6–35 times lower cost than comparison systems. This is a vendor benchmark: Sail supplied both the inference engine and the agent harness, so the result does not isolate contributions from the model, search strategy, token budget, and scheduler. It makes abundant-token inference worth testing; it does not predict automatic gains for your agent.
Discount and latency figures on the completion-window page are generally targets or ranges relative to “traditional providers,” not an SLA. The service terms explicitly provide no general uptime commitment. Before production adoption, measure three workload-specific curves: completion-time distribution by window, total cost per successful task, and retry amplification after provider failures.
The model catalog and API maturity are also changing quickly. Passing a Responses API test today does not guarantee that the beta Messages surface reproduces every Anthropic behavior. The provider reserves the right to change beta features. Hide provider metadata behind an adapter and maintain a second routing path for critical models.
Overall
Sail's interesting idea is not another OpenAI-compatible endpoint. It makes “this request can wait” a first-class scheduling signal. That matches the actual economics of background agents better than minimizing TTFT on every call, while Sailboxes address the compute wasted during those waits.
Start with one genuinely parallel workload and run the same model through asap, balanced, and flex. Record full wall-clock time, success rate, retries, and the token bill, then compare cost per successful task—not cost per million tokens. Sail's architecture becomes your cost advantage only when the agent is truly patient.
References
Loading...