Reviews · NOVEMBER 19, 2026
Gemini 4 Neon arrives at $3/$15—but task costs complicate the pitch
Google’s new model offers a million-token context window and strong knowledge-work scores. Plotly’s analytics test shows why small teams should still benchmark a cheaper alternative.
Google began rolling out Gemini 4 Neon November 18 at $3/$15 per million input/output tokens. Cached input costs $0.15, a 95% discount. Access includes the paid Gemini API tier, Gemini for business, and Pro and Ultra access in the Gemini app and Google Antigravity.
Adapted from Gemini 4 Carbon, Neon combines spreadsheets, PDFs, charts, audio and video, Google says. Its million-token combined input-output window permits 64,000 default synchronous output tokens; million-token output requires background long decode.
Benchmark strengths, qualified
Vals AI’s evaluation reports:
| Index | Score | Rank | Cost/test | |---|---:|---:|---:| | Vals | 70.87% | Second | $7.40 | | Finance | 68.32% | First | $4.88 | | Legal | 78.36% | First | $6.10 |
Earnings Modeling Bench v1.1 scores 76.46%; Legal Research Bench, 56.25%. Neon reaches seven of 14 score–cost Pareto frontiers: no cost-reported model matches its score cheaper or beats it for equal cost.
Google self-reports 32% lower task costs than Carbon on the Artificial Analysis Intelligence Index at default high thinking. AutomationBench scores are Neon 56.3%, Claude Opus 5.5 42.5%, GPT-6.2 Sol 41.3%. Professional-chart-reading test Chartography puts Neon at 73.5%, Sol at 75.1%.
Google’s evaluation document specifies pass-at-one unless stated otherwise, mostly highest-thinking Gemini results, and generally provider-reported competitor scores. It distinguishes provider-harness computations from external evaluations. The model card warns of hallucinations; knowledge stops March 2026.
Plotly’s perfect score costs more
Plotly tested through OpenRouter on a private, synthetic wind-energy-fleet exam. Its multi-turn harness emphasizes data and charts, not writing. Four thinking levels are available: low, medium, high (default), xhigh.
| Model | Correct/43 | Run cost | Median seconds/question | |---|---:|---:|---:| | Neon | 43 | $17.03 | 46 | | Gemini 4 Hydrogen | 43 | $6.36 | 32 | | GPT-6.1 Sol | 43 | $2.61 | 18 |
Neon’s input/output list prices are twice Hydrogen’s. Against Carbon, Neon used 20.1 million versus 29.5 million tokens and 369 versus 444 calls, yet Carbon cost $11.55. Plotly attributes that difference to token pricing.
Neon spent roughly $9.50 on uncached input, $2.50 cached, $5 output; median first-token latency was 2.9 seconds, generation about 159 tokens/second. It found five hidden patterns, caught five statistics traps and rejected five impossible requests.
On November 19, Plotly corrected its initial 42-of-43 score after fixing a rubric and manually checking all 43 answers. Each result represents one run; other models’ repeat runs varied by up to three points. It isn’t a repeatability guarantee.
Test acquisition work directly
Neither evaluation establishes acquisition performance. Owners shouldn’t assume benchmark leadership saves money. Related reading: pricing and coding-benchmark analysis.
"Compare cost per verified, usable draft—not just token prices or benchmark rank."
That’s the operator takeaway, not an attributed testimonial. A suggested pilot gives Neon and a cheaper alternative identical public materials and qualification criteria: B2B prospective accounts, or consumer-business distributors and referral partners. Compare factual accuracy, staff review time and spend per usable personalized outreach draft. These aren’t demonstrated outcomes; people must verify claims and approve sending.