Poolside
Listed at https://poolside.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Laguna S 2.1: $0.09/M input; Laguna XS 2.1: $0.06/M input; cached input $0.009/M 与 $0.003/M（分别为 10× 缓存折扣） | per million tokens |
| Output | Laguna S 2.1: $0.18/M output; Laguna XS 2.1: $0.12/M output | per million tokens |
🤖 Supported Models (4)
✨ Pros
- ✓Poolside is an AI coding-focused lab; the Laguna family spans from XS 2.1 (33B/3B, runs on a 36GB-RAM Mac) to S 2.1 (118B/8B, datacenter-class) for agentic coding workloads
- ✓Laguna S 2.1 supports 1M-token context and native reasoning (thinking off / max), tuned for long-horizon agent tasks with multi-step verification loops
- ✓On Poolside's reference benchmarks: Laguna M.1 SWE-bench Verified 65.4%, Terminal-Bench 2.0 32.7%; XS 2.1 improves over XS.2 by +5.4% on SWE-bench Multilingual and +1.8% on Terminal-Bench 2.0
- ✓OpenAI-compatible API; weights released under OpenMDW-1.1 (BF16 + FP8 / INT4 / NVFP4 / GGUF / MLX quantizations); local inference supported on vLLM, SGLang, Ollama, llama.cpp, MLX, and NVIDIA TensorRT-LLM
- ✓OpenRouter hosts Laguna S 2.1 at $0.09/M input, $0.18/M output, $0.009/M cached input; Laguna XS 2.1 at $0.06/M input, $0.12/M output — below Anthropic Sonnet 5 and GPT-6 main tiers
- ✓No first-party marketing pricing page; verified via OpenRouter rates + docs/release-notes/models.md; a single NVIDIA DGX Spark is sufficient to host Laguna S 2.1 locally
⚠️ Cons
- ×No mainland China endpoint — proxy required; the official console and chat.poolside.ai do not serve the region directly
- ×Only two reasoning settings (off / max); no fine-grained xhigh / high / medium / low / minimal / none effort controls, which hurts cost-sensitive reasoning workflows
- ×OpenMDW-1.1 is a custom license (not Apache 2.0 / MIT); commercial boundaries need legal review
- ×Models are recent (M.1 Apr 2026, XS 2.1 Jun 2026, S 2.1 Jul 2026); IDE / Copilot-replacement ecosystem is less mature than Anthropic or OpenAI
- ×No first-party marketing pricing page; rates live on OpenRouter and docs/release-notes/models.md, so budget evaluation requires manual cross-checks
🎯 Best For
Long-context (1M) + native-reasoning + open-weight agentic coding workloads; single-machine local deployment (36GB-RAM Mac or DGX Spark); teams already routing through OpenRouter add it with zero friction
💰 Pricing & Plans
| Model / Access path | Input ($/1M tokens) | Output ($/1M tokens) | Notes |
|---|---|---|---|
| Laguna S 2.1 (OpenRouter) | $0.09 | $0.18 | 1M context, 118B/8B MoE; 10% off the $0.10/$0.20 list price |
| Laguna S 2.1 - cached input | $0.009 | - | 10x cheaper than standard input; 10% off the $0.01 cache-read list |
| Laguna XS 2.1 (OpenRouter) | $0.06 | $0.12 | 256K context, 33B/3B MoE; 40% off the $0.10/$0.20 list price |
| Laguna XS 2.1 - cached input | $0.003 | - | 40% off the $0.05 cache-read list price |
| Laguna S 2.1 :free / XS 2.1 :free | $0 | $0 | Rate-limited free variants on OpenRouter; 256K context, 32K max completion |
| Vercel AI Gateway | $0.09 | $0.18 | poolside/laguna-s-2.1 at the same rates, billed through Vercel |
| Poolside-hosted inference | Not published | Not published | No first-party price list; keys from platform.poolside.ai, contract pricing |
| Self-managed Poolside Inference | Quote | Quote | Helm deployment on AWS EKS, OpenShift, upstream Kubernetes or on-prem GPUs |
🔧 API & Developer Experience
- •OpenAI-compatible Chat Completions API: Base URL https://inference.poolside.ai/v1 exposes GET /v1/models and POST /v1/chat/completions, so any OpenAI SDK works by swapping base_url; OpenRouter serves the same models at https://openrouter.ai/api/v1
