Granite Switch 4.1 30B (revision 4.1.1, decision adapter)v1.7 · 10 s rule

Composite (secondary)

47.4 · Not ranked on this board

Cost and measurement conditions

$0.060 per 1,000 decisions (estimated)

Sibling estimate (*est.): the declared measured base is ibm-granite/granite-4.1-30b; granite-4.2-30b is the sibling reference at USD 0.16 per million input tokens and USD 0.65 per million output tokens. This is not a listing or tariff for the measured base and not an author API quote. The frozen price catalog has no exact 4.1-30b listing. Native server usage includes every full-prompt chunk and every generated label-read token. On the frozen 1,489-decision common set: 554,274 input + 1,489 output tokens, yielding USD 0.0602093284083277 per 1,000 decisions. Across all 1,500 decisions: 818,484 input + 1,500 output tokens, yielding USD 0.08795496 per 1,000. Granite is outside the six calibrated Gemma/Qwen throughput classes; under the existing v1.8 method the sibling estimate is retained and the GPU-time component is unavailable. No cross-family throughput mapping, observed-loop extrapolation or pod bill is used. Estimated, not a bill. USD 0.0602093284 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.08795496 over all 1,500 items).

p50 latency: 0.3 s

Self-hosted latency estimate under the shared method (assumption, not measured)

Measured 31 October 2026 on one evaluator-operated Verda RTX PRO 6000 Blackwell Server Edition GPU, nominal 96 GB, in FIN-03, Finland (EU), PAYG non-spot. Text and the decision adapter only; Guardian and speech-to-text were not evaluated. The composed checkpoint ibm-granite/granite-switch-4.1-30b@eb206725d03c0f1c409ce3aa512afe325fdf8731 identifies its decision adapter as revision 4.1.1 and ran with generative-computing/granite-switch@43ccd291a176507ae4ed5af293f8f2d947c0d451. No base-plus-adapter recomposition or direct PEFT substitute. Stock vLLM 0.30.0, Torch 2.13.0+cu130, CUDA 13.0, Transformers 5.17.0, tokenizers 0.23.3, Triton 3.7.1, BF16, max-num-seqs=1, chunked prefill=2048 and gpu-memory-utilization=0.90. TRITON_ATTN was selected explicitly; FlashInfer sampling, all-reduce and autotune were disabled to avoid an unavailable nvcc path. Backend selection can change roundoff and performance; no cross-backend numerical equivalence is claimed. The author Decision Index recipe includes a fixed system prompt; this measurement used no system prompt, followed by a State: user turn and a DECISION user turn containing the whole question, ordered option descriptions and an instruction to reply with one option label. The local wire shim reads raw first-generated-position label logprobs, not the sampled answer token. Explicit label reads use chunks of at most 128 IDs over the identical full prompt and join them using one stable global softmax. Every actual full-prompt and generated label-read token from every chunk is billed in native usage. Native output is one generated token per actual call, not the tokenless convention. All 1,500 requests returned HTTP 200: 818,484 native input and 1,500 output tokens, zero usage mismatches and zero native-behavior deviations. Semantic context is 131,072 tokens; no prompt truncation or native refusal occurred. Replay variability: 7 of 50 repeated requests were bit-identical, with ONE argmax flip and maximum absolute probability difference 0.031222338329561394. Exact probability reproducibility is not established. Repeats were recorded only; no scored answer was averaged, repaired, replaced or retried. The public same-engine label re-extraction check was fixed before measurement at a 5e-2 absolute ceiling with exact label sets, own-winner consistency and a near-tie allowance. This passed; it is not a cross-backend equivalence test. Median measured latency was 0.07952452649999486 seconds and maximum 21.930013085999917 seconds; five decisions exceeded ten seconds and count as wrong answers under the unchanged ten-second rule. The price uses the labelled sibling estimate because Granite is outside the six calibrated Gemma/Qwen throughput classes; no GPU-time component is available.