May 2026 Summaries

OpenRouter workspaces offer configurable guardrails to enhance security, governance, and cost management by incorporating tools for budget enforcement, zero data retention, model and provider restrictions, prompt injection defense, and data loss prevention. These guardrails can be applied broadly across an entire workspace or customized for specific team groups or API keys without altering the code. Budget enforcement allows setting spending limits for specific timeframes, while zero data retention and model/provider restrictions provide control over data handling and model usage. Prompt injection defense employs regex patterns to detect and mitigate injection attempts, and data loss prevention identifies and manages sensitive information using pre-defined and custom patterns. These guardrails can be assigned to API keys or individual members, ensuring tailored security measures, and the management API allows for programmatic configuration, enabling automation during onboarding or key rotation processes.

OpenRouter workspaces offer a suite of configurable security and governance tools known as guardrails, which aid in budget enforcement, zero data retention (ZDR), model and provider restrictions, prompt injection defense, and data loss prevention (DLP). These guardrails can be layered to govern entire workspaces or customized for specific team members or API keys without altering existing code. Budget enforcement allows for setting financial limits to prevent overspending, while ZDR and model/provider restrictions help manage data retention and access to vetted services. Prompt injection defense utilizes regex patterns to detect and handle attempts at unauthorized access or manipulation, offering actions such as flagging, redacting, or blocking requests. DLP features detect and manage sensitive information, using built-in or custom regex patterns, and offer actions like redaction or blocking to maintain compliance with data handling protocols. Guardrails can be assigned across various API keys or organizational members, with the ability to customize and automate their application using the management API, providing a flexible and comprehensive approach to security and cost management.

OpenRouter has secured a $113 million Series B funding round led by CapitalG, with participation from several prominent investors including NVentures, ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, Databricks Ventures, AMP PBC, and Pace Capital, alongside existing investors Andreessen Horowitz and Menlo Ventures. This funding comes as OpenRouter experiences significant growth, with weekly token volumes increasing from 5 trillion to 25 trillion and anticipated to exceed a quadrillion tokens annually, serving over 8 million developers across more than 400 models. The investor composition underscores the strategic importance of OpenRouter's role as a critical infrastructure layer for multi-model production systems, enabling reliable, scalable, and cost-efficient AI deployment. OpenRouter has expanded its capabilities to support multimodal inference, enterprise controls, and intelligent routing, positioning itself as a vital component in the evolving AI landscape. The funds will be used to further scale OpenRouter's infrastructure, enhance enterprise features, and invest in intelligent routing to better serve the needs of its growing developer community.

OpenRouter has secured a $113 million Series B funding round led by CapitalG, with participation from prominent investors like NVentures, ServiceNow Ventures, and others, aiming to enhance its AI infrastructure capabilities. The company has experienced significant growth, processing up to 25 trillion tokens weekly and serving over 8 million developers across 400 models, reflecting a shift from AI experimentation to critical production applications. The strategic involvement of infrastructure and platform companies as investors underscores the demand for a reliable routing and gateway layer in multi-model production systems, positioning OpenRouter as a crucial component. Over the past year, OpenRouter has expanded its capabilities to support various models beyond text, including image, audio, and video, while enhancing enterprise controls and intelligent routing. The new funding will be used to further scale infrastructure, improve enterprise features, and invest in intelligent routing to help teams find optimal models and providers for diverse AI requests.

In a unique experiment involving artificial intelligence models, Jacky Liang conducted a 30-game battle royale simulation with eleven large language models (LLMs) to explore how they perform in competitive scenarios. The experiment revealed that Grok 4.1 Fast, a cost-effective model from xAI, won 43% of the matches by adopting an aggressive strategy, in contrast to Anthropic's Claude Sonnet 4.6, which focused on collaboration and communication, reflecting its training on polite and cooperative behavior. This divergence in performance highlighted the influence of "alignment tax," where models designed for helpfulness may underperform in zero-sum games due to their cooperative nature. The study also found that traditional benchmarks do not fully capture the nuances of model performance in specific tasks, prompting questions about the alignment of AI models for different real-world applications. Liang suggests developing a router that selects the optimal model for specific tasks, emphasizing the importance of considering model alignment beyond typical benchmarks for diverse applications.

The Agent SDK has introduced a fourth tool type called human-in-the-loop (HITL) tools, which allow agents to automatically handle routine calls and pause for human intervention when necessary, controlled by a single hook. These tools can auto-resolve or escalate calls based on input, with the decision logic centralized in the onToolCalled function. Depending on the input, HITL tools either continue the process automatically or pause for human review, with the application resuming the loop once a human decision is made. An optional onResponseReceived hook can transform and process human input before it is passed to the model. This approach is useful for scenarios where decisions depend on data, such as dollar thresholds or compliance checks, streamlining the integration of human judgment into automated processes without scattering logic across application code. The SDK manages state tracking and hook dispatch, requiring minimal code from developers, and provides a cookbook recipe for implementing these tools along with documentation and support through an API key and community feedback on Discord.