- •Native reasoning, on by default: Responses carry reasoning_content alongside content; turn it off per request with chat_template_kwargs.enable_thinking=false. Poolside ships only two settings (off and max) - there is no xhigh/high/medium/low effort ladder
- •Tool calling with parallel execution: The tools array is OpenAI-style and parallel_tool_calls defaults to true; set it to false to force one call per response. In agentic loops, keep reasoning_content from earlier turns in the request history or later steps lose their reasoning
- •Generation and streaming controls: max_tokens defaults to 32768 and accepts up to 262144, temperature runs 0-2, plus top_k and min_p. Streaming uses SSE with running usage totals on every chunk and the final total on the closing chunk; logprobs and stop_reason are not returned
- •Poolside Agent CLI (pool) as the native harness: The models were trained with reinforcement learning inside this agent, so they perform best in it: pool for interactive sessions, pool exec for one-shot CI tasks, with plan mode, skills, subagents, hooks, MCP servers and turn rewind
- •Editor, agent and CI integrations: ACP editors (JetBrains, Zed, Neovim/CodeCompanion), VS Code and Visual Studio extensions (Poolside Assistant, Cline, Kilo Code, GitHub Copilot), desktop apps (Goose), automation (GitHub Actions, n8n) and MCP over stdio, Streamable HTTP or SSE
- •Self-hosted inference and gateways: Deploy the Helm stack on AWS EKS, OpenShift, upstream Kubernetes or on-prem single-node GPU hosts, then front it with Bifrost or LiteLLM for virtual keys, budgets, rate limits and request logging; vLLM Prometheus metrics are exposed on /metrics
🧠 RL-Trained In-House, Shipped Open-Weight
Poolside trains every Laguna model from scratch on its own data and infrastructure, then post-trains it with reinforcement learning inside its own agent harness - the same pool loop you can run yourself. Laguna S 2.1 was trained on 30T tokens and reaches 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE with 118B total and only 8B active parameters, competitive with models an order of magnitude larger. What happens after training is the unusual part: the weights ship open, so you can pull BF16, FP8, INT4, NVFP4, GGUF or MLX checkpoints from Hugging Face and serve them with vLLM, SGLang, Ollama, llama.cpp, MLX or TensorRT-LLM, or install the full inference stack on your own GPUs. A 36GB-RAM Mac runs XS 2.1; a single NVIDIA DGX Spark runs S 2.1.
🌐 China Access & Latency
Poolside serves inference from a single US region: inference.poolside.ai resolves to 35.244.129.243, a Google LLC address in Kansas City, Missouri, and answers with a via: 1.1 google header. Measured from an Alibaba Cloud Malaysia host on 30 October 2026, an unauthenticated /v1/models call returned TTFB in 254-279ms (median 263ms) - fast for a US origin, but every mainland request still crosses the Pacific. Poolside publishes no China endpoint, region or availability commitment anywhere in its docs, so treat direct calls from the Chinese mainland as proxy-dependent and expect roughly 400-900ms per request behind a proxy: fine for background agent runs, visible for interactive completions. The console is friendlier - platform.poolside.ai sits behind CloudFront with a Sepang, Malaysia edge at 237-675ms. Three workarounds matter for China-facing teams. Route through OpenRouter or the Vercel AI Gateway, both Cloudflare-fronted with wide domestic reach, at identical model rates. Better still, self-host: the Helm-based inference stack deploys to your own Kubernetes or on-prem GPUs in Hong Kong, Singapore or Shanghai, served over the same OpenAI-compatible /v1 base URL, with code never leaving your infrastructure. Or download the open weights and run fully offline via vLLM, SGLang, Ollama or MLX.