Model comparison · accepted catalogue facts 2026-11-22

Gemma 3 4B vs Gemma 4 31B

Neither model has both an equal-or-higher measured score and an equal-or-lower estimated price in this pair. Retaining your current model is a valid outcome.

Both models belong to the Google Gemma family. Inclusion does not recommend either model. This is a direct pair comparison; global frontier membership is not its verdict.

Gemma 3 4B

LMArena general · Elo · higher is better

Scale starts at 1040 Elo

1290.8

Balanced effective USD / 1M tokens · lower is better

$0.043/M

Gemma 4 31B

LMArena general · Elo · higher is better

Scale starts at 1040 Elo

1442.5

Balanced effective USD / 1M tokens · lower is better

$0.205/M

In-Depth Analysis

Gemma 3 4B vs Gemma 4 31B: Measured Technical Breakdown

Comprehensive comparison across measured human-preference evaluation, effective production token costs, architectural limits, and migration trade-offs.

Measured Capability (LMArena)
Gemma 4 31B leads by 151.7 Elo

Gemma 3 4B: 1290.8 ±9.23 · Gemma 4 31B: 1442.5 ±7.41 Statistically significant preference separation

Cost Comparison Across Canonical Production Workloads

Rates represent effective cost per 1M tokens under standard token blends. Cached prompt reads are factored according to workload context reuse.

Workload & RatioGemma 3 4BGemma 4 31BCost DifferenceAdvantage
Balanced 3 tokens in per 1 out$0.043 /1M tokens$0.205 /1M tokens$0.163 (79% diff)Gemma 3 4B costs less
Summarise Huge input, tiny output$0.035 /1M tokens$0.153 /1M tokens$0.118 (77% diff)Gemma 3 4B costs less
Chat Long cached system prompt$0.038 /1M tokens$0.244 /1M tokens$0.206 (84% diff)Gemma 3 4B costs less
Code gen Output dominates$0.044 /1M tokens$0.296 /1M tokens$0.252 (85% diff)Gemma 3 4B costs less
Agentic Context replayed each turn$0.030 /1M tokens$0.179 /1M tokens$0.149 (83% diff)Gemma 3 4B costs less

Architectural & Operational Specifications

SpecificationGemma 3 4BGemma 4 31BOperational Note
Context window131,072 tokens262,144 tokensGemma 4 31B has a 2.0× larger context window
Max output generation117,964 tokens16,384 tokensGemma 3 4B allows longer single-turn completions
Base input price$0.040 / 1M tokens$0.140 / 1M tokensGemma 3 4B input rate is cheaper
Base output price$0.050 / 1M tokens$0.400 / 1M tokensGemma 3 4B output rate is cheaper
Cached input price$0.020 / 1M tokens$0.140 / 1M tokensGemma 3 4B offers lower cached prompt cost
Reasoning & thinking tokensStandard generationSupported; standard rates—
Input modalitiestext, imageimage, text, video—
Tool calling / function useNot listedSupportedGemma 4 31B declares function calling / tool support
Open weights & licenceProprietary (API-only)Open weights (apache-2.0)Gemma 4 31B allows independent hosting or local deployment
Provider & reseller marketGoogle · 4 provider offersGoogle · 18 provider offers (7.5× spread)—

When to Choose Gemma 3 4B

  • Higher maximum completion limit (117,964 tokens).
  • Direct integration with Google tooling and API features.
Risks when moving from Gemma 4 31B → Gemma 3 4B:
  • Context window: 262,144 → 131,072 tokens.
  • Tool use is not supported by the candidate.
  • Reasoning is not supported by the candidate.
  • video input is not supported by the candidate.

When to Choose Gemma 4 31B

  • Higher measured general capability (1442.5 Elo on human evaluation).
  • Larger context window (262,144 tokens) for long-document or repo-level context.
  • Open weights availability allows private self-hosting and on-premise execution.
  • Broader multimodal input support: image, text, video.
  • Direct integration with Google tooling and API features.
Risks when moving from Gemma 3 4B → Gemma 4 31B:
  • Output limit: 117,964 → 16,384 tokens.

Compare · model decision

Choose two or three models

Inspect the differences, estimate your workload and keep the evidence. All standard catalogue models are available, including unrated ones.

