TL;DR
Claude Code is powerful but expensive. It burns through tokens fast, and Anthropic does not give you a proper cost dashboard. AI gateways solve this by sitting between Claude Code and the provider, logging every request with cost, latency, and token data. This post covers the top 5 enterprise AI gateways you can use to track and control Claude Code costs: Bifrost, OpenRouter, Helicone, LiteLLM, and Cloudflare AI Gateway.
Why Tracking Claude Code Costs Is Hard
If you have been using Claude Code for a while, you already know the problem. It is fast, capable, and chews through tokens at a rate that can surprise you at the end of the month.
Here is what makes cost tracking difficult:
- No native cost dashboard. Anthropic's billing page shows you total spend, but it does not break things down by session, task, or team member.
- Token-heavy workflows. Claude Code sends large context windows with every request. A single coding session can rack up thousands of input tokens before you even notice.
- No per-project visibility. If you have multiple teams or projects using Claude Code, there is no built-in way to see who is spending what.
- No budget enforcement. You cannot set spending limits per developer, team, or project natively.
You need something that sits between your Claude Code instance and Anthropic's API, capturing every request and giving you the data you need.
That is what AI gateways do.
If you want to get started with one right away, Bifrost is an open-source option that works with Claude Code by changing the base URL. More on that below.
How AI Gateways Solve This
An AI gateway acts as a proxy. Instead of Claude Code talking directly to Anthropic, it talks to the gateway first. The gateway forwards the request to the provider, and on the way back, it logs everything: tokens used, cost, latency, status, and more.
This gives you:
- Per-request cost tracking with full audit trails
- Budget controls so teams cannot overspend
- Rate limiting to prevent runaway usage
- Analytics dashboards for cost trends over time
The setup is straightforward. You change the base URL that Claude Code points to, and the gateway handles the rest.
Top 5 Enterprise AI Gateways for Claude Code Cost Tracking
1. Bifrost (by Maxim AI)
Bifrost is a fully open-source LLM gateway written in Go. It is designed for production use with performance as a priority, adding only 11 microseconds of latency overhead per request.
Claude Code compatibility: Works with Claude Code by changing the base URL. This is documented in their release notes: "You can now use Bifrost seamlessly with tools like LibreChat, Claude Code, Codex CLI, and Qwen Code by simply changing the base URL."
Cost tracking features:
- Log store: A persistent, queryable audit trail that captures cost, latency, tokens, input, output, and status for every request. Supports SQLite and PostgreSQL backends.
- Aggregated stats: Total requests, success rate, average latency, total tokens, and total cost, all queryable through a search API.
- Model catalog: Auto-synced pricing data from all providers, refreshed every 24 hours. This means cost calculations stay accurate without manual updates.
- Cache-aware cost calculation: If you use semantic caching, Bifrost calculates costs correctly for cache hits vs. misses.
- Four-tier budget hierarchy: Customer, Team, Virtual Key, and Provider Config. You can set dollar-amount budgets with reset durations at each level.
- Rate limiting: Token-based and request-based throttling at the virtual key level.
- Observability: Live monitoring, request logs, metrics, and analytics built in.
Strengths:
- Open-source (you can self-host and audit the code)
- 11 microsecond latency overhead
- OpenAI-compatible API format (drop-in replacement)
- Works with 1000+ models across providers with fallback support and intelligent routing
Limitations:
- Newer project compared to some alternatives
- Community is still growing
Check out the docs or the GitHub repo.
2. OpenRouter
OpenRouter is a unified API gateway that provides access to hundreds of AI models through a single endpoint. It handles routing, pricing, and usage tracking across providers.
Claude Code compatibility: Supports Anthropic models. You change the base URL in your Claude Code configuration to OpenRouter's endpoint and use their API key.
Cost tracking features:
- Per-request cost logging with token breakdowns
- Usage dashboard with spending history
- Credit-based system with balance tracking
- Model-level cost comparisons
Strengths:
- Wide model selection across providers
- Transparent pricing with per-token rates displayed upfront
- Easy to switch between models without config changes
- Active community and good documentation
Limitations:
- Hosted service only, no self-hosted option
- Adds a margin on top of provider pricing
- Limited governance features (no budget hierarchies or team-level controls)
3. Helicone
Helicone is an observability platform for LLM applications. It focuses on logging, monitoring, and cost tracking across providers.
Claude Code compatibility: Works as a proxy. You change the base URL to route requests through Helicone, which then forwards them to Anthropic.
Cost tracking features:
- Automatic cost calculation per request
- Usage dashboards with filtering by model, user, and time range
- Rate limiting and caching
- Custom properties for tagging requests by project or team
Strengths:
- Clean, developer-friendly UI
- Easy setup with minimal code changes
- Has an open-source version available
Limitations:
- The open-source version has fewer features than the managed service
- Advanced features require the paid plan
- Less focus on governance and budget enforcement compared to gateway-first tools
4. LiteLLM
LiteLLM is an open-source proxy that provides an OpenAI-compatible interface for 100+ LLM providers. It is popular for unifying API calls across providers.
