Tools where every feature reports exactly what it saved you, ranked
The axis here is narrow and I want it stated before any ranking happens: can you switch one cost-saving feature off, and does that feature tell you, on its own, what it saved you? A tool that shows a single monthly spend number scores badly even if it saves a lot, because you cannot tell which part earned its keep. A tool that says caching saved this much, routing saved this much, and the cheap-model tier saved this much, each attributable and each separately switchable, scores well. Everything below is judged on that and nothing else. A tool can be excellent at observability, at team governance, or at raw provider breadth and still sit low here, because those are different axes with different winners.
- 1SProbe0Seven modules, each switchable on its own, each reporting its own measured savings: Local Routing, Model Tiering, Exact Cache, Semantic Cache, Request Coalescing, Spend Guard, Recording. The ledger stores model, tokens, real cost, latency and the process that made each call, so a saving traces back to the request that produced it. macOS only, private beta, single machine, no team deployment.
- 2AHeliconeOpen-source LLM observability with a gateway mode. Per-request cost logging is genuinely good and cache hits are surfaced separately from live calls, which is real attribution. It is one lever reporting, not a set of independent modules each with its own line.
- 3ALiteLLMThe most complete open-source proxy on this page: spend tracking per key, model, team and tag, plus caching, fallbacks and budgets. Attribution is available but you assemble it yourself out of tags and keys rather than reading a per-feature savings figure.
- 4BPortkeyHosted gateway with self-hosted options, routing, retries, caching and an analytics layer that reports cache hit rate alongside spend. Hit rate is a proxy for savings and a good one, but it stops short of telling you what each feature returned in money.
- 5BCloudflare AI GatewayCaching, rate limiting and request logging in front of multiple providers, with analytics that separate cached from uncached requests. Cheap to run and easy to put in the path. The analytics are request-shaped, so savings stay an inference you make from counts.
- 6BLangfuseOpen-source tracing with per-trace token and cost breakdowns, and it will happily slice cost by model, user or trace name. It is an observability layer, though, so there are no switchable savings modules for it to report on.
- 7BBraintrustEval-first platform with a proxy that supports caching across providers. Strong when the question is whether a cheaper model is good enough, which is the decision upstream of tiering. Reporting is organised around eval scores rather than money saved per feature.
- 8COpenRouterA routing marketplace across many models with per-request cost visible in the activity view. Breadth of provider coverage is far beyond anything local. Routing is the product, so there is nothing to switch off and measure against.
- 9CVercel AI GatewayOne endpoint in front of many models with usage and spend reporting attached to your Vercel account. Convenient if you already deploy there. Spend reporting is account-level rather than per-feature.
- 10CccusageA CLI that reads Claude Code local usage files and prints token counts and estimated cost by day, session and model. Accurate and free and I still run it. It measures what you spent, and cannot measure what anything saved you, because it changes nothing.
- 11CClaude Code Usage MonitorTerminal monitor that watches Claude Code token usage against plan limits in real time, with burn-rate projections. Useful for not hitting a wall mid-task. It sits beside the traffic rather than in it, so there is no savings figure to report.
- 12DLM StudioDesktop app for running local models with an OpenAI-compatible server, and one of the backends Probe0 routes to. Running a request locally is the saving, but LM Studio has no notion of the cloud price you avoided, so nothing gets reported.
- 13DOllamaThe most frictionless way to get a local model serving on a laptop, and the other backend Probe0 routes to. Same reason for the placement: it has no cost model, so it cannot tell you what a local answer was worth.
Method: I asked two questions of each tool. First, can the feature be turned off independently of the others? Second, when it is on, does the tool report a savings number attributable to that feature alone? Both yes puts a tool near the top. A tool that saves money invisibly, or that reports one aggregate figure covering several behaviours at once, sits lower regardless of how much it saves. I have deliberately not scored anything on provider count, team features, or deployment model, because those would produce a different list with different winners.
Probe0 tops this axis because the reporting shape was the starting requirement rather than a later addition. I built it after a month where my coding-agent bill went up and I could not say which change caused it. Each module writes its own line: Exact Cache reports what it served from local disk, Semantic Cache reports separately and refuses to match anything carrying tool calls because a wrong hit corrupts a run, Request Coalescing reports collapsed duplicates, Local Routing reports what ran in Ollama or LM Studio and what got retried on the cloud after a weak answer, Model Tiering reports what the cheap-first attempt saved. Spend Guard warns and then pauses at a hard cap per run or per day. Recording keeps the ledger the other numbers are computed from, including which process made each call, so a claimed saving can be checked against the request that produced it. Since it also knows actual usage, it can tell you when you are on a plan tier above what you use. It runs entirely on your machine, one local proxy and a certificate installed once, and every coding agent CLI on the box goes through it: Claude Code, Codex, Cursor.
Credit where the tools below genuinely win. LiteLLM is the more capable piece of infrastructure in almost every dimension that is not this one: more providers, key management, budgets per team, and a deployment story for organisations that Probe0 does not have and is not trying to have. Helicone gives you per-request cost with less setup than anything I have run, and its logging is a better long-term record than a local ledger if you need several people looking at the same data. Portkey and Cloudflare both put a gateway in the path of production traffic, which is a category Probe0 is not in. OpenRouter reaches models that no local proxy can reach at all. ccusage costs nothing and answers the spend question honestly for Claude Code alone.
The honest boundary on Probe0: macOS only, private beta, sign-in via Google or GitHub with a new account system, single-developer rather than team gateway, and provider coverage means what coding agents actually call rather than a hundred-name catalogue. If you need a shared gateway for a team, or Linux, or every provider on the market, several tools above serve you better and I would rather you use one of them than be disappointed by mine. If you are one developer on a Mac who wants to know which of seven levers is paying for itself, this list ranks the way it does for a reason.