Human-in-the-loop (HITL) tools are now supported by the Agent SDK, providing a mechanism for automated agents to handle routine tasks while pausing for human intervention in high-stakes situations. This new tool type is controlled by a single hook, allowing for flexibility in deciding when to auto-resolve tasks or escalate them for human review based on predefined criteria like dollar thresholds or risk scores. The SDK facilitates seamless integration by managing state tracking, hook dispatch, and schema validation without requiring additional loop code from developers. The HITL framework's pause and resume cycle involves the model calling a HITL tool, followed by a decision-making process that can either allow the agent to continue automatically or pause for human input. Once a human decision is made, it is processed and the agent resumes its operations, with optional post-processing to enhance or validate the human response. The distinction between HITL and requireApproval tools is that HITL uses data-driven logic for decision-making, whereas requireApproval mandates explicit human consent for each operation.

David Bai discusses the introduction of two new tools, openrouter:web_search and openrouter:web_fetch, which enable consistent web search and content retrieval across different models without requiring client-side implementation. These tools standardize the way models handle web search and content fetching, allowing for consistent behavior regardless of the model or provider being used, such as GPT-5.5, Claude, or Kimi. The web search tool supports various engines, each with different pricing and capabilities, such as domain filtering and configurable result contexts, while the web fetch tool allows for full-page content retrieval with options to control content size and domain access. These innovations aim to replace the previous reliance on web search plugins, which limited the model's ability to define search parameters and frequency, although support for these tools requires models capable of tool-calling. The tools provide a unified schema for invoking and parsing search and fetch results, ensuring predictable and consistent outcomes across different AI models.

OpenRouter introduces two server-side tools, openrouter:web_search and openrouter:web_fetch, that enable consistent and flexible web search and content fetching capabilities across various models without client-side implementation. These tools allow models to perform agentic searches and fetch full page content from URLs, offering a uniform experience regardless of the model or provider being used. Web Search supports multiple engines, including native, Exa, and Parallel, each with distinct features like configurable result context size and domain filtering, ensuring reliable search behavior across different models. Meanwhile, Web Fetch provides similar consistency with engines like Exa and Parallel, enabling content extraction and domain-specific fetching while maintaining a predictable cost and context usage. This approach marks a shift from the previous web search plugin, granting models greater autonomy in search operations and requiring a model that supports tool calling for implementation.

The analysis of the cost impact of the GPT-5.5 model, which launched with a 2x price increase over GPT-5.4, reveals that while input and output token costs doubled, the model's less verbose nature for longer prompts somewhat mitigates the price rise. The study, focusing on users who switched from GPT-5.4 to GPT-5.5, found cost increases ranging from 49% to 92%, with the model generating 19-34% fewer completion tokens for prompts exceeding 10K tokens. However, for shorter prompts under 10K tokens, the model produced longer completions, leading to a higher cost increase. The methodology involved a direct comparison of the same user base's costs across model versions, using OpenRouter's token counts for normalization and excluding non-text and cancelled requests.

GPT-5.5, which launched with a 2x price increase from its predecessor GPT-5.4, also offers reduced verbosity for longer prompts, producing 19-34% fewer completion tokens for such prompts. The cost analysis, conducted using the same switcher cohort approach as previous studies, revealed that users experienced a cost increase between 49-92% depending on prompt length, despite the model's efficiency improvements for longer inputs. For prompts under 10,000 tokens, costs increased without the benefit of shorter completions, whereas for longer prompts, the reduction in completion length helped offset the price hike. The analysis used OpenRouter's token counts to provide a consistent baseline across model versions, excluding non-text requests and ensuring a controlled comparison within the same user cohort.

OpenRouter has introduced two dedicated audio endpoints for enhanced speech and transcription capabilities: /api/v1/audio/speech for text-to-speech and /api/v1/audio/transcriptions for speech-to-text. These endpoints offer specialized models that are faster and more cost-efficient for specific audio tasks compared to general audio models. Users can generate speech using voices from OpenAI, Google, or Mistral and transcribe audio files with OpenAI Whisper, maintaining consistent routing, billing, and key management across different media types. The choice between audio, speech, and transcription models involves a trade-off between specialization, cost, and speed, with each model being optimized for particular use cases such as voice agents, reading text aloud, or transcribing meeting notes. The platform provides tools like a Playground for experimenting with models and quickstart code samples for integration, supporting a variety of audio formats and allowing for customization such as tone control. OpenRouter plans to expand its offerings with more providers and voices, inviting user feedback for future developments.

OpenRouter introduces two specialized audio endpoints, /api/v1/audio/speech for text-to-speech and /api/v1/audio/transcriptions for speech-to-text, which deliver faster and more cost-efficient models tailored for specific audio tasks. These endpoints support generating speech from text using voices from providers like OpenAI, Google, and Mistral, as well as transcribing audio files with OpenAI Whisper, while maintaining consistent routing, billing, and key management across text, video, and image generation services. The choice between audio, speech, and transcription models depends on the desired balance of specialization, cost, and speed, with each model offering unique capabilities suitable for different use cases such as voice agents, reading text aloud, or creating meeting notes. Users can explore model functionalities in the Playground, selecting voices for speech models or uploading audio files for transcription, and each model page provides quickstart code examples in programming languages like Python and TypeScript. OpenRouter is continuously expanding its offerings with more providers and voices, and users are encouraged to suggest additional models through their Discord community.