
Gemini 4 Neon Effort Levels: What Low, Medium and High Actually Cost
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 189 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 204 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 132 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 77 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 118 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 114 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 93 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 211 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 168 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
One prompt, forty input tokens, and nine bills. Simon Willison ran the pelican-on-a-bicycle test through Gemini 4 Carbon, Gemini 4 Neon and Gemini 4 Hydrogen at low, medium and high effort, and published the token counts, the dollar costs and the wall-clock times that came back — in a blog entry dated 18 November 2026 and as a comparison grid on X the following morning. The nine results are the most concrete public measurement yet of what a Gemini 4 thinking-level setting actually does to a bill, and they say something the launch material does not: on a short, easy prompt the effort rung moves the cost more than the choice of model does, and the ranking between two models can flip when you change the rung.
His verdict on the output is deliberately unquantified. The Neon pelicans "look better than Hydrogen and worse than Carbon," detail and polish drop as you descend the family and, less sharply, as you descend the ladder, and every one of the nine images is still recognisably a pelican on a bicycle. You cannot score that. You can, however, count what it cost.
Nine bills for one forty-token prompt
The prompt is identical across all nine runs, the input is 40 tokens in every case, and the only things that vary are the model and the thinking setting. Every figure below is from Willison's own run at Google's published per-token rates, and the costs are his measurements rather than estimates.
• Gemini 4 Hydrogen, low — 2,996 output tokens, 2.25 cents, 10 seconds.
• Gemini 4 Hydrogen, medium — 3,964 output tokens, 2.98 cents, 17 seconds.
• Gemini 4 Hydrogen, high — 5,311 output tokens, 3.99 cents, 19 seconds.
• Gemini 4 Neon, low — 4,974 output tokens, 7.47 cents, 43 seconds.
• Gemini 4 Neon, medium — 7,846 output tokens, 11.78 cents, 54 seconds.
• Gemini 4 Neon, high — 9,691 output tokens, 14.55 cents, 1 minute 11 seconds.
• Gemini 4 Carbon, low — 9,153 output tokens, 9.16 cents, 1 minute 34 seconds.
• Gemini 4 Carbon, medium — 11,020 output tokens, 11.03 cents, 1 minute 53 seconds.
• Gemini 4 Carbon, high — 20,305 output tokens, 20.31 cents, 4 minutes 11 seconds.
The arithmetic behind those bills is worth stating because it explains everything that follows. At 40 input tokens the input charge is roughly one hundredth of a cent and can be ignored, so the bill is almost exactly output tokens multiplied by the output rate: Carbon at $10 per million output tokens, Neon at $15, Hydrogen at $7.50. Nothing about the effort setting changes a price. It changes how many tokens the model emits before it stops, and on this prompt that multiplier ranges from 1.77x to 2.22x.
The rung moves the bill more than the model does
Take the cheapest rung of the most expensive family. Hydrogen at low costs 2.25 cents. Neon at high costs 14.55. That is a 6.5x spread across one family and one prompt, and 3.3x of it comes from the model while the remaining 1.95x comes from the rung. The same comparison inside a single model is just as large: Neon low to high is 1.95x, Carbon low to high is 2.22x, Hydrogen low to high is 1.77x.
Put the two effects side by side and the practical point arrives quickly. Dropping Carbon from high to medium saves 9.28 cents on this prompt, which is 64% of what Neon costs for the same bird at the same rung. If your default is "the biggest model at the deepest rung," a one-step change in that default is worth more, in cash, than switching models entirely.
The direction of the cross-model comparison is also not the one the price list suggests. Neon's $3 and $15 are 50% above Carbon's current $2 and $10 — the introductory rate Carbon and Argon carry until 30 November, after which both double to $4 and $20. Yet on this prompt Neon at high is the cheaper of the two, by 28%, because Carbon emits 2.1 times as many output tokens to draw the same bird. Price per token is the wrong number to compare. Tokens per answer is the one that lands on the invoice.
Where Neon and Carbon swap places
The pair does not order consistently across the ladder, and the middle rung is where it breaks.
• At low — Neon 7.47 cents against Carbon 9.16, so Neon is 18% cheaper.
• At medium — Neon 11.78 cents against Carbon 11.03, so Neon is 6.8% dearer.
• At high — Neon 14.55 cents against Carbon 20.31, so Neon is 28% cheaper.
Carbon's token count is the reason. It rises steeply from low to medium (9,153 to 11,020) and then nearly doubles to high (20,305), while Neon climbs more evenly (4,974, 7,846, 9,691). At medium the two token counts are close enough that Neon's 1.5x price premium decides the bill; at high, Carbon's verbosity overwhelms the price difference. Anyone who ran one comparison at one rung and generalised would have got the answer backwards for the other two.

