Changelog

🧠 Self-Learning Routing

Edgee can now learn from your actual AI usage to recommend more cost-efficient routing strategies.

• AI classification: Explore your activity by task difficulty, type of work, and spending to understand where your AI budget goes.
• Personalized routing recommendations: Get model recommendations based on your team's usage, with estimated savings before making changes.
• Continuous optimization: Enable self-learning to receive updated routing recommendations every week as your usage evolves.
• Compare strategies: See how different model combinations affect costs, model switching, and task difficulty coverage.

Learn more about the approach in our blog post.

🔐 Security & Infrastructure

• Two-factor authentication (2FA): Add an extra layer of security to your Edgee account.
• Static IP on demand: Request a static IP for environments that require IP allowlisting.

✨ New Model

Gemini Carbon joins the Edgee model catalog.

🔐 Security & Secret Protection

• Secret detection (beta): Edgee scans requests for API keys, tokens, and credentials, flagging secrets before they reach LLM providers. Monitor detections in the Console, configure how secrets are handled, and create custom detection rules for your organization's sensitive data. Read the docs.

🔑 API Keys & Usage

• Service keys: Create organization-owned API keys for automation and shared workloads, without tying usage to an individual. Admins can manage and rotate keys, with improved search and filters across key types.

💻 Coding Agents

• Automatic Claude fallback: Switch to a configured fallback model as your Claude subscription approaches its usage limit, helping avoid interruptions and unexpected overages.
• Continuous conversations: Resume coding-agent sessions across launches and find all related activity in one place. New conversations keep their histories separate.
• More models in Claude Code: Access your organization's enabled models directly from Claude Code's /model picker when launched through Edgee.
• Improved agent compatibility: More reliable tool handling, reasoning continuity, image support, and error reporting across providers, especially Amazon Bedrock.

🤖 Models & Providers

• New models: Fable 5.2 and Mythos 5.2 join the Edgee model catalog, with Mythos 5.2 available using your own Anthropic credentials.
• Provider status & alerts: Check real-time provider health under Models → Status, with your organization's providers listed first. Get notified of outages and degradations through Alerts → Provider status.
• More provider options: Kimi K3.1 Pro is now available through Fireworks, and GLM 5.3 through Amazon Bedrock using your own credentials.

🛠️ Console Improvements

• Easier credential management: Clearer navigation and controls for API keys and provider credentials.
• Bug fixes: Improved activity-log rendering, clipboard copying, and changelog previews.

💻 Console

• Provider management: Manage providers and model pricing, monitor provider status, and control whether new models and providers are enabled by default.
• Better usage insights: Track session savings, reroutes, and usage across members, squads, integrations, and embedding requests.

🚀 Gateway

• More routing control: Pin healthy provider routes, configure fallback providers, and customize reasoning effort for individual models.
• Claude Code improvements: More reliable fallback handling and accurate usage tracking for Claude Code sessions.

✨ Models

• New models: Meta Muse Wonder, Kimi K3.1 Pro, and GPT-6.1 Terra, with expanded provider availability.

🧩 Embeddings API

Generate embeddings through the OpenAI-compatible POST /v1/embeddings endpoint, with support for batches, custom dimensions, and base64 output. Use the same API keys, budgets, and usage tracking as chat, with fallback only between providers serving the same model.

✨ New Models

• OpenAI: text-embedding-3-small and text-embedding-3-large.
• Google: Gemini Embedding 001.
• DeepSeek: V4.1 Flash 1017, plus V4.1 Flash on Google Vertex AI.

⚡ Better Prompt Caching

• Lower costs in Claude Code: System messages added mid-conversation no longer invalidate the prompt cache, reducing repeated cache writes.
• More consistent cache reuse: Conversations stay with the same healthy provider when health scores fluctuate slightly.

🛠️ Compatibility Improvements

Improved tool handling for Gemini on Vertex AI, reasoning-effort support for DeepSeek on Azure, and compatibility with newer Claude Code clients on Bedrock Mantle.

🚀 Codex: More Models, Right Where You Work

Use Claude, Kimi, GLM, and other Edgee models directly from the native model picker in Codex CLI and Desktop, without leaving your workflow.

• Team controls: Admins can select and reorder up to 15 models or use the automatically managed catalog.
• Flexible billing: Use Edgee credits or your own provider API keys (BYOK) for models outside your ChatGPT subscription.
• Conversation portability: Reopen saved Codex conversations without Edgee.

💻 Console

• Plugin management: Install Edgee-provided plugins directly from the Console. Admins can also create, update, and distribute plugins through MCP, with multi-organization support.
• Debug log retention: Debug logs are now retained for 30 days by default. Enterprise organizations can configure retention from 1 to 30 days, including for existing logs.
• Expanded model availability: Gemini 4 Argon is available through Gemini API and Vertex AI, while DeepSeek V4.1 Flash now also supports Vertex AI.

⚡ Gateway & Routing

• More reliable Codex routing: MCP tools remain available when routing Codex requests to alternative models.
• Faster provider failover: Automatically switch to the next provider when a rate limit is reached.
• Improved caching: Better cache preservation across supported Anthropic and Amazon Bedrock routes.

