Weave Router is a hosted coding-agent router with Apache-2.0 source. How it picks models, where it beats Nadir, and how Nadir differs, cited from its docs.
Questions answered on this page
What is the difference between Nadir and Weave Router?
Nadir charges nothing on your own keys, changes no base URL, and receives only the latest prompt for a decision. Weave Router takes over the endpoint and charges 5% on your own keys. Weave Router is Weave's hosted routing endpoint for coding agents, with one-line installers for Claude Code, Codex CLI, opencode and pi, and an Apache-2.0 self-hostable repository with about 5,600 GitHub stars. It routes every LLM call in an agent session, subagents included, and passes Claude and ChatGPT subscription logins through when a plan covers the turn.
How does Weave Router decide which model to use?
Per request, a scorer weighs each call against your settings, anchored on the model the agent asked for, and moves up a tier as readily as down. It prices the provider-cache eviction a switch would cause, pins warm sessions and filters by context window. The repository's default policy is an in-process cluster scorer. The optional HMM policy needs Google's gemini-embedding-2 embeddings when self-hosted, and the hosted Codex installer selects it. The managed service runs Weave's production model, which its FAQ says is not the policy published in the repository. Nadir decides from a trained difficulty classifier over the prompt, the candidate menu, policy, cache, context, and session economics.
What does Weave Router do after the model answers?
Nothing checks the answer itself. The 2.0 launch post describes an escalation classifier that watches a task as it runs and moves it to a frontier model when a cheaper one is stalling, looping or missing the point. Nadir's decision-only mode returns before generation. On the managed, non-streaming path a complete response can enter a reference-free verifier and escalate when it misses the configured bar. Streaming bypasses that post-generation cascade.
What does Weave Router cost compared with Nadir?
Weave Router: Prepaid credits are debited at upstream cost, with a 5.5% top-up fee (minimum $0.80). Bring your own provider keys and Weave debits a 5% platform fee on that spend from a funded credit balance, so BYOK still needs a balance. Router seats are billed separately from the Weave platform: Code Max is $10 a month and limited to open-source models, Code Boost is $100 a month with no open-source restriction, and Enterprise is custom. Nadir has no base fee. Credits for Nadir-managed keys draw down at provider list price, with a 5.5% platform fee (8% on Business) added when you buy them. On your own keys, your provider bills you directly and Nadir adds no fee. Routing savings carry no fee, and shadow and recommendation-only deltas are advisory estimates that are never billed.
What does Weave Router do well?
Claude and ChatGPT subscription pass-through, with a strict mode. Routes every call in the loop, subagents included, with cache cost priced in. Published end-to-end Codex benchmarks and an Apache-2.0 repository.
Why pick Nadir over Weave Router?
Your own keys cost nothing extra. Weave debits a 5% platform fee on bring-your-own-key spend and needs a funded credit balance before that mode turns on. Nadir adds no fee on your own keys on any plan, and has no base fee or seat price. On managed keys the top-up fee is 5.5% at both. Every hosted Nadir decision comes with a receipt for model, effort, cache, policy and economics, plus an outcome endpoint for measured follow-through. Weave's own FAQ says benchmarks run against its public repository will not match its managed router. No base-URL change, and no change in who runs inference. Nadir Auto asks for a decision and your provider runs the request. Weave's own docs say Claude Code Remote Control is unavailable while its custom base URL is active. Nadir receives the latest prompt, allowed model IDs and cache accounting, never the transcript, answers or credentials. Nadir publishes no universal deployed savings or quality percentage. The 60% cost and 98% quality-preserved RouterBench result, measured on 11,420 held-out triples, is a reference-assisted research ceiling rather than the live reference-free path.
How do I trial Nadir against Weave Router?