Claude Code compatibility: Supports Anthropic models through its proxy. You set LiteLLM's endpoint as the base URL.
Cost tracking features:
- Spend tracking per API key, team, and user
- Budget limits with alerts
- Request logging with cost data
- Admin dashboard for monitoring
Strengths:
- Open-source with an active community
- Supports a wide range of providers
- Good for teams already using OpenAI's API format
Limitations:
- Written in Python, which can add more latency compared to Go-based alternatives
- Stability issues have been reported during high-traffic scenarios
- Configuration can get complex for advanced setups
5. Cloudflare AI Gateway
Cloudflare AI Gateway is part of Cloudflare's developer platform. It provides caching, rate limiting, and analytics for AI API calls.
Claude Code compatibility: Supports Anthropic as a provider. You route requests through your Cloudflare AI Gateway endpoint.
Cost tracking features:
- Request logging with token counts
- Analytics dashboard with cost estimates
- Caching to reduce repeated API calls
- Rate limiting per gateway
Strengths:
- Runs on Cloudflare's edge network (low latency globally)
- Free tier available
- Minimal setup if you are already on Cloudflare
Limitations:
- Limited governance features (no budget hierarchies or virtual keys)
- Less granular cost controls compared to dedicated AI gateways
- Fewer advanced features like fallbacks or load balancing for AI workloads
How to Set Up Bifrost with Claude Code
Setting up Bifrost with Claude Code takes a few steps. Follow the quickstart guide or read on. The core idea is that you point Claude Code's base URL to your Bifrost instance instead of directly to Anthropic.
Step 1: Deploy Bifrost
Clone the repo and run it locally or deploy it to your infrastructure:
git clone https://github.com/maximhq/bifrost.git
cd bifrost
go run .
Step 2: Configure your Anthropic provider
Add your Anthropic API key to Bifrost's provider configuration through the Web UI or config file. Bifrost will handle authentication and routing.
Step 3: Point Claude Code to Bifrost
Change the base URL in your Claude Code configuration to your Bifrost endpoint:
http://localhost:8080/openai
Since Bifrost uses an OpenAI-compatible API format, Claude Code works with it out of the box.
Step 4: Create a virtual key (optional but recommended)
Set up a virtual key in Bifrost with budget limits and rate controls. This lets you enforce spending limits per developer or team without touching Claude Code's configuration.
Once connected, every Claude Code request flows through Bifrost. You get full cost tracking, request logs, and budget enforcement in the Web UI.
Check the Bifrost docs for detailed setup instructions.
Comparison Table
| Feature | Bifrost | OpenRouter | Helicone | LiteLLM | Cloudflare AI Gateway |
|---|---|---|---|---|---|
| Open-source | Yes | No | Partial | Yes | No |
| Claude Code support | Yes (base URL) | Yes (base URL) | Yes (base URL) | Yes (base URL) | Yes (base URL) |
| Per-request cost logging | Yes | Yes | Yes | Yes | Yes |
| Budget hierarchies | 4-tier (Customer/Team/VK/Provider) | No | Limited | Yes (key/team/user) | No |
| Rate limiting | Yes (token + request) | No | Yes | Yes | Yes |
| Auto pricing sync | Yes (every 24h) | Yes | Manual | Community-maintained | N/A |
| Self-hosted | Yes | No | Partial | Yes | No |
| Latency overhead | 11 microseconds | Not published | Not published | Higher (Python) | Low (edge network) |
| Web UI | Yes | Yes | Yes | Yes | Yes |
| Cache-aware costing | Yes | No | No | No | No |
Conclusion
Claude Code is a productivity multiplier for developers, but without proper cost tracking, it can become an expensive black box. AI gateways give you the visibility and control you need.
If you want a self-hosted, open-source solution with minimal latency overhead and proper budget hierarchies, Bifrost is worth looking at. It works with Claude Code with a base URL change, gives you a persistent audit trail for every request, and lets you set budgets at the customer, team, and virtual key levels.
For teams that prefer a managed service, OpenRouter and Helicone are solid options with polished UIs. LiteLLM is a good open-source alternative if you are already in its ecosystem. And Cloudflare AI Gateway works well if you need basic analytics with minimal setup.
Pick the one that fits your stack, set it up, and stop guessing what Claude Code is costing you.
Top comments (1)
A surprising insight is that many teams overlook the cost-saving potential of optimizing prompt engineering. In our experience with enterprise teams, refining prompts and leveraging agents to automate routine tasks can drastically reduce token consumption. It's not just about the AI output but also how efficiently you generate that output. Focusing on workflow integration with lightweight agents often yields better ROI than just tracking token usage alone. - Ali Muwwakkil (ali-muwwakkil on LinkedIn)