· via dev.to (home feed)
Gemini 4 Neon tops coding agent index and cuts per-task cost 32% vs Carbon
Google's mid-tier Gemini 4 Neon ranks first on the Artificial Analysis Coding Agent Index and claims 32% lower per-task cost than Gemini 4 Carbon, but most headline scores were run at a higher thinking setting.

What happened
Google launched Gemini 4 Neon on November 18, 2026, a mid-tier addition to the Gemini 4 family priced at $3 per million input tokens and $15 per million output tokens. According to a dev.to benchmark breakdown, the model ranks first on the Artificial Analysis Coding Agent Index with a score of 58.6, and Google says it completes tasks for 32% less than Gemini 4 Carbon when both run at the default high reasoning setting.
Coding results lead, with harness caveats
Neon's strongest published numbers sit in agentic coding. It scored 65.2% on Terminal-Bench 4.1, a run carried out by Artificial Analysis with the public Antigravity CLI harness (version 1.5.9) using accuracy figures adjusted for benchmark gaming, though as the dev.to breakdown points out, that harness is Google's own tooling. FrontierCode 2.0 came in at 46.2%, one of the few results produced at the production-default high setting rather than a higher one.
The SWE-sweep results show how much the harness matters: 39.0% inside Google's Antigravity, where sub-agents fan out across a codebase, versus 12.8% on the plainer mini-SWE agent. Both figures are Google-computed, and in that same comparison Opus 5.5 ran on Claude Code while GPT-6.2 Sol ran on Codex.
On agent benchmarks, Neon placed second on the Agent Arena Index at 1,336, built from more than 4,200 real agent sessions, and first on Zapier's AutomationBench at 56.3% and on Andon Labs' Vending-Bench 2 with $16,843 in simulated earnings. Google-reported safety figures include a 1.98% refusal rate on Terminal-Bench 4.0 tasks, a 0.0% cheating score on the same suite, and a 99.2% pass rate on its agent honesty evaluation.
The 32% cost claim has an asterisk
Google's methodology states that, unless noted, results were produced at the maximum xhigh thinking level as single-attempt pass@1 runs, with competitor scores taken from the providers themselves. The headline 32% saving against Carbon, however, is measured at high. The dev.to analysis flags the mismatch directly: buyers can have the top-of-table scores or the cheaper production runs, but not both from the same numbers, and the gap needs to be measured on their own prompts. Google additionally claims roughly 50-80% lower per-task cost than leading competitors on the Vals Index and Agent Arena at high, and at least 36% lower at xhigh against rivals running at maximum settings.
Not first everywhere
Google's own launch post put Neon third on the broader Artificial Analysis Intelligence Index, behind larger and slower models, and second on the Vals Index at 70.9%. A second dev.to comparison adds that the same Artificial Analysis page showed Neon sixth among 96 models by November 19, a reminder that launch-week rankings churn quickly. That post scores Claude Fable 5.2 first on Vals at 71.31%, Neon second at 70.87% and GPT-6.2 Astra eighth at 67.66%; on the Artificial Analysis Intelligence Index it lists 61 for Neon at xhigh against 62 for Astra and 65 for Fable. Google itself still recommends Carbon for the hardest long-horizon projects.
Cheap per token, not always cheap per task
Neon undercuts rivals on unit price, roughly 62.5% below the $8/$40 listings of Astra and Fable, but per-task economics tell a different story. On the Artificial Analysis test set, per-task cost came to $6.44 for Neon, $4.00 for Astra and $26.54 for Fable. The reason is verbosity: Neon generated 283 million output tokens across the index at xhigh, against Astra's 65.8 million. Run at high instead, Neon produced 123 million tokens and lost a single point. Wall-clock time per task was 21.5 minutes for Neon, 13.9 for Astra and 41.8 for Fable. Neon also caps output at 65,536 tokens per request, and per the comparison post its API rejects legacy parameters such as thinking_budget and frequency_penalty with HTTP 400 errors while ignoring custom temperature, top_k and top_p values.
Why it matters
The launch illustrates the widening gap between leaderboard numbers and production economics. Top scores measured at xhigh on a vendor's own harness do not transfer to default-setting traffic, and per-token price is a poor proxy for per-task spend when models differ more than fourfold in how much they write. Teams choosing between frontier models should benchmark their own prompts at both thinking levels and log token counts rather than pass/fail alone, because as the two dev.to analyses show, those numbers can flip the ranking entirely.
- #gemini
- #benchmarks
- #coding-agents
- #llm