🛠️ Claude Mod Improvements

Improved compatibility with other mods, corrected savings reporting, and added Escape-key support to collapse the panel when focused.

🔀 Edgee Claude Mod: Switch Models Without Leaving Claude Code

Introducing Edgee Claude Mod, letting you switch between 200+ models and monitor your session directly in Claude Code, without losing your conversation context.

• Instant model switching: Search 200+ models and route requests to another model within the same session, with one-click access back to your Claude plan.
• Live session insights: Track requests, token usage, cache hits, and routing directly in your terminal.
• Cost savings visibility: See how much you save through compression and routing, with a live savings breakdown.

Launch Claude Code through the Edgee CLI to get started.

🎬 Watch the demo

⚡ New Model: Claude Haiku 5.5

Claude Haiku 5.5 joins the Edgee model catalog, offering a new option for fast, efficient AI workloads.

🚦 Routing Strategies

• Manual strategies: Use the new guided builder to configure permanent routing, model-specific spending caps, or staged routing as budgets are consumed.
• Smart Routing (Beta): Automatically choose a model for each turn based on task complexity, with configurable models for each complexity level.
• Assignment & gradual rollout: Assign strategies to organizations, squads, members, or individual keys, and roll them out to a percentage of your team before expanding adoption.
• Subscription-aware routing: Reuse supported provider subscriptions, including GitHub Copilot, when switching models. Smart Routing can also preserve the selected agent’s subscription.
• Fallbacks & provider selection: Configure fallback chains across strategy types. Routing targets now default to automatic provider selection, including your BYOK providers.
• Temporary session reroutes: Switch models for an individual session through Edgee’s MCP tools or at launch with edgee launch --reroute …. Active overrides appear in the Console.

Learn more about Routing Strategies

💻 Coding Agents

• Desktop apps (Beta): Edgee now works beyond the terminal with Claude Desktop, Codex Desktop, and GitHub Copilot Desktop.
• Claude Code integration: The new /edgee interface adds an inline model picker, session token totals, cache statistics, and recent requests. Requires Claude Code 2.1.287 or newer.
• Expanded agent support: New support for DeepSeek Harness and Cursor’s OpenAI-compatible provider setup, plus OpenCode v2 compatibility and improved Copilot connectivity and tool calls, including GitHub Copilot for IntelliJ IDEA.

✨ Models

• New model & expanded provider coverage: Mistral Large 4 joins the model catalog, with expanded provider coverage including Kimi K3 on Amazon Bedrock.

🔭 Console & Observability

• Better team visibility: Filter activity, logs, sessions, and exports by squad. The dashboard now combines member and squad selection into a single scope picker.
• Clearer request logs and savings: Failed attempts are grouped under the final request. Routing and compression savings now include percentage views, with more accurate cache-aware calculations.

⚙️ Compression

• More compression controls: Independently override tool-result trimming, tool-surface reduction, and output brevity per request. Tool-surface reduction is enabled by default for newly enrolled Claude and Codex keys.

💳 Billing & Account

• Enterprise billing & account controls: Enterprise customers can purchase credits by invoice and access invoices directly. Eligible organizations can also request a one-time $5 test credit, and users can revoke all API access tokens at once.

🔒 Privacy & Usability

• Public pages removed: Public leaderboard, profile, and session pages have been removed.
• Console & docs improvements: Improved mobile Console navigation, with documentation now integrated directly into the Edgee website.

🤖 Models

• GPT-6.1 Sol: OpenAI’s latest model, with improved agentic coding, computer use, and complex professional workflows. Use it with openai/gpt-6.1-sol.

• Claude Sonnet 5.5: Anthropic’s latest Sonnet model, with stronger coding, agentic workflows, and knowledge work. Use it with anthropic/claude-sonnet-5-5.

🧩 Plugins

• Your team’s conventions, everywhere: Package shared skills and tools into plugins and distribute them to your organization, squads, or individual members across supported coding agents. Learn more.

💻 Console

• See what drives your savings: Understand exactly how much compression and routing each contribute to your savings across Activity, Sessions, and Logs.
• A refreshed Console: We’ve redesigned key parts of the Console for a more consistent experience, with clearer navigation and simpler workflows across Activity, Sessions, Logs, Integrations, API Keys, Alerts, and Organization Settings.
• Simpler team management: Members and seats now live in one place, with streamlined squad management and drag-and-drop member assignment.
• Improved alerts: Alerts are easier to manage, with richer Slack and email notifications that surface the key context you need.
• Billing, all in one place: Manage billing details and invoices from a centralized Billing page, with one billing profile across subscriptions and credit purchases.

