Meter-SPW
[ LLM Cost Optimizer ]
Our pricing is performance-based: the router fee* only applies when verified cost savings are achieved.
You only pay when we deliver real savings. No savings, no router fee*
Every dollar saved is fully transparent and auditable, down to each request.
Satisfaction guaranteed — if approved, we'll issue a full refund of fees paid for each subpar routing decision.
See our latest product update, supported models, or FAQ
Meter cross-vendor mode benchmark 09/2026
Compare Competition
Baseline
Metric Spotlight
Cost savings × quality margin retained, relative to the $9.36 baseline cost and our selected 5/10 quality floor.
C = total cost; Q = mean quality; i = system; bar = selected floor. Scores 0% at or above baseline cost, or at or below the quality floor.
Higher is better. 100% CQE means an ideal system matching baseline quality but costs $0.
About this metric Based on hypervolume (Zitzler–Thiele, 1998), inspired by RouteLLM and RouterBench; CQE scores each mode against the baseline and selected 5/10 quality floor.
Trade-off Lab
Explore the trade-offs between cost, quality, and latency for the systems you picked.
- Best Overall Meter-SPW · Cheapest
- Best CQE Meter-SPW · Cheapest 68%
- Best Speed Meter-SPW · Cheapest 21,946 ms
Capability Profile
Evidence & Methodology
Transparent. Reproducible. Verifiable.
- Quality
- Judged by frontier LLMs: Fable 5.1 and Astra 6 independently, each scoring 1–10 in 0.5 increments against a structured rubric, with cross-validation, an in-house confidence algorithm, and a tiebreaker for disagreements greater than 1 point.
- Run-to-run variance
- Quality scores and pricing are a calculated sample from an average of 2-3 reruns for each mode. The same model can score and price differently on a rerun due to the probabilistic nature of LLMs.
- Model Pool
- All models of Meter are US models. The label "US models" means all popular models by US providers. The label "all models" means all possible models within that routing system without any restriction.
- Meter's supported models:
- See supported models
- Prompt Category
- Expert-level difficulty across 7 categories: Coding, CS Theory, Math, Factual, Science, Reasoning, Creative
- Sample prompts:
- See sample questions
- More questions?
- See the FAQ
Cheapest variants
Cheapest 92% saved
Cheapest++ 95% saved
2× Cheapest 96% saved
4× Cheapest 98% saved
Model distribution
| System | Model | Avg total tokens | Percentage |
|---|---|---|---|
| Baseline gpt-5.6-solcalled direct | gpt-5.6-sol | 3,183 | 100% |
| Meter-SPWcheapest | gpt-5.6-luna | 2,116 | 66.4% |
| grok-4-1-fast-reasoning | 890 | 20% | |
| gpt-5.6-terra | 2,745 | 12.9% | |
| gemini-2.5-flash-lite | 1,533 | 0.7% | |
| OpenRouter autohigh, US models | openai/gpt-5.6-sol | 3,381 | 62.1% |
| anthropic/claude-sonnet-4.6 | 3,539 | 30.7% | |
| openai/gpt-5.5 | 5,619 | 5% | |
| anthropic/claude-sonnet-5 | 11,620 | 1.4% | |
| openai/gpt-5.6-luna | 2,425 | 0.7% | |
| Microsoft Foundrycost, all models | grok-4-1-fast-reasoning | 1,179 | 55.7% |
| gpt-oss-120b | 4,522 | 17.9% | |
| gpt-4o-mini-2024-07-18 | 1,222 | 4.3% | |
| gpt-5.6-luna-2026-07-09 | 2,769 | 3.6% | |
| gpt-5.4-nano-2026-03-17 | 4,505 | 2.9% | |
| gpt-5.2-2025-12-11 | 2,965 | 2.1% | |
| gpt-5.6-terra-2026-07-09 | 2,424 | 1.4% | |
| gpt-4.1-2025-04-14 | 1,415 | 1.4% | |
| Llama-4-Maverick-17B-128E-Instruct-FP8 | 1,856 | 1.4% | |
| gpt-5.5-2026-04-24 | 3,988 | 1.4% | |
| o4-mini-2025-04-16 | 1,379 | 1.4% | |
| gpt-5-2025-08-07 | 5,668 | 1.4% | |
| deepseek-v3.2 | 976 | 1.4% | |
| gpt-5.4-mini-2026-03-17 | 10,164 | 0.7% | |
| gpt-5.6-sol-2026-07-09 | 939 | 0.7% | |
| gpt-5-nano-2025-08-07 | 10,076 | 0.7% | |
| gpt-4.1-mini-2025-04-14 | 1,517 | 0.7% | |
| gpt-4o-2024-11-20 | 1,136 | 0.7% |
Sample questions
-
Code write
c62-code-write-001-rust-transactional-build-graph
In stable Rust, implement `BuildGraph` with `add_target(name)`, transactional `set_dependencies(target, deps)`, `mark_changed(target)`, and `rebuild_plan()`. Replacing dependencies must either commit completely or return a concrete directed cycle without changing the graph. A rebuild includes every changed target and all transitive dependents, orders dependencies before dependents, and breaks otherwise-free ties by target creation order. Reject unknown names and duplicate dependency entries. Return compilable code, complexity bounds, and four focused tests including rollback after a cycle.
-
Code write
c62-code-write-011-cuda-segmented-scan
Implement an inclusive segmented prefix sum in CUDA C++ where `head[i]=1` starts a segment. Handle arbitrary `n` using multiple blocks, with the explicit constraint that the first pass launches at most 1024 blocks so block summaries fit in one second-pass block. Provide the three kernels plus compact host orchestration, a CPU oracle, and one randomized test loop.
