我们完成了1.13 亿美元 B 轮融资，现在我们每月路由的 token 量达到 100 万亿。以下是 5 月上线的其他所有内容。
工作区护栏
为通过你的工作区路由的每一个请求提供集中式安全与治理。设置按成员和按密钥的支出限额，将流量锁定到模型和提供商允许列表，强制执行零数据保留，拦截针对 30 多种源自 OWASP 模式的提示词注入，并在 PII 到达提供商之前将其脱敏。将规则叠加为一道护栏，或将其限定到特定的 API 密钥和成员，无需更改代码。
语音与转录 API
通过你已在使用的同一个 API 密钥，为任何应用添加语音能力。语音转文本已上线，支持 Whisper、GPT-4o Mini Transcribe 和 Voxtral；文本转语音在模型 API 中提供 supported_voices。两者都内置了提供商故障转移和上游错误透传。
模型融合
将你的提示词并行路由到多个模型，并把它们的响应综合成一个质量更高的单一答案。模型融合现已作为 API 插件、服务器工具以及聊天室编辑器中的功能提供。你可以在一次调用中获得一个专家集成，而无需依赖单个模型。
模型对比
最多可并排对比五个模型的价格、上下文长度和基准测试得分。重建后的对比页面包含“Highlight best”开关、按提供商着色的 Intelligence、Coding 和 Agentic 指标基准图表，以及可快速添加模型的交互式槽位卡片。
私有模型（企业版）
通过标准的 completions 和 responses API，将请求路由到你自己的自定义、微调或专用模型端点。你的私有模型享有与平台上任何公共模型相同的护栏、可观测性和计费。仅在企业版方案中提供。
Pareto Code Router
设置 min_coding_score，即可路由到满足你质量门槛的最便宜的具备代码能力的模型。你的编程智能体不再为“够用就好”的代码多付钱。可在插件设置中按工作区配置默认值。
企业版与工作区控制
面向大规模运行 OpenRouter 的团队的一系列发布：
- IP 允许列表强制执行。带有 IP 允许列表的 API 密钥现在会以
403主动阻止来自未授权 IP 的请求，从仅观察模式升级而来。文档 - BYOK 管理 API。 以编程方式跨工作区列出、创建、更新和删除自带密钥（BYOK）凭据。密钥现在按优先级分组，支持拖拽重新排序，并针对失败请求提供一键“测试密钥”。API 文档
- 可观测性目标 API。 通过管理密钥管理 Datadog、Langfuse、LangSmith 及其他可观测性集成的 CRUD 端点。API 文档
- 按提供商划分的 ZDR 控制。 针对非前沿、Anthropic、OpenAI 和 Google 提供商分别提供零数据保留开关，让你可以按提供商满足合规要求，而无需限制整个模型目录。
- 跨工作区复制护栏。 通过“复制到……”菜单，只需点击几下即可在所有工作区统一安全策略。
本月还发布了
预设 API。 直接从推理请求体创建或对预设进行版本管理，现在支持 Anthropic Messages 和 Responses 外观，以及 TypeScript 和 Python SDK 支持。文档
人在回路工具。 一种新的 SDK 工具类型，会暂停执行并等待人工输入后再返回结果，适用于需要在任务中途进行人工判断的智能体。博客文章
会话 ID 提供商粘性。 共享同一 session_id 的请求现在会路由到同一提供商，并在多轮对话中固定到同一具体模型，从而提升多轮智能体工作流的缓存命中率。文档
自动路由器 cost_quality_tradeoff。 用 0 到 10 的整数取代旧的二元开关，在使用自动路由器时对成本与质量进行更精细的控制。文档
重新设计的模型页面。 全新的模型页面页眉、带有 /responses 和 /messages 端点的分步 API 标签页、全屏模型选择器，以及用于内联测试的 playground 侧边面板。
日志中的请求标签页。 在生成日志之外提供完整的请求级下钻，支持请求 ID 过滤和时间选择器简写（15min、1h、3d）。日志
改进的编码智能体归因。 Cursor、GitHub Copilot、Cline、RooCode、Kilo Code、Zed 和 OpenCode 现在都能在活动日志中被正确识别，让你可以看到是哪些工具驱动了你的用量。
API 密钥的用量与预算。 按护栏层级展示支出图表和预算进度，直接显示在每个 API 密钥上。
每日排行榜数据集。GET /api/v1/datasets/rankings-daily 按每日 token 用量返回前 50 名模型，供程序化分析使用。
新模型
5 月共发布 20 个模型，涵盖文本、语音、图像、视频和编程：
- Anthropic Claude Opus 4.8：Anthropic 最新的 Opus，支持会话中途系统提示，另有一个快速变体
- Google Gemini 3.5 Flash：Google 最新的 Flash 模型
- xAI Grok 4.3：xAI 最新的前沿模型
- xAI Grok Imagine Video：xAI 的视频生成模型
- xAI Grok Build 0.1：xAI 的代码生成模型
- Qwen Qwen3.7 Max：Qwen 最新的 max 档模型
- Recraft V3、V4、V4 Pro：三个新的图像生成模型
- Mistral Voxtral Mini Transcribe：Mistral 的语音转文本模型
此外还有：Gemini 3.1 Flash Lite、GPT Chat Latest、CoBuddy（免费）、Ring-2.6-1T（免费）、Perceptron Mk1 等。
以上所有内容现已上线。浏览完整模型目录，或在 Discord 上告诉我们还缺什么。
We closed our $113M Series B, and we’re now routing 100 trillion tokens a month. Here’s everything else that shipped in May.
