Sail Research

Sail Research sells inference and compute infrastructure tuned for long-horizon AI agents rather than chat latency. It pairs an OpenAI- and Anthropic-compatible API over open-weight models with Sailboxes, persistent Linux VMs that auto-sleep while an agent waits. It targets engineering teams whose agent token and sandbox bills have outgrown general-purpose GPU clouds.

Sail Research operates an inference and compute platform built specifically for long-horizon AI agents, the workloads that run for hours or days and consume billions of tokens on a single task. The company emerged from stealth on 25 June 2026 with $80 million in combined Seed and Series A funding at a $450 million valuation, backed by Kleiner Perkins, Sequoia Capital, Redpoint, Theory Ventures, Vine Ventures, CRV, A* and Abstract Ventures, plus angels including Alphabet chair John Hennessy, Intel CEO Lip-Bu Tan and Together AI chief scientist Tri Dao. The platform has two halves. The first is an inference stack built on customised open-source engines including vLLM, serving open-weight models such as GLM-5.3, DeepSeek V4 Flash and Pro, Kimi K3 and K2.6, gpt-oss-120b, Gemma 4 and Qwen through OpenAI- and Anthropic-compatible APIs, so migrating requires only a new base URL and key. Its distinguishing feature is completion windows: the caller declares how much latency it will tolerate, Default, Balanced or Flex, and Sail schedules the work across its fleet accordingly, advertising roughly 5-35%, 45-65% and 60-80% savings respectively. The second half is Sailboxes, generally available since 14 July 2026: full Linux VMs with persistent disk, local NVMe, Docker support and no runtime limits, which live-migrate across Sail's fleet based on live resource usage and auto-sleep while an agent waits on inference, so customers pay for work rather than for a reservation. Sail reports 90.72% accuracy on the BrowseComp-Plus research benchmark at up to a tenth of rivals' inference cost, and names Parallel, Detail.dev, Jack and Jill, and Quadrillion Labs as production customers.

Category

Infrastructure & Cloud

Pricing

Usage-based, Contact for pricing

Target Market

CTOs, Enterprise Developers, Platform Engineers, AI/ML Engineers

Deployment

API-based, Cloud-only

Headquarters

San Francisco, United States

• ✓
  Completion windows

  Callers declare latency tolerance per request — Default, Balanced or Flex — and Sail schedules against fleet capacity, cutting token cost by roughly 5-35%, 45-65% or 60-80%.

• ✓
  Sailboxes

  Full Linux VMs with persistent state, independent disks, local NVMe and Docker support that run for hours or days with no runtime limit.

• ✓
  Auto-sleep and live migration

  Sailboxes suspend while an agent waits on inference and migrate across the compute fleet by live resource usage, so idle waiting is not billed as reserved capacity.

• ✓
  Drop-in compatible APIs

  OpenAI Chat Completions, OpenAI Responses and Anthropic Messages endpoints mean migration is a base-URL and API-key change rather than a rewrite.

• ✓
  Open-weight model catalogue

  Serves GLM-5.3 and 5.3-Flash, DeepSeek V4 Flash and Pro, Kimi K3 and K2.6, gpt-oss-120b, Gemma 4 and Qwen with per-model input, cached and output rates.

• ✓
  Fine-tuning and RL support

  Deploys LoRA fine-tunes and supports reinforcement-learning rollout sampling, including a Tinker integration, so custom adapters run on the same stack.

• ✓
  Zero data retention by default

  ZDR is the default posture, with HIPAA and SOC 2 compliance and enterprise BAA, MSA and DPA terms plus regional data processing available.

• •
  Cost-optimising a deep-research agent

  Route a multi-hour browsing and synthesis agent through Flex windows to cut token spend by up to 80% where results are consumed asynchronously.

• •
  Hosting autonomous coding agents

  Give each agent a Sailbox with a real filesystem and Docker so it can build, test and iterate across days without losing working state.

• •
  Parallel experiment fleets

  Fork Sailboxes to run many agent variants concurrently, as Quadrillion Labs does for the Qualia Cloud autonomous research platform.

• •
  Migrating off a general-purpose inference provider

  Swap the base URL on an existing OpenAI- or Anthropic-shaped client and compare cost per token on identical traffic before committing.

• •
  Serving custom LoRA adapters at agent scale

  Deploy domain-tuned adapters and sample RL rollouts on the same inference fleet instead of standing up separate GPU infrastructure.

Metered per million tokens with separate input, cached-input and output rates that vary widely by model — DeepSeek V4 Flash lists at $0.09 input and $0.18 output while Kimi K3 lists at $3.00 and $15.00. The real lever is the completion window: Balanced saves roughly 45-65% and Flex 60-80% against the default ASAP rate, but not every model supports every window. Sailbox compute is billed separately on usage with auto-sleep, and $5 in free credits arrives monthly. Volume discounts, HIPAA BAAs, regional processing and SLAs are Enterprise-only and require a sales conversation.

This page was written from 8 sources, 7 on domains other than sailresearch.com.

1. 1.sailresearch.com — sailresearch.comvendor
2. 2.docs.sailresearch.com — pricing
3. 3.docs.sailresearch.com — docs.sailresearch.com
4. 4.prnewswire.com — sail research raises 80 million to build max efficiency infr
5. 5.prnewswire.com — sail research launches sailboxes the first cloud environment
6. 6.siliconangle.com — sail research raises 80m optimize long horizon ai agents
7. 7.thenextweb.com — sail research 80m ai agent inference
8. 8.hn.algolia.com — hn.algolia.com