Changelog

Batch API and Gemini-compatible endpoint

Submit large jobs as a batch for 35% off list price, and call our chat models through Google’s native Gemini API with the google-genai SDK.

Batch API

• Run large jobs at 35% off. The new Batch API lets you submit many requests as a single asynchronous job and pay 35% off list price. Upload a JSONL file of requests to /v1/files, create the batch with /v1/batches, then poll for the finished results file. Each line targets /v1/chat/completions or /v1/embeddings.
• See which models are eligible. Batch-eligible models show a “Batch API” note on their model page and a batch_supported field on GET /v1/models.

Gemini-compatible endpoint

• Use our models with the Google GenAI SDK. EmpirioLabs now speaks Google’s native Gemini API. Point the google-genai SDK (Python or JavaScript), or any Gemini-native client, at https://api.empiriolabs.ai, authenticate with your EmpirioLabs API key, and call any chat model through generateContent and streamGenerateContent.
• Every parameter carries over. System instructions, temperature, max output tokens, thinking budget, tools and function calling, JSON output, and image input all translate automatically. Pricing and billing match /v1/chat/completions.
• Find it in the API reference. The new Gemini section documents generateContent, streamGenerateContent, and countTokens.

Structured output and Merge Gateway support

Constrain chat model responses to JSON or a JSON Schema with response_format, and use EmpirioLabs models through Merge Gateway.

Structured output

• Structured output with response_format. Constrain a chat model to return JSON. Send {"type":"json_object"} for any valid JSON object, or {"type":"json_schema", "json_schema": {...}} to force output that matches a schema you provide. It works on /v1/chat/completions, /v1/completions, /v1/messages, and /v1/responses (through the Responses text.format field).
• See which models support it. Every model page and GET /v1/models now shows a Structured output field: JSON Schema (schema enforced), JSON Mode (valid JSON), or Not supported.
• In the playground. Chat models that support it now have an Off / JSON mode / JSON schema control, with a box to paste your schema.

Now available on Merge Gateway

• EmpirioLabs is a supported model vendor on Merge Gateway. If your team routes LLM traffic through Merge Gateway, you can now select EmpirioLabs models there and call them alongside your other providers.

Compose video productions and playground refinements

Compose turns a topic or script into a finished short-form video with AI script, voiceover, scenes, captions, and music. Plus a cleaner playground chat view.

What’s new

• Compose: finished videos from a single brief. Generation Templates now includes Compose, which produces a complete short-form video instead of a single clip. Give it a topic, or paste your exact script, and it writes the script, records a natural voiceover, generates the scenes, burns in word-timed captions, and can add music. Planning and rendering are separate steps: planning returns an editable storyboard, so you can adjust narration, captions, or visuals per scene before the final render. Recipes cover faceless reels, talking heads, product ads, story and explainer formats, music montages, and more.
• Compose in the playground and the API. Pick a recipe in the playground under Templates, then Compose, or call POST /v1/videos/compose with mode: "plan" and then mode: "render", polling /v1/jobs/{job_id} for the result. List available recipes with GET /v1/videos/compose/recipes. Each pipeline step bills its underlying model at the normal catalog price, and the production appears as a single entry in your usage logs. See the Generation Templates guide for details.
• Playground interface refinements. The chat view now uses a single header toolbar: model info on the left, with the save-state indicator, creative templates, chat history, and Clear docked together on the right. The conversation area also spans the full page width, giving long responses and code blocks more room.

Hosted Agents improvements

A round of Hosted Agents improvements: yearly plans at 10% off, clearer subscription management with renewal dates and one-click cancel, plus reliability hardening for saved state and recovery.

We shipped a batch of Hosted Agents improvements.

Plans and subscriptions

• Yearly plans, 10% off. Hosted Agent plans can now be billed yearly at a 10% discount, alongside the monthly option. Choose monthly or yearly when you deploy an agent. Yearly pricing appears on the pricing page and in the deploy modal.
• Clearer subscriptions. The hosted agents page now lists your active subscriptions at the top, each with its plan and renewal date, so you always know what you are paying for even when no agent is currently running.
• Cancel anytime. Cancel a subscription to stop it renewing. The agent keeps running until the end of the period you already paid for, then it is removed, and you can resume any time before then. Deleting an agent removes only the agent; your subscription is separate and keeps billing until you cancel it.