Workspace Guardrails
Centralized security and governance for every request routed through your workspace. Set per-member and per-key spend limits, lock traffic to a model and provider allowlist, enforce zero data retention, block prompt injection against 30+ OWASP-derived patterns, and redact PII before it reaches a provider. Layer the rules into one guardrail, or scope them to specific API keys and members, with no code changes.
Speech and Transcription APIs
Add voice to any application through the same API key you already use. Speech-to-text is live with Whisper, GPT-4o Mini Transcribe, and Voxtral; text-to-speech exposes supported_voices in the models API. Provider failover and upstream error passthrough are built into both.
Model Fusion
Route your prompt to multiple models in parallel and synthesize their responses into a single, higher-quality answer. Model Fusion is now available as an API plugin, a server tool, and in the chatroom composer. You get an ensemble of experts in a single call instead of relying on one model.
Model Comparison
Compare up to five models side by side on pricing, context length, and benchmark scores. The rebuilt comparison page includes a “Highlight best” toggle, provider-coded benchmark charts for Intelligence, Coding, and Agentic metrics, and interactive slot cards to quickly add models.
Private Models (Enterprise)
Route to your own custom, fine-tuned, or dedicated model endpoints through the standard completions and responses API. Your private models get the same guardrails, observability, and billing as any public model on the platform. Available exclusively on the Enterprise plan.
Pareto Code Router
Set min_coding_score and route to the cheapest code-capable model that clears your quality bar. Your coding agents stop overpaying for good-enough code. Configurable defaults per workspace in plugin settings.
Enterprise & Workspace Controls
A set of releases for teams running OpenRouter at scale:
- IP allowlist enforcement. API keys with an IP allowlist now actively block requests from unauthorized IPs with a
403, upgraded from observe-only mode. Docs - BYOK management API. Programmatically list, create, update, and delete bring-your-own-key credentials across workspaces. Keys are now grouped by priority with drag-and-drop reordering and a one-click “Test Key” for failed requests. API docs
- Observability destinations API. CRUD endpoints for managing Datadog, Langfuse, LangSmith, and other observability integrations via management key. API docs
- Per-provider ZDR controls. Separate Zero Data Retention toggles for non-frontier, Anthropic, OpenAI, and Google providers, so you can meet compliance requirements per provider without restricting your entire model catalog.
- Copy guardrails across workspaces. Standardize safety policies across all workspaces in a few clicks via the “Copy to…” menu.
Also shipped this month
Presets API. Create or version a preset directly from an inference request body, now with Anthropic Messages and Responses skins, plus TypeScript and Python SDK support. Docs
Human-in-the-loop tools. A new SDK tool type that pauses execution and waits for human input before returning results, for agents that need human judgment mid-task. Blog post
Session-id provider stickiness. Requests sharing a session_id now route to the same provider and pin to the same concrete model across turns, improving cache hit rates for multi-turn agentic workflows. Docs
Auto router cost_quality_tradeoff. A 0 to 10 integer replacing the old binary toggle for finer control over cost versus quality when using the auto router. Docs
Redesigned model pages. New model page header, step-by-step API tab with /responses and /messages endpoints, full-screen model selector, and playground side panel for inline testing.
Requests tab in logs. Full request-level drill-down alongside generation logs, with request ID filtering and time picker shorthand (15min, 1h, 3d). Logs
Improved coding agent attribution. Cursor, GitHub Copilot, Cline, RooCode, Kilo Code, Zed, and OpenCode are now properly identified in activity logs so you can see which tools drive your usage.
Usage & Budgets on API keys. Spend charts and budget progress by guardrail layer, directly on each API key.
Rankings daily dataset.GET /api/v1/datasets/rankings-daily returns top-50 models by daily token volume for programmatic analysis.
New models
20 models launched in May, spanning text, speech, image, video, and coding:
- Anthropic Claude Opus 4.8: Anthropic’s latest Opus with mid-session system support, plus a fast variant
- Google Gemini 3.5 Flash: Google’s newest Flash model
- xAI Grok 4.3: xAI’s latest frontier model
- xAI Grok Imagine Video: Video generation from xAI
- xAI Grok Build 0.1: xAI’s code generation model
- Qwen Qwen3.7 Max: Qwen’s latest max-tier model
- Recraft V3, V4, V4 Pro: Three new image generation models
- Mistral Voxtral Mini Transcribe: Mistral’s speech-to-text model
Plus: Gemini 3.1 Flash Lite, GPT Chat Latest, CoBuddy (free), Ring-2.6-1T (free), Perceptron Mk1, and more.
Everything above is live now. Browse the full model catalog, or tell us what’s missing on Discord.