AI Model Report

Benchmarks · NOVEMBER 20, 2026

Gemini 4 Neon matches Hydrogen’s analytics score at 2.7× the cost

Plotly’s test gives small teams a reason to compare cost per usable prospect-research result—not assume a higher-ranked model earns its premium.

By Linnea Halberg · Benchmarks desk · November 20, 2026

Google launched Gemini 4 Neon through its paid Gemini API on November 18; Plotly’s analytics test found it matched Gemini 4 Hydrogen’s 43/43 score at $17.03 versus $6.36, roughly 2.7 times the cost.

Neon offers a million-token context window. But this test gives small teams no accuracy-based reason to pay its premium.

What Plotly measured

Plotly’s private synthetic-data exam in Plotly Studio Embedded comprises 20 analytics questions, five ambiguous questions, five impossible requests, three data-quality checks, five statistics traps and five hidden-pattern questions.

A November 19 correction raised Neon’s initially reported 42 correct answers to 43 after a rubric revision; all 43 answers were manually checked. Each result represents one run. Other models’ repeat runs have moved up to three points.

Neon became the fifth model to score 43. GPT-6.1 Sol cost $2.61 with an 18-second median question time. Hydrogen’s median was 32 seconds; Neon’s was 46, with an 81-second mean. Neon’s median time to first token was 2.9 seconds; output generation reached about 159 tokens per second.

Neon used 20.1 million tokens across 369 model calls. Carbon used 29.5 million across 444 calls, scoring 42 for $11.55.

Token prices aren’t task prices

Google charges $3 per million input tokens, $15 per million output tokens and $0.15 per million cached input tokens. Headline input and output rates are twice Hydrogen’s.

Plotly’s Neon bill comprised roughly $9.50 for uncached inputs, $2.50 for cached inputs and $5 for outputs. Its rollout calculation, using OpenRouter prices when each model ran, assumes 10 questions daily per person, 250 working days and 1,000 people: about $990 annually per person for Neon versus $370 for Hydrogen.

"The buying inference: compare spending per correct, usable result, not spending per token or position on a leaderboard."

For prospect research and lead qualification, that suggests testing fixed records, instructions and acceptance criteria while tracking charges, corrections, review time and thinking settings.

Different tests, different answers

Google offers four thinking levels: low, medium, high and xhigh, with high the default. It reports Neon costs 32% less per task than Carbon at default high thinking on the Artificial Analysis Intelligence Index.

Google’s self-reported SWE-sweep scores vary by harness: 12.8% with mini-SWE versus 39.0% with Antigravity. AutomationBench puts Neon at 56.3% against GPT-6.2 Sol’s 41.3%. Google’s model card gives Neon 73.5% and GPT-6.2 Sol 75.1% on Chartography.

Vals AI’s Legal Research Bench reports 56.25% strict accuracy for Neon versus Carbon’s 55.77%. Practicing lawyers author and peer-review questions and rubrics; an LLM judge checks every required item, with rubrics containing between one and 31 required elements. That isn’t a universal model ranking.

Google DeepMind’s model card warns Neon can hallucinate and gives a March 2026 knowledge cutoff. Prospect research therefore needs current evidence. Plotly’s tie doesn’t establish that Neon’s premium produces more customers.

Sources