{"slug": "gemini-4-neon-vs-carbon-vs-hydrogen-vs-argon-hydrogen-answers-twice-as-fast-neon", "title": "Gemini 4 Neon vs Carbon vs Hydrogen vs Argon: Hydrogen Answers Twice as Fast, Neon Costs the Most", "summary": "A developer at apimodels.app measured production latency across Google's four Gemini 4 models over 30 days and found Gemini 4 Hydrogen roughly twice as fast as Argon and Carbon for answers of 500 to 4,000 tokens, with a median of 23 seconds versus 44 and 49 seconds in the 1,500–4,000 token band. Hydrogen is also the cheapest at $0.825/$4.125 per 1M tokens, while Neon tops Google's self-reported Coding Agent Index but costs $3/$15. Argon was fastest only for answers over 4,000 tokens, and Carbon's slowest tenth of short replies exceeded eight minutes.", "body_md": "The short answer: use [Gemini 4 Hydrogen](https://apimodels.app/models/gemini-4-hydrogen) for agent loops and bulk coding. For answers of 1,500 to 4,000 tokens it took a median 23 seconds on our platform, against 44 for Argon and 49 for Carbon, and at $0.825 / $4.125 per 1M tokens it is the lowest-priced. Use [Gemini 4 Argon](https://apimodels.app/models/gemini-4-argon) when unsure and [Gemini 4 Carbon](https://apimodels.app/models/gemini-4-carbon) for the hardest debugging. [Gemini 4 Neon](https://apimodels.app/models/gemini-4-neon) tops the coding-agent index but costs $3 / $15; on low thinking it cost half as much in our test.\n\nIn about six weeks Google released four Gemini 4 models: the flagship Argon, announced on 30 September and on the API from 13 October; Carbon, opened to all paid developers on 10 November; Hydrogen on 12 November; and Neon on 18 November. The names are all chemical elements, which tells you nothing about speed or price.\n\nA word on who we are: we run apimodels.app, a platform where one API key calls the whole Gemini 4 family, Claude, GPT and other image, video, audio and language models. The latency figures come from production calls on that platform over the last 30 days, sample size not disclosed, with our own test accounts removed. Prices were checked on 19 November 2026.\n\nWe will not paste another benchmark table here. This piece answers three practical questions: which one is fast, which one is worth the money, and which job goes to which model. Speed comes from real calls on our platform over the last 30 days. Capability comes from Google's published numbers, labelled as such. Neon has been live for a day, so instead of production data we ran our own six-task test on its thinking levels.\n\nAt the same answer length, Gemini 4 Hydrogen is about twice as fast as Argon and Carbon from 500 to 4,000 tokens; only for answers over 4,000 tokens is Argon the fastest. Averages mislead here. Some people run long agent jobs on Hydrogen while others ask Argon short questions, so answer lengths vary a lot. We split real requests into four groups by answer length and looked at the median and the 90th percentile in each. Neon is too new to include.\n\nThe middle two groups belong to Hydrogen. For answers of 500 to 1,500 tokens its median was 15 seconds against 25 for Argon and 30 for Carbon. For 1,500 to 4,000 tokens the gap widens to 23 seconds against 44 and 49, and its 90th percentile was only 49 seconds, so even its slow requests stay reasonable.\n\nThe ends of the chart tell a different story. For short answers under 500 tokens, all three medians sit between 9 and 16 seconds, but Carbon's slowest tenth took more than eight minutes: it can think for a long time before writing a short reply. For answers over 4,000 tokens Argon was actually the fastest, with a 94-second median against 113 for Hydrogen and 137 for Carbon.\n\nWhat this means in practice: if your program is an agent that calls the model step after step with a few thousand tokens per step, moving to Hydrogen roughly halves the run. That is how people on our platform use it, too. Hydrogen requests show far more long outputs and cache reuse than the other two, the signature of an agent loop rereading the same codebase. For one long report in a single call, Argon is not slow. Carbon is a poor fit anywhere a person is waiting on screen.\n\nBy Google's published numbers, Gemini 4 Neon ranks first on the Artificial Analysis Coding Agent Index, Hydrogen ties Argon on the composite intelligence index, and Carbon ranks first on CursorBench 4.0 and Perplexity's WANDR. All of these are self-reported by Google, and each model was shown on different tests, so they cannot be ranked side by side. What they do show is the job Google built each one for.\n\nThe standout is Hydrogen. On Google's own composite index it ties Argon, while running about twice as fast and costing a quarter less per token. For most teams the most cost-efficient way to start is to run your own tasks on Hydrogen and send only the ones it gets wrong up a tier.\n\nMoving Gemini 4 Neon from its default high thinking level to low cut the cost of our six tasks from 24.7 cents to 12 cents, with all 37 checks still passed. On high it will think even before replying \"OK\". On 19 November we ran six auto-graded tasks at both levels: write an interval-merging function, fix a bug in a parser for durations like \"1h 30m 15s\", pull the final order out of an email thread as JSON, analyse a table of monthly revenue by region, solve a friction-on-a-slope physics problem, and write a short Chinese product blurb with strict length and format rules. On high, the Chinese writing task alone burned 7,397 thinking tokens.