BitdepthIndependent grades for inference endpoints
google/gemma-4-31b-it · 20 endpoints

Gemma 4 31B

First look at Gemma 4 31B: 20 endpoints on OpenRouter, 2 nights measured. Grades start at 7 nights.

Last run Nov 11, 2026
2 nights measured
0
match the reference
20
first look, not yet ranked
0
outside the reference
The board

Every endpoint, named.

#EndpointDeclaredString matchSelf-agreeNightsDrift$ in / out per MQuality-adjustedState
First look: fewer than 7 nights, not ranked
coreweave/fp4fp454.2%n/a2n/a$0.10 / $0.34$0.30First look, 2 nights
friendliunknown62.5%n/a2n/a$0.14 / $0.40$0.33First look, 2 nights
venice/fp4fp454.2%n/a2n/a$0.12 / $0.36$0.33First look, 2 nights
chutes/fp4fp454.2%n/a2n/a$0.12 / $0.37$0.34First look, 2 nights
dekallmunknown59.1%n/a2n/a$0.14 / $0.40$0.35First look, 2 nights
novita/bf16bf1658.3%n/a2n/a$0.14 / $0.40$0.35First look, 2 nights
deepinfra/fp8/flexfp854.2%n/a2n/a$0.16 / $0.32$0.37First look, 2 nights
crusoe/bf16bf1654.2%n/a2n/a$0.14 / $0.40$0.38First look, 2 nights
parasail/fp8fp854.2%n/a2n/a$0.15 / $0.40$0.39First look, 2 nights
nextbit/fp4fp454.2%n/a2n/a$0.15 / $0.40$0.39First look, 2 nights
deepinfra/turbofp454.2%n/a2n/a$0.14 / $0.51$0.42First look, 2 nights
primeintellect/fp4fp454.2%n/a2n/a$0.20 / $0.34$0.43First look, 2 nights
deepinfra/fp8fp854.2%n/a2n/a$0.20 / $0.40$0.46First look, 2 nights
modalunknown54.2%n/a2n/a$0.34 / $0.10$0.52First look, 2 nights
deepinfra/fp8/priorityfp854.2%n/a2n/a$0.30 / $0.60$0.69First look, 2 nights
sambanovaunknown54.5%n/a2n/a$0.38 / $1.15$1.05First look, 2 nights
siliconflow/fp8fp856.5%n/a2n/a$0.75 / $1.00$1.44First look, 2 nights
io-netunknown54.2%n/a1n/an/a / n/an/aFirst look, 1 night
modelrun/fp4fp454.2%n/a1n/an/a / n/an/aFirst look, 1 night
ktunknownn/an/a1n/a$0.10 / $0.34n/aFirst look, 1 night

Cite: Bitdepth · bitdepth.dev/m/gemma-4-31b-it · as of Nov 11, 2026

Reference: plurality among endpoints declaring bf16 or fp16. State rule: outside when self-agreement exceeds agreement with the reference by more than 15 points, or when the endpoint's mean on the model's first primary column is more than 8 points below the median of endpoints with 7+ nights; no_data under 7 nights; stale when the data is more than 3 days old.

Price against agreement

Cheap and wrong sits low on the left.

Each dot is one endpoint. Across: blended price per million tokens. Up: agreement with the reference.

Matches referenceFirst lookOutside reference
Blended price against agreement with the reference for 17 endpoints serving Gemma 4 31B.45%55%65%$0$0.25$0.5$0.75Blended price per million tokenscrusoe/bf16: $0.21 blended, 54.2%novita/bf16: $0.21 blended, 58.3%sambanova: $0.57 blended, 54.5%siliconflow/fp8: $0.81 blended, 56.5%deepinfra/fp8: $0.25 blended, 54.2%parasail/fp8: $0.21 blended, 54.2%deepinfra/fp8/flex: $0.20 blended, 54.2%deepinfra/fp8/priority: $0.38 blended, 54.2%friendli: $0.21 blended, 62.5%coreweave/fp4: $0.16 blended, 54.2%modal: $0.28 blended, 54.2%deepinfra/turbo: $0.23 blended, 54.2%dekallm: $0.21 blended, 59.1%chutes/fp4: $0.18 blended, 54.2%nextbit/fp4: $0.21 blended, 54.2%venice/fp4: $0.18 blended, 54.2%primeintellect/fp4: $0.23 blended, 54.2%

Cite: Bitdepth · bitdepth.dev/m/gemma-4-31b-it#price · as of Nov 11, 2026

Night by night

Does an endpoint change?

Exact match with the reference each night, one line per endpoint. × marks a night with no row.

Matches referenceFirst lookOutside reference
Exact match with the reference each night for 20 endpoints serving Gemma 4 31B.708090100Nov 10Nov 11crusoe/bf16novita/bf16sambanovasiliconflow/fp8deepinfra/fp8parasail/fp8deepinfra/fp8/flexdeepinfra/fp8/priorityfriendlicoreweave/fp4io-netmodaldeepinfra/turbodekallmchutes/fp4modelrun/fp4nextbit/fp4venice/fp4primeintellect/fp4

Cite: Bitdepth · bitdepth.dev/m/gemma-4-31b-it#nightly · as of Nov 11, 2026

For providers

Graded by Bitdepth.

An endpoint that matches the reference can show this badge on its docs or pricing page. It links back to its row here, and it updates every night.

Graded by Bitdepth badge sample

Badge code for a matching endpoint, after the first ranked read:

<a href="https://bitdepth.dev/m/gemma-4-31b-it"><img src="https://bitdepth.dev/badge/gemma-4-31b-it/PROVIDER.svg" alt="Graded by Bitdepth"></a>
Method

How this page is measured.

The reference

For each model, the answer most endpoints declaring full precision (bf16 or fp16) give at temperature zero. Where fewer than two declare it, the plurality answer across all endpoints. Method.

Agreement and self-agreement

Agreement is how often an endpoint gives the reference answer. Self-agreement is how often it gives its own answer twice. An endpoint that agrees with itself far more than with the reference is serving something different.

How rank works

Endpoints that match the reference are ranked by quality-adjusted price: the blended price, (3 × input + output) ÷ 4 per million tokens, divided by agreement. A cheap endpoint that misses the reference is never shown as the cheapest.

First look

An endpoint needs 7 nights before it is ranked. Until then it shows its night count and its numbers, with no rank.

Drift

The mean of the last three nights against the first three. A move of more than 10 points reads as improving or worsening; anything smaller is stable.

How to cite

Every chart and table carries a cite line with the date. Copy it with the figure. Wrong model name or missing endpoint? Tell us.

Numbers against a stated reference. Not a claim about intent.Listed prices from OpenRouter, read nightly; grades are Bitdepth's.