Side by side · November 2026

Gemini 4 Neon vs Gemma 4 31B

Gemini 4 Neon scores higher on the Models at Work leaderboard (76 vs 54 out of 100). On independent tests alone, Gemini 4 Neon is ahead (99 vs 37). Gemma 4 31B has no list price we could verify, so the two cannot be compared on cost. On the Artificial Analysis Intelligence Index, one of the independent tests, Gemini 4 Neon leads 61 to 4.

Updated Nov 19, 2026 · every number links to who measured it
Gemini 4 Neon · GoogleGemma 4 31B · Google
Leaderboard rank#4#31
Score (0–100)7654
Kind of workHardest problemsRun it yourself
Price per 1M tokens$3 to read · $15 to writeNot verified
Reads at once1M tokens (about 750,000 words)256K tokens (about 192,000 words)
Where you can get itGoogle AI Studio, Gemini for business, Gemini app (Google AI Pro and Ultra), GitHub CopilotAmazon Bedrock, Gemini Enterprise Agent Platform (Google Cloud)
Can a company host it itselfNoYes (open weights)
Artificial Analysis Intelligence Index614
Vals Finance Index68.31% (rank 1)Not measured
Vals Tax Agent Bench54.89% (rank 1)Not measured
Vals Legal Research Bench56.25% (rank 1)Not measured
Arena Text scoreNot measured1452 (rank 75)
Best forLegal and finance analysis, Working across spreadsheets, PDFs and charts together, Multi-step researchReading scanned documents, Fixed, repeatable steps in a process, Keeping data on your own machines

Cost of 1,000 jobs

1,000 of theseGemini 4 NeonGemma 4 31B
Support repliesthe customer message, the account notes and two help articles read; a 300-word reply written$15No list price
Documents processeda two-page invoice or form read; the fields pulled out$11No list price
Calls summariseda 40-minute call transcript read; a summary and next steps written$39No list price
Contracts revieweda 40-page contract read against a checklist; a memo of issues written$113No list price
Drafts writtena brief and one example read; a 600-word email or proposal section written$17No list price

Worked out from each vendor’s list price and a fixed size for each job. A product built on a model charges its own price on top. See what every model costs for the same jobs.

Choose Gemini 4 Neon for

  • Legal and finance analysis
  • Working across spreadsheets, PDFs and charts together
  • Multi-step research

First on all three Vals AI tests of finance, tax and legal research work a day after launch (68.3%, 54.9%, 56.3%), at $3 in and $15 out per million tokens: 50% above Gemini 4 Carbon until 30 November, 25% below it from 1 December.

Caveats: Released 18 November 2026 and still rolling out: its score rests on four independent tests read a day after launch and no scored reports from people running it. Artificial Analysis calls it 'notably slow per task and very verbose': it wrote 283 million tokens (the unit vendors bill by) across the index's tasks against a median of 65.1 million, so the bill can run above the list price. The other rankings in Google's launch post (Surge AI, Zapier's AutomationBench, Arena's agent index) are relayed by Google and are not counted. In the Gemini app it is for Google AI Pro and Ultra subscribers. Google's post names Gemini for business but not the Gemini Enterprise Agent Platform on Google Cloud, where Carbon and Hydrogen are sold. It ranks 4th, level with Gemini 4 Carbon on 76 and within three points of the three models above it: a tie. It has the board's best test results; price holds it back until Carbon's rises on 1 December.

Choose Gemma 4 31B for

  • Reading scanned documents
  • Fixed, repeatable steps in a process
  • Keeping data on your own machines

Google's small open model for fixed, rule-following steps: free to use commercially and able to run on one strong workstation, but 4 on Artificial Analysis's index rules out reasoning.

Caveats: Slow on a desktop: users report 12 to 16 tokens (word-pieces) a second on Apple machines, and about 70GB of memory to read its longest documents. Users say the rival Qwen models handle long documents and multi-step tasks a little better; Gemma is better at sticking to instructions. Amazon Bedrock lists it at $0.14 in / $0.40 out per million tokens in US regions if you would sooner rent than run it.

What people running them say

Gemini 4 Neon: The launch thread on Hacker News, a developer forum, drew 856 points in under a day. Commenters mostly ask which of Google's four Gemini 4 models to use for what; one who has tried it says it 'feels like it might actually be as good as Opus 5.5', another notes it costs more than GPT-6.2 Sol for a similar Artificial Analysis score. No reports yet from operations teams, and too little to score.

Gemma 4 31B: Hacker News (a developer forum) users who run models on their own computers treat it as the reference point: 'the new baseline for local models' (soganess, who saw it use 70GB of memory on an Apple M5 Max); 'particularly good at pipeline/automation tasks' and better at following rules than far larger Qwen models; 'very good at OCR' (reading text from images). One user had it catch a bug Opus 4.7 missed. Counterpoints: Qwen 3.6 'handles context a little better', peers 'predominantly run Qwen', and about 11 to 16 tokens (word-pieces) a second on Macs feels slow.

Questions people ask

Which is better, Gemini 4 Neon or Gemma 4 31B?

Gemini 4 Neon scores higher on the Models at Work leaderboard (76 vs 54 out of 100). On independent tests alone, Gemini 4 Neon is ahead (99 vs 37). Gemma 4 31B has no list price we could verify, so the two cannot be compared on cost. On the Artificial Analysis Intelligence Index, one of the independent tests, Gemini 4 Neon leads 61 to 4.

When should I choose Gemini 4 Neon over Gemma 4 31B?

Choose Gemini 4 Neon for legal and finance analysis, working across spreadsheets, pdfs and charts together, multi-step research. First on all three Vals AI tests of finance, tax and legal research work a day after launch (68.3%, 54.9%, 56.3%), at $3 in and $15 out per million tokens: 50% above Gemini 4 Carbon until 30 November, 25% below it from 1 December.

When should I choose Gemma 4 31B over Gemini 4 Neon?

Choose Gemma 4 31B for reading scanned documents, fixed, repeatable steps in a process, keeping data on your own machines. Google's small open model for fixed, rule-following steps: free to use commercially and able to run on one strong workstation, but 4 on Artificial Analysis's index rules out reasoning.

Scores come from the Models at Work leaderboard: Score is 0 to 100 and computed, not typed. 40% is independent testing (Artificial Analysis, Arena, Scale, Epoch, Vals and others; each test is scaled against the best score on this board, and tests of real work, such as finished freelance projects or professional tasks, count double the exams). 30% is price: the vendor's list price for a typical mix of reading and writing, on a fixed scale from $0.05 to $60 per million tokens, so a model that costs a fifth as much gains real ground. 15% is what people report after running it, from Signals and linked threads. 15% is fit: how much text it can take in at once and where a business can get it (the products and clouds that carry it), with a small credit for open weights. A missing component drops out and the rest are reweighted. A model with no independent test yet is provisional and ranks below every measured one. Within three points is a tie. Vendor-published figures are labelled and never counted. Prices are vendor list prices where published; a typical mix assumes three parts reading to one part writing.