\n\nHalf the price, yet the total time for all six was about the same: 280 seconds on high, 288 on low. The difference is when the answer starts. Artificial Analysis measured time to first answer token at 40.6 seconds on high and 13.9 seconds on low. Our advice: set Neon to low for everyday coding and raise it only for problems it cannot crack. In the OpenAI-compatible format, setting reasoning_effort to low does it; the top xhigh level is only available through thinkingLevel in Gemini's native format.\n\nA typical request that reads 20,000 tokens and writes 1,000, with no cache, costs about 2.1 cents on Gemini 4 Hydrogen, 2.75 cents on Argon or Carbon (5.5 cents from 1 December) and 7.5 cents on Neon:\n\nThe formula is input tokens times the input price plus output tokens times the output price. For Hydrogen that is 20,000 × $0.825 / 1M + 1,000 × $4.125 / 1M, about 2.1 cents. Argon and Carbon come to 2.75 cents, or 5.5 cents from 1 December, and Neon to 7.5 cents. Per 1,000 such requests: about $20.6 on Hydrogen, $27.5 on Argon or Carbon ($55 from December) and $75 on Neon. Prices as of 19 November 2026.\n\nTwo things push a real bill above this. Thinking tokens are charged at the output rate, and Carbon and Neon on its default high level think the most. Cache hits save money (cached input is $0.04125 per 1M on Hydrogen and $0.15 on Neon), but Google decides when a cache hit happens, so budget as if it will not. Failed requests are not charged.\n\nThe most common setup is Hydrogen as the workhorse, with the small share of requests it gets wrong escalated to Argon or Carbon. The call names are gemini-4-hydrogen, gemini-4-argon, gemini-4-carbon and gemini-4-neon. On APIMODELS they share one API key and one endpoint, so switching is a one-field change; request formats are in the [LLM docs](https://apimodels.app/docs/llm).\n\nFor the same workhorse-or-flagship question on the OpenAI side, see our [GPT-6 Astra test](https://apimodels.app/blog/gpt-6-astra-api-tested): on the same everyday tasks Astra, at 3.4 times the price, tied the workhorse. The lesson rhymes with this one: use the fast model day to day and bring in the big one for hard jobs.\n\nIt depends on the job. Hydrogen is best for agent loops and bulk work: on APIMODELS it answers about twice as fast as Argon at the same length and costs the least. Argon is the safe default, Carbon suits the hardest multi-step reasoning, and Neon has Google's top coding scores but the highest price per token, so trial it first.\n\nPer 1M tokens, input / output: Hydrogen $0.825 / $4.125, Argon and Carbon $1.10 / $5.50 (from 1 December 2026, $2.20 / $11.00), and Neon $3.00 / $15.00. The first three are 45% below Google's list price and Neon is at list. Thinking tokens bill at the output rate, and failed requests are free.\n\nIn real calls on APIMODELS over the last 30 days, Hydrogen took a median 23 seconds for answers of 1,500 to 4,000 tokens against 44 for Argon, and 15 against 25 seconds for 500 to 1,500 tokens. For answers over 4,000 tokens it flips: Argon's median was 94 seconds and Hydrogen's 113.\n\nStart on low. In our six auto-graded tasks on 19 November 2026, the default high level cost 24.7 cents and low cost 12 cents, with all 37 checks passed both times and similar total time. Artificial Analysis measured time to first answer token dropping from 40.6 seconds to 13.9. Raise it only for problems low cannot solve.\n\nYes. On APIMODELS all four use the same API key and endpoint. The call names are gemini-4-hydrogen, gemini-4-argon, gemini-4-carbon and gemini-4-neon, so switching is a change to the model field, and the same key also calls Claude, GPT and image or video models.\n\n*This article was first published on the [APIMODELS blog](https://apimodels.app/blog/gemini-4-neon-vs-carbon-vs-hydrogen).*", "url": "https://wpnews.pro/news/gemini-4-neon-vs-carbon-vs-hydrogen-vs-argon-hydrogen-answers-twice-as-fast-neon", "canonical_source": "https://dev.to/super_lewis/gemini-4-neon-vs-carbon-vs-hydrogen-vs-argon-hydrogen-answers-twice-as-fast-neon-costs-the-most-3664", "published_at": "2026-11-19 15:11:05+00:00", "updated_at": "2026-11-19 15:13:49.560567+00:00", "lang": "en", "topics": ["large-language-models", "ai-tools", "mlops", "ai-agents"], "entities": ["Google", "Gemini 4 Hydrogen", "Gemini 4 Argon", "Gemini 4 Carbon", "Gemini 4 Neon", "apimodels.app", "Artificial Analysis Coding Agent Index", "CursorBench 4.0"], "also_reported_by": [], "takeaway": "Gemini 4 Hydrogen cut median latency roughly in half versus Argon and Carbon for 500–4,000 token answers (23s vs 44s and 49s at 1,500–4,000 tokens) at $0.825/$4.125 per 1M tokens, but Argon was fastest above 4,000 tokens at a 94-second median.", "recommendation": {"advice": "Route agent-loop and bulk coding calls with a few thousand tokens per step to Gemini 4 Hydrogen to roughly halve run time and token cost, and reserve Argon for single calls producing over 4,000 tokens.", "playbook": "llm-cost", "playbook_url": "https://wpnews.pro/playbooks/llm-cost"}, "alternates": {"html": "https://wpnews.pro/news/gemini-4-neon-vs-carbon-vs-hydrogen-vs-argon-hydrogen-answers-twice-as-fast-neon", "markdown": "https://wpnews.pro/news/gemini-4-neon-vs-carbon-vs-hydrogen-vs-argon-hydrogen-answers-twice-as-fast-neon.md", "text": "https://wpnews.pro/news/gemini-4-neon-vs-carbon-vs-hydrogen-vs-argon-hydrogen-answers-twice-as-fast-neon.txt", "jsonld": "https://wpnews.pro/news/gemini-4-neon-vs-carbon-vs-hydrogen-vs-argon-hydrogen-answers-twice-as-fast-neon.jsonld"}}