-
CS theory
c62-code-theory-005-differential-privacy
Prove the privacy guarantee of the Gaussian mechanism using privacy loss, then compare basic composition, advanced composition, zCDP/Renyi-DP accounting, and privacy amplification by subsampling. Explain why adaptively chosen queries still compose and convert one concrete rho-zCDP budget to `(epsilon,delta)`.
-
CS theory
c62-code-theory-016-succinct-rank
Prove how a bit vector supports O(1) rank and select with o(n) extra bits on a word RAM. Lay out superblocks, subblocks, lookup tables, parameter sizes, and explain why storing every prefix count would cost too much. State the model assumptions and select's additional indexing idea.
-
Math
c62-math-014-reed-solomon
Over F_17, encode message polynomial m(x)=3+5x+2x^2 at evaluation points 1,...,7. Suppose received values have errors at two unknown positions. Demonstrate syndrome/Berlekamp-Welch decoding by constructing equations, locate and correct errors for a concrete corrupted word you specify, and state the decoding radius.
-
Math
c62-math-012-stability-bdf
Derive the region of absolute stability of BDF2 for y'=lambda y, and decide whether BDF2 is A-stable and L-stable. Carefully define stability for a two-step method using the root condition, and compare with backward Euler as z tends to minus infinity.
-
Factual
c62-factual-005-sovereign-debt
Explain sovereign-debt restructuring without a global bankruptcy court. Cover collective-action clauses, pari passu litigation, holdouts, aggregation across bond series, IMF lending-into-arrears policy, domestic-law versus foreign-law bonds, and why restructuring nominal principal alone may not restore sustainability.
-
Factual
c62-factual-020-survey-measurement
Explain total survey error: coverage, sampling, nonresponse, measurement, processing, and adjustment error. Analyze why weighting can reduce bias yet increase variance, how mode and question order change responses, and how to validate a cross-national latent construct with differential item functioning.
-
Science
c62-science-007-ecology-metapop
Four identical habitat patches have extinction probability e=0.20 per year when occupied. Each empty patch is colonized independently with probability 1-(1-c)^k where c=0.15 and k is the number of occupied other patches after extinction. Construct the Markov transition matrix for occupied-patch count and find or numerically characterize its stationary distribution.
-
Science
c62-science-006-population-genetics
At one biallelic locus, genotype fitnesses are w_AA=1, w_Aa=1, w_aa=0.8 and mutation A→a occurs at 1e-5 with negligible back mutation. Derive the recessive mutation–selection balance, estimate q, and compute expected affected fraction. Explain when drift invalidates this deterministic result.
-
Reasoning
c62-reasoning-016-insurance-adverse-selection
An insurer offers high- and low-coverage contracts to privately informed high- and low-risk customers. Draw incentive-compatibility and participation constraints, derive the qualitative Rothschild–Stiglitz separating outcome, and explain why a seemingly profitable pooling contract can unravel.
-
Reasoning
c62-reasoning-002-sequential-trial
A trial checks significance every week and stops the first time p<0.05, but reports an ordinary fixed-sample p-value. Explain optional-stopping inflation, construct a valid correction using alpha spending or an e-value/test martingale, and compare what each guarantees under an unbounded horizon.
-
Creative
c62-creative-012-palimpsest
Write two interleaved diary entries from occupants of the same room exactly 100 years apart. Alternate paragraphs, six per diarist. Each later diarist paragraph unknowingly completes a sentence begun in the preceding earlier paragraph, while preserving two distinct voices and revealing a shared danger.
-
Creative
c62-creative-007-deposition
Write 18 question–answer exchanges from a civil deposition. The witness never lies and answers every question literally, yet conceals that the allegedly stolen violin was theirs. By exchange 18, the attorney should prove ownership using only earlier exact wording. No speech may exceed 18 words.
Same-vendor mode Benchmark 06/2026
| System | Total cost | vs baseline | Mean quality (of 10) | Mean latency |
|---|---|---|---|---|
| Baseline gpt-5.5 called direct | $1.42 | — | 9.14 | 17,354 ms |
| Meter-SPW router (No cache mode) | $0.64 | −55% | 9.09 | 17,428 ms |
| NotDiamond router | $1.61 | +14% | 8.69 | 18,527 ms |
| OpenRouter router | $1.70 | +20% | 8.68 | 23,917 ms |
| Microsoft Foundry router | $2.32 | +64% | 9.15 | 28,174 ms |
- MS Foundry routes 53% of prompts to gpt-5 thinking with 2x longer output, costing 64% higher than baseline gpt-5.5
60% of prompts never needed the expensive model.
All routers share the same pool of similar models
| Router | Model | Avg total tokens | Percentage |
|---|---|---|---|
| Baseline gpt-5.5 | gpt-5.5 | 1,019 | 100% |
| Meter-SPW | gpt-5-mini | 1,527 | 50% |
| gpt-5.5 | 940 | 40% | |
| gpt-5-nano | 1,348 | 10% | |
| Microsoft Foundry | gpt-5 | 3,187 | 53% |
| gpt-5.5 | 1,302 | 22% | |
| gpt-5-mini | 2,174 | 17% | |
| gpt-5.4 | 1,665 | 8% | |
| OpenRouter | openai/gpt-5.5 | 1,497 | 58% |
| anthropic/claude-4.6-opus | 525 | 42% | |
| NotDiamond | gpt-5.5 | 1,263 | 58% |
| claude-opus-4-6 | 637 | 42% |