Two frontier models launched today, and the menu got bigger and pricier — not simpler.

A free, MIT, fully local CLI that reads your LLM call logs and tells you which calls a cheaper model could have handled. It runs entirely on your machine — no network, no LLM calls, just tokenizing, applying provider pricing, and doing arithmetic — so nothing about your prompts leaves the box. Feed it two ways: point it at existing OpenAI-format JSONL, or run `frugon capture` as a local proxy that logs your traffic with no added latency. `uvx frugon analyze ./logs.jsonl` and you get a per-model breakdown; the demo run flagged $205/mo in potential savings (37%) on a $549 baseline, using the 30–50% range from the RouteLLM research it's built on. Reach for it before you touch any gateway config — it's the audit that tells you whether routing is even worth it for your specific workload, rather than assuming it is. Delete the eyeball-the-dashboard guesswork about which endpoints are burning tokens. The honest part is the caveat Frugon writes itself: "quality is not verified — 'within tolerance' is an offline estimate; run `--measure` to confirm it on your real outputs before you switch." That `--measure` flag samples your traffic through candidate models using your own keys sent straight to the provider, never through Frugon — the verification step most cost dashboards quietly skip. Tradeoff: three weeks old, 53 stars, and the savings it estimates are a starting point to test, not a promise.

A single Go binary (Apache-2.0) that sits between your coding agents and every provider, keeping your keys and traffic inside your own network. It routes on a three-tier cheap/mid/frontier system: policy floors force high-risk tasks to a frontier model, `X-Foreman-Task-Type` / `X-Foreman-Risk` headers let the caller hint, and otherwise it "picks the cheapest healthy model the tier allows." It translates protocols (Claude-format clients can run against OpenAI-compatible upstreams with full tool-call translation), pins a session to one model so you don't thrash the prompt cache, and fails over on a circuit breaker. Every request writes a ledger entry with measured tokens and cost, and `foreman trace <id>` shows exactly why it picked what it did and what it rejected. `brew install northwood-systems/tap/foreman`, then `foreman init && foreman serve`. Reach for it once Frugon's audit says routing is worth it and you'd rather not hand a hosted proxy your keys and prompts. Delete the OpenRouter-or-LiteLLM-plus-a-hosted-dashboard stack where a third party sees every request. The tell that it's built by someone serious: it "never publishes a savings number" — the cost figures come from your actual ledger, not a marketing estimate. Tradeoff: this is one day old (created yesterday, 16 stars) — the freshest bet on the slate. The routing and failover read production-minded, but nobody's run it at scale yet, and anything that can't cross protocols (images, documents) gets an explicit 400 rather than a silent flatten — correct behavior, but you will hit it.

The hosted alternative to Foreman is a crowded desk, and today three of them launched. **Auriko** (auriko.ai) bills itself as a "trading desk for LLM calls" — ex-quant-trader founders scoring every provider on cost, latency, throughput, and success rate, claiming ~30% off; closed and hosted, your traffic through their box. **Opper AI** (opper.ai) is the grown-up version: one key to 300+ models, EU-hosted with no prompts stored by default, per-team spend caps, 50k+ developers, no token markup (3% on top-ups) — also closed, but a real company with real traction. **Gate** (Constellation) is the security-first entry — #1 across 16 prompt-injection benchmarks (97.4% F1, methodology on arXiv) with lossless prompt compression bolted on for a claimed 20–50% token savings — again a hosted proxy you point `base_url` at. All three want to sit between you and your model; that convenience is the tradeoff. Off the thread but worth knowing: **Flint** (github.com/microsoft/flint-chart, MIT) is Microsoft Research's intermediate visualization language — an agent emits a compact, human-editable chart spec and Flint compiles it to Vega-Lite, ECharts, or Chart.js, deriving the scales, axes, and spacing so the model stops hand-tuning low-level config and produces charts that render reliably and stay editable. `npx -y flint-chart-mcp` drops it in as an MCP server, or `npm install flint-chart` for the library. The cleanest first-party answer yet to "my agent's charts come out broken."