Catalogue facts
2026-11-22

2 / 3 selected · 0 requirements
Open public comparison link
Public workload and requirements

These settings are included in the share link. The fixed-request scenario and private evaluation data below are excluded.

A ↔ B

Neither model has both an equal-or-higher measured score and an equal-or-lower estimated price in this pair. Retaining your current model is a valid outcome.

Estimated blend: 75% input, 25% output; 0% cached input. Base context tier. Prompt length selects the billing tier. An unlisted cache-read rate uses that tier’s full input rate in this blend; missing required input or output rates prevent an estimate. The fixed-request estimate below validates actual context and output limits and requires a published cache-read rate when cache hits are assumed.

What changes between these models

Quality first, then estimated price. LMArena scores are human-preference measurements, not a benchmark of your application. A missing score is unrated.

  • Gemma 3 4B · Meets the stated model requirements
  • Gemma 4 31B · Meets the stated model requirements
Material differences and unknown evidence. Deltas compare each model with model A. Unknown facts remain visible.
Measured factA Gemma 3 4BB Gemma 4 31B
LMArena general score (Elo)1,290.81,442.5Δ vs A: +151.7 Elo
Score interval half-width (± Elo)9.237.41Δ vs A: -1.82 Elo
LMArena document score (Elo)Unknown1,434.3
Estimated effective USD / 1M tokens$0.0425$0.205Δ vs A: +0.1625 USD/M
Base input USD / 1M tokens$0.04$0.14Δ vs A: +0.1 USD/M
Base output USD / 1M tokens$0.05$0.4Δ vs A: +0.35 USD/M
Base cached-input USD / 1M tokens$0.02$0.14Δ vs A: +0.12 USD/M
Context window (tokens)131,072262,144Δ vs A: +131,072 tokens
Output limit (tokens)117,96416,384Δ vs A: -101,580 tokens
Input modalitiestext, imageimage, text, video
Tool useNot supportedSupportedΔ vs A: +1
ReasoningNot supportedSupportedΔ vs A: +1
Open weightsUnknownSupported
Recorded licenceUnknownapache-2.0
Distinct sellers in the catalogue215Δ vs A: +13
Serving precision differs across offersNot supportedSupportedΔ vs A: +1
Recorded retirement dateUnknownUnknown
3 shared measured facts
Shared measured facts. Deltas compare each model with model A. Unknown facts remain visible.
Measured factA Gemma 3 4BB Gemma 4 31B
Benchmark reasoning effortdefaultdefault
Model providerGoogleGoogle
Retirement evidenceNone announced; not a guaranteeNone announced; not a guarantee

Provider counts and precision flags describe catalogue offers. They do not establish the precision, availability, region, tool behaviour or limits of your chosen endpoint. Open each model’s provider table before switching.

Complete price ladders and billing conditions

Gemma 3 4B

Rates below are USD per 1M tokens. A threshold selects that whole input/output rate pair.

Input-token thresholdInputOutputCached input
Base, up to and including first threshold$0.04$0.05$0.02

Gemma 4 31B

Rates below are USD per 1M tokens. A threshold selects that whole input/output rate pair.

Input-token thresholdInputOutputCached input
Base, up to and including first threshold$0.14$0.4$0.14
Migration preflight · compare each candidate with model A

Keeping Gemma 3 4B is a valid decision. No pairwise comparison certifies a drop-in replacement.

Gemma 3 4B → Gemma 4 31B

Known loss or failed requirement

  • Output limit: 117,964 → 16,384 tokens.
  • Provider endpoint, serving precision, availability and task success need your own verification. Catalogue model capabilities do not certify an endpoint.
Check alternatives to Gemma 3 4B

Estimate a fixed request shape

Private to this tab unless you save locally or explicitly export private details. These are estimates, not observed invoices. Cross-model token counts and task success are assumptions.

0.1 extra attempts means 10 extra full attempts per 100 requests. Cache writes, tool charges, taxes and unlisted charges are excluded.

  • Gemma 3 4BEstimated total: $0.105 1,000 full attempts · base rates
  • Gemma 4 31BEstimated total: $0.48 1,000 full attempts · base rates
Prompt and output sensitivity · sampled scenarios

Every point recalculates the entire request shape. Tier-boundary samples include one token below, at and above each threshold. Ordering between samples is not guaranteed. Lowest known cost excludes unknown estimates and is not a replacement recommendation.