🤖 Models

• Claude Opus 5.5: Anthropic’s latest Opus model, with stronger coding, agentic workflows, and complex knowledge work. Use it with anthropic/claude-opus-5-5.
• GPT-6 Sol: OpenAI’s latest model for complex coding and agentic workflows, balancing intelligence and cost. Use it with openai/gpt-6-sol.
• GPT-6 Luna: OpenAI’s most efficient GPT-6 model, built for focused, high-volume workloads. Use it with openai/gpt-6-luna.
• Grok 4.7: xAI’s latest Grok model, with improved coding, reasoning, and agentic capabilities. Use it with xai/grok-4.7.

💰 Cost visibility and savings

• Routing savings: Savings from Edgee routing are now visible across activity dashboards, request logs, and individual sessions, making it easier to track their impact at every level.

⚡ Edgee CLI and coding agents

• GitHub Copilot support: Launch GitHub Copilot Desktop and IntelliJ IDEA Copilot through the Edgee CLI with Edgee-managed configuration.

🔭 Integrations and observability

• Snowflake destination: Snowflake is now available as an observability data destination.
• Bulk model and provider management: Models and providers can now be enabled or disabled in bulk.
• Dedicated Observability section: GitHub and data integrations have moved from Settings to a dedicated Observability section for easier access and management.

🛡️ Reliability

• Better agent error handling: Agents now receive the correct errors from reroute and fallback chains, improving automatic recovery when requests fail.
• Streaming errors: Error codes are now preserved for streaming requests instead of being masked.
• Claude Code model selection: Permission-classifier requests now preserve their intended model selection.

Forward hosted-gateway usage logs to your own observability and data stack, while keeping Edgee’s built-in observability. Supports OTLP, Amazon S3, ClickHouse, BigQuery, and Snowflake.

Learn more

⚡ CLI

• Oh My Pi support: Launch Oh My Pi with Edgee models and shared Pi credentials.
• OpenCode relay: Route OpenCode requests through Edgee without replacing existing provider credentials.

💻 Console

• DeepSeek V4.1 Flash: Now available on Edgee, bringing DeepSeek's latest model with faster inference, native multimodal capabilities, and improved agent performance. Use it with deepseek/deepseek-v4.1-flash.

🚀 Gateway

• Reasoning controls: Model-specific reasoning settings are applied consistently across direct and fallback requests.
• Usage attribution: Usage reporting now preserves reroute and BYOK attribution more accurately.

Routing & request visibility

More visibility into fallbacks

See fallback activity directly in sessions and identify which fallback model was used in coding-assistant status lines.

Reasoning & cache visibility

Request logs now show reasoning-token usage alongside cache reads and writes. Reasoning configuration is also preserved across supported API dialects.

Models & providers

Richer model discovery

Sort models by performance and efficiency, compare SWE-bench classifications and cache pricing, and check Zero Data Retention and data-center availability.

New models & provider routes

New models available through Edgee include GPT-6 Astra, Gemini 3.8 Flash, Muse Spark 1.3, GLM 5.3 Flash, Fable 5.1, and MAI-Code-1.1-Flash, alongside additional provider routes.

Organization controls

Multiple usage-log export destinations

Organizations can now configure multiple destinations for usage-log exports.

Encrypted debug logs

Organization-wide encryption is now available for debug logs.

Coding assistants

Copilot CLI

Route Copilot CLI through Edgee while preserving native Copilot billing.

OpenCode

OpenCode now supports generated reasoning variants.

Edgee now supports three more coding agent from the CLI: Pi, Kimi Code, and Kilo Code.

Route your coding sessions through Edgee to reduce token costs with lossless compression and get full observability over every session.

Get started with one command:

Pi

edgee launch pi

View Pi docs

Kimi Code

edgee launch kimi

View Kimi Code docs

Kilo Code

edgee launch kilo

View Kilo Code docs

🚀 Gateway

15 new providers

Edgee now supports 15 additional providers: Baseten, Cerebras Cloud, DeepInfra, Groq, Inceptron, IONOS AI Model Hub, Meta, MiniMax, Nebius Token Factory, Nscale, OVHcloud AI Endpoints, Poolside, TensorX, Together AI, and Umans AI.

Expanded model support

This week we added:

• DeepSeek V4 Flash with a 1M context window
• Qwen 3.7 Max and Qwen 3.8 Max
• Muse Spark 1.2
• Direct xAI routes for Grok 4.6, Grok Build, Grok 4.20, and Grok 4.3

💻 Console

Better filtering for usage and logs

Usage and Logs can now be filtered by team member, coding agent, or API key. Filters can also be combined to isolate exactly the activity you’re looking for.

Smarter exports

Exports now carry over the same filters and savings calculations shown in the dashboard.

If an export exceeds one million rows, you’ll get a warning so you can narrow the selected time range.

Better organization controls

Organization administrators can now disable automatic Edgee MCP injection across the organization.

Improved Models experience

We’ve redesigned the Models experience to make it easier to discover and manage models and providers.

You can now enable or disable models and providers, with improved filtering to quickly find what you need.

⚡ CLI

Better developer experience

The CLI now includes improved onboarding, automatic shell integration, richer session summaries, and a more informative status line.

Improved reliability

We’ve improved compatibility across macOS, Linux, and Windows and fixed several launch and relay issues when working with coding agents.

View the full CLI release notes