Pick the work. Choose 10 to 20 real coding tasks that have tests or a review bar you trust, and record the model, effort and spend each one used before the trial. Run one router at a time. Turn Weave off for the trial session (`npx @weave-os/router off --claude`, or the matching flag for your harness), then install Nadir Auto with `curl -fsSL https://getnadir.com/install/claude-auto.sh | sh` for Claude Code or `codex-auto.sh` for Codex, and start `nadir-auto` or `nadir-codex-auto`. Inference stays with your provider. Check what ran. `nadir-auto --status` reads local receipts of each recommendation, the model that served the turn and cache accounting. Mark finished work with /pass or /fail, or send your test results to the outcome endpoint. Compare completed work, not token price: accepted changes, full-session cost, retries and elapsed time, with seat fees counted on both sides. Streaming bypasses Nadir's verifier, so judge quality with your own tests. Then pick one router per team, not both.
What Nadir is
Nadir is an LLM router. Nadir sends each prompt to the cheapest model
that can handle it. A small decision model
reads each prompt against your model menu and picks the lowest-cost model
it expects to handle it. When the decision model is uncertain or
unavailable, the request goes to the strongest model on the menu. The
routing step makes no call to your model provider.
Nadir runs two ways. The decision API returns a model, reasoning-effort,
cache, context, and policy recommendation without calling a model
provider, beside the gateway you already run. That is how a shadow-mode
evaluation works, and its projected savings stay advisory. The OpenAI
compatible managed proxy executes the route, and migration is a two-line
change: point the base URL at api.getnadir.com and set model to auto. On
that path an optional verifier can score a complete non-streaming answer
and escalate to a stronger model when it misses the configured bar.
Streaming bypasses post-generation verification. BYOK is supported on
every tier.
For coding agents, Nadir Auto is a terminal launcher for Claude Code or
Codex that asks Nadir for the main-session model before each turn, with
your configured model as the fallback and cost ceiling. Delegation
integrations for Claude Code, Codex, or Cursor instead recommend a model
tier for subagent work, and the agent decides; the Claude plugin can also
move a main-thread turn to Haiku when Nadir's cache check prices it
cheaper, unless its switch_main_model option is off. Either way inference runs
on your own provider account, so Nadir holds no provider keys on that path.
What the numbers are, and what they are not
Nadir publishes each evaluation with its scope. On checkable code,
run-check-escalate solved 392 of 395 common HumanEval and MBPP problems
(99.2%), graded by running the canonical tests; that applies only to tasks
with runnable deterministic tests. Nadir-Tumbler posts an arena_score of
72.3 on RouterArena's public scorer, 5th of 23 routers, which measures the
routing decision on RouterArena's own model pool. A reference-assisted
RouterBench evaluation over 11,420 held-out triples produced a 60% lower
projected cost than always-Opus with about 98% retained quality and a 1.7%
catastrophic-route rate. That experiment gave the verifier the
expensive-model reference answer, which production does not have, so it
is a research ceiling and not the deployed path.
None of these is a production guarantee, a universal savings rate, or a
forecast for any particular workload. Customer savings are reported from
measured execution against a declared baseline, and customer quality only
from outcome-labelled traffic. Projected savings and realized savings are
separate artifacts and are never blended.
Design-partner program
Three rungs, picked by risk appetite. Rung 0 Shadow runs advisory decision
calls alongside live traffic and returns a projected receipt, with nothing
in the request path changed. Rung 1 Hosted is the two-line swap on a
production slice and returns a realized receipt. Rung 2 On-prem is a
supervised six-week proof of concept inside the partner's VPC, where no
prompt, response, or usage reaches Nadir. The commitments are the same at
every rung. Apply for a rung directly:
Rung 0 Shadow,
Rung 1 Hosted, or
Rung 2 On-prem.
Not sure which fits? Start at
getnadir.com/contact/?reason=design-partner.
Licensing
NadirClaw is the self-hosted core, source-available under the PolyForm
Noncommercial License. Source-available is the correct label; NadirClaw is
not open source. Nadir Route's hosted plan has no base fee and charges a
variable fee only on measured savings from requests Nadir executed.