Vary input; hold output, requests, cache and retries fixed.
Input tokensGemma 3 4BGemma 4 31BCoverage
2,000$0.105Lowest known$0.48All computed
4,096$0.18884Lowest known$0.77344All computed
32,768$1.33572Lowest known$4.78752All computed
131,072ineligibleInput plus output exceeds the published context window.$18.55008Lowest knownIncomplete
Vary output; hold input, requests, cache and retries fixed.
Output tokensGemma 3 4BGemma 4 31B
128$0.0864$0.3312
2,048$0.1824$1.0992
8,192$0.4896$3.5568

Bring your observed evaluation results

JSON is parsed locally. Prompts, filenames, individual responses and extra fields are not retained. Import only aggregate measurements for selected models. Task/version labels do not establish that two evaluations used the same dataset.

Expected format
[{"model":"provider/exact-model-id","task":"support-answer","version":"rubric-1","attempts":100,"successes":84,"costUsd":2.4,"latencyMs":950}]

Use the exact catalogue ID. Replace the example values with observed measurements. Omit costUsd or latencyMs when they are unknown. No public score is inferred from these observations.

Keep the decision and its evidence

A portable record retains selected facts, complete tier ladders, public requirements and the publication manifest. It also embeds the full permitted public catalogue to verify the input artifact checksum. Files are larger for this reason (up to 2 MB); local storage is limited to 12 records. Team review is local to this browser; there is no hosted collaboration.

Saving is opt-in and includes private review fields. Shared comparison links never include them. Version 1 Check references remain separate and are not overwritten or guessed into this format.

Saved local decisions · 0

No version 2 decision records are stored in this browser.

Questions this page answers

Is Gemma 3 4B a better-and-cheaper replacement for Gemma 4 31B?

Neither model has both an equal-or-higher measured score and an equal-or-lower estimated price in this pair. Retaining your current model is a valid outcome.

Which model is higher rated: Gemma 3 4B or Gemma 4 31B?

Gemma 4 31B scores higher on LMArena general (1442.5 ±7.4) compared to Gemma 3 4B (1290.8 ±9.2), a margin of 151.7 Elo points. The measured confidence intervals do not overlap, demonstrating a statistically significant preference lead on human evaluations.

Which model is more cost-effective: Gemma 3 4B or Gemma 4 31B?

On a standard balanced workload (3 prompt tokens per 1 generated token), Gemma 3 4B costs $0.043 per 1M tokens versus $0.205 for Gemma 4 31B. Gemma 4 31B is approximately 4.8× the price of Gemma 3 4B. Base input rates are $0.040 for Gemma 3 4B and $0.140 for Gemma 4 31B; base output rates are $0.050 for Gemma 3 4B and $0.400 for Gemma 4 31B.

How do their context windows and generation limits compare?

Gemma 3 4B provides a 131,072-token context window with a maximum output limit of 117,964 tokens. Gemma 4 31B provides a 262,144-token context window with a maximum output of 16,384 tokens. Gemma 4 31B offers a 2.0× larger context capacity, making it suitable for larger repositories and multi-document workflows.

What capabilities are lost when switching between Gemma 3 4B and Gemma 4 31B?

Switching from Gemma 3 4B to Gemma 4 31B involves: Output limit: 117,964 → 16,384 tokens. Conversely, moving from Gemma 4 31B to Gemma 3 4B involves: Context window: 262,144 → 131,072 tokens; Tool use is not supported by the candidate; Reasoning is not supported by the candidate; video input is not supported by the candidate. Run test suites on your specific prompt templates and tool declarations before cutting over production traffic.

Can Gemma 3 4B or Gemma 4 31B be self-hosted or deployed privately?

Gemma 4 31B provides open weights, enabling self-hosting on private GPU infrastructure, local quantization, and private data pipelines. Managed inference is also offered across third-party API providers (4 for Gemma 3 4B, 18 for Gemma 4 31B).

What is included in this comparison?

Accepted model capabilities, context and output limits, complete price-tier ladders, independent LMArena measurements and missing evidence. Provider endpoint precision and application success still need verification.

Can I compare a different workload or keep a decision record?

Yes. Change the public workload, prompt length and requirements in the comparison workspace. Fixed-request scenarios, aggregate evaluations and review notes stay local unless explicitly exported.

Evidence & Ask