· via dev.to (home feed)
Google launches Gemini 4 Neon at $3/$15 per million tokens, aimed at agentic workloads
Google launched Gemini 4 Neon on November 18, a mid-tier model tuned for fast agentic task completion at flat $3/$15 per million tokens, while Carbon and Argon introductory rates double on December 1.

Google launches Gemini 4 Neon
Google introduced Gemini 4 Neon on November 18, 2026, according to a dev.to write-up — the fourth release in the Gemini 4 family after Argon (October 13), Carbon (November 10) and Hydrogen (November 12). The pitch, as the article relays it, is not peak intelligence but near-frontier judgment that carries multi-step tasks through to completion at a cost and latency that holds up under all-day, high-volume use.
Where Neon sits in the lineup
Per the article, Hydrogen, Neon and Carbon share a single request schema, a unified 1,048,576-token context window and long decode up to 1M output tokens, so model IDs can be swapped without code changes. Hydrogen is the fast, cheap option at $1.50 input and $7.50 output per million tokens, and the only tier with a free tier, whose traffic is used to improve Google's products. Carbon is the deliberate, heavier tier. Neon sits between them at moderate speed, with thinking levels from low to xhigh and high as the default. One migration trap the article flags: Carbon lacks an xhigh setting and defaults to medium thinking, so the level should be set explicitly when moving between models. Pricing stays flat across the full context window regardless of prompt length.
Pricing and the December 1 crossover
Neon has no introductory rate: a flat $3 input and $15 output per million tokens, with cached input at $0.15. Carbon and Argon's launch pricing of $2/$10 lapses on November 30, 2026, after which both double to $4/$20, with cached input rising from $0.10 to $0.20. Neon is therefore 1.5x Carbon's token price before the cutoff and 0.75x after it. The article's practical warning is that high-frequency Carbon workloads — code review, refactoring, document extraction, agent loops — double in cost on December 1 if left untouched.
Google's published comparison, as reported, puts Neon's cost per task 32% below Carbon's at the default high thinking level, despite the higher rates. The article's author works this back to Neon billing roughly 45% of Carbon's tokens for equivalent work, while clearly labelling that figure a personal calculation rather than a Google disclosure.
Benchmarks: agentic wins, general intelligence third
The launch figures show a deliberate trade. On general intelligence Neon scores 61 on the Artificial Analysis Intelligence Index — third overall, measured at the xhigh thinking level — with Google's own framing placing it behind larger, slower models. On agentic work it fares better: first place on the Artificial Analysis Coding Agent Index (58.6), Zapier's AutomationBench (56.3%) and Andon Labs' Vending-Bench 2 ($16,843), plus second on the Vals Index (70.9%) and Agent Arena (1,336 across more than 4,200 real agent sessions). The most instructive number is SWE-sweep: the same model scores 12.8% under the mini-SWE harness and 39.0% under Antigravity's, meaning the harness alone tripled the result — evidence, the article argues, that investing in agent tooling can matter more than swapping models. It also cautions that ranks cover only comparable models, thinking levels vary between scores, honesty metrics are Google-internal, and public benchmarks carry contamination risk.
Token accounting and caching
Neon's output price includes thinking tokens, and max_output_tokens caps thinking plus visible output together, making the thinking level the primary cost lever rather than prompt length. Caching is implicit-only: it activates on prompts of at least 4,096 tokens and discounts hits by 95%, while the Gemini 3-era explicit cache APIs (cachedContents and caches.create) are unsupported, so leftover fields should be removed. Stable material belongs at the prompt prefix, since one changed character breaks the cache. The Batch API halves costs, and Flex and Priority inference are supported. The article adds that Gemini 3.8 Flash moves to Hydrogen's price point on January 1, 2027, and that Google Antigravity's Orchestrator pairs Neon as planner with Hydrogen as executor.
Why it matters
Neon's launch is less a capability leap than an economics play. The December 1 price doubling turns model selection into a budget decision with a hard deadline: workloads that stay on Carbon double in token cost, while Neon's flat rates and apparent token efficiency position it as the default for repetitive agentic tasks. For teams building agents, the SWE-sweep harness gap may be the bigger lesson — a threefold swing from tooling alone dwarfs typical model-to-model gains. One caveat: every figure here comes from a single dev.to article and could not be cross-checked against Google's own documentation, so pricing and benchmarks should be verified directly before committing spend.
- #gemini
- #llm
- #api-pricing
- #ai-agents