Reliability

• Durable agent memory. We strengthened how an agent’s saved state is stored and restored. An interrupted save can no longer leave an agent unable to reload its history, and restoring state now retries automatically through transient storage issues.
• More reliable creation. Creating a new agent retries automatically through brief platform maintenance, so it no longer fails when it overlaps with an update.
• Automatic recovery. Agents reconcile and come back on their own after routine platform updates.

Your connected channels and tools are unaffected.

Model catalog release dates and generation updates

Model pages, API model data, generation workflows, and multi-language support now show fresher product information with clearer release history.

What’s new

• Model release dates are now visible. Model release dates now appear in the public catalog, pricing cards, model detail pages, /v1/models, /v1/models/{id}, and model docs. When an exact release date is not verified, the field stays blank or shows -.
• New models launched. Kimi K2.7 Code, GLM 5.2, and Kling 3.0 Turbo launched.
• Generation templates expanded. POV and Lip Sync templates are available, with better support for compatible video-generation workflows.
• Multi-language support expanded. Marketing, docs, and platform surfaces now have broader localization coverage across the supported language list.

GPU Cloud and Hosted Agents are now available

GPU Cloud and Hosted Agents are out of early access and available to every account, with full API coverage and in-dashboard chat for deployed models.

What’s new

• GPU Cloud is generally available. Deploy a managed GPU instance in a click, then serve any Hugging Face model behind an OpenAI-compatible endpoint, run a one-click template (JupyterLab, ComfyUI, Web Terminal, Ollama), or bring your own CUDA Docker image. Pricing is per second, the rate is locked when you launch, and you reach the workload through the authenticated connect path at api.empiriolabs.ai/v1/gpu/connect/{instance_id}/....
• Chat with your deployed model in the dashboard. Any GPU instance running an OpenAI-compatible model gets a built-in chat page, so you can test prompts, attach images or audio, and stream responses without leaving the dashboard.
• Hosted Agents is generally available. Deploy a private OpenClaw or Hermes agent that lives in your chat apps, runs code safely in an isolated sandbox, generates media across image, video, speech, and more, browses the web, and connects your tools through remote MCP connectors. Each agent is its own monthly subscription billed to your credits.
• Full API coverage. Anything you can do in the dashboard you can do programmatically: deploy and manage GPU instances under /v1/gpu/*, and deploy, message, and configure agents (model, skills, connectors, channels, and access) under /v1/hosted-agents/*. Usage and spend for both appear in /v1/account/usage.

Get started

Open GPU Cloud or Hosted Agents in the dashboard, or read the GPU Cloud and Hosted Agents guides.

Playground improvements and creative templates

A batch of playground quality-of-life updates plus the new creative templates feature for image and video generation.

What’s new

• Creative templates for image and video generation. Pre-curated effect recipes you can apply with a single click in the playground or by passing template: "<slug>" to /v1/images/generations or /v1/videos/generations. Each template picks a recommended model, ships a supported model list, and applies sensible default parameters so you can ship a polished generation without tuning every knob.
• Live cost estimate in the playground. A running estimate now appears next to the Send button before you fire a request, so you can see what each call will cost against your live balance.
• Save generated media from your usage logs. A Save button on usage log entries downloads the file (image, video, audio) for any past generation, so you can keep your output locally before the standard retention window expires.
• Edit, regenerate, and delete on text chat messages. Hover any user or assistant bubble in the playground to edit a prompt and resubmit, regenerate the last response, or mark messages for deletion.

Integration Setup Support

We added new integration setup guidance to help teams connect EmpirioLabs with coding agents, IDE extensions, CLIs, and OpenAI-compatible tools more quickly.

What’s new

• Added an Integrations guide with setup paths for popular agent tools, including OpenCode, Claude Code, Cline, Qwen Code, Codex CLI, Aider, Continue, OpenHands, Hermes Agent, goose, Zed, and Kilo/Roo/Cursor-style IDEs.
• Added a setup helper that can generate selected local configuration files for project-level and user-level workflows, with configurable tool selection, model choice, and an optional smoke test.
• Clarified the core connection values for OpenAI-compatible clients, Anthropic-style Messages clients, API keys, model IDs, and live model catalog checks.