What the ladder buys on a hard evaluation
Cost is only half the trade, and the drawing prompt is a poor instrument for the other half because there is no correct pelican. Artificial Analysis supplies the missing axis: it scores Gemini 4 Neon separately at each of its four settings on the Intelligence Index v5.0.1, and the spread is far larger than anything the price ladder shows.
• Index score — 44 at low, 55 at medium, 60 at high, 61 at xhigh, as read from the tracker's page on 19 November 2026.
• Time to first answer token — 13.9 seconds at low, 20.1 at medium, 40.6 at high, 56.5 at xhigh.
• Time per Intelligence Index task — 3.7 minutes at low, 8.3 at medium, 14.8 at high, 21.5 at xhigh.
• Output speed — 97 to 110 tokens per second, rising slightly with the rung, which is the one dimension where the ladder is nearly flat.
Seventeen index points between Neon's cheapest and deepest setting is a bigger gap than separates Neon at xhigh from Gemini 4 Hydrogen's published 54 on the same index. That is the strongest argument for the ladder: on tasks with a right answer, the rung is not a polish setting, it is a capability setting. The same tracker publishes a cost-per-task figure only for the xhigh variant, $6.44, and that single point is the one most likely to be quoted out of context. It belongs to the most expensive configuration of the model, not to the model.

Why a bird is a fair test of the rung
The pelican prompt has one property that makes it unusually good for this question and one that makes it useless for a different one. It is open-ended, so every rung produces something and none of them fails outright — which means the cost curve can be read without being confounded by retries, and the quality curve can be read as a gradient rather than a pass rate. It is also unscored, so it can tell you that low effort drew a simpler bicycle and cannot tell you whether that matters for your workload.
Google's own documentation is more direct about the risk at the top of the ladder than a demo grid can be. It warns that xhigh "can cause overthinking or exhaust max_output_tokens on simpler prompts such as text or JSON extraction," and recommends evaluating at every level and settling on the lowest one that clears your quality bar. Willison tested Neon at low, medium and high for the pelicans and used the separate extra setting in Antigravity for a Three.js animation prompt instead, so xhigh has no bird in this comparison. His own note is that Argon dropped out of the grid, which is a reminder that this is one practitioner's selection and not a published evaluation set.
Read together, the two sources give a usable rule. On easy, open-ended generation, the rung buys refinement you may not need and bills you 1.8 to 2.2 times the tokens for it. On hard reasoning, the same rung buys fifteen index points, and the 5.8x increase in time per task is the price of that capability rather than a cost of indecision.
Picking a rung instead of a model
The habit worth breaking is choosing between Gemini 4 Carbon, Gemini 4 Neon and Gemini 4 Hydrogen. All three take the same request parameter, so the real decision is a pair: which model, at which rung, for which class of work. Three rules follow from the numbers above.
Set the rung per task class, not per project. Text extraction, formatting, short classification and JSON repair are the cases Google names as overthinking risks; they are also the cases where a low or medium rung produced a recognisable answer for a fifth to a third of the tokens. Deep multi-step coding, long tool-using sessions and anything where the index's fifteen points show up are where the expensive rung earns its bill.
Measure tokens, not seconds, when you compare. The time ladder and the cost ladder run in the same direction here but not at the same slope: Neon's high setting is nearly three times slower than its low setting to first token and four times slower to finish an index task, while the bill moves 1.95x. Latency budgets and cost budgets will pull you to different rungs, and only one of them shows up on the invoice.
Re-run the comparison after 1 December 2026. Carbon and Argon double to $4 and $20 per million tokens on that date while Neon stays at $3 and $15, and the pelican numbers scale linearly with the output rate. Carbon at high would then cost 40.6 cents for the same bird against Neon's 14.55 — a 2.8x gap instead of a 28% one — and Carbon at medium would cost 22.0 cents against Neon's 11.78. Every conclusion drawn from this week's rate card has a shelf life of twelve days.
Routing the rung, not just the model
Because the effort setting is a request parameter rather than a separate deployment, the whole ladder is reachable from one integration, which makes the routing decision cheap to change. OrcaRouter exposes 200+ models behind a single key at provider list price with no markup, so a vendor price step is live on our side the day it is live on Google's — the 1 December doubling of Carbon and Argon included, with nothing to renegotiate. The routing DSL can send a task class to a specific model and rung, which is how a per-class policy like the one above becomes configuration rather than a code change, and automatic failover means a rung you are still evaluating never becomes the only path to production.
One honest limit: our catalogue currently carries google/gemini-4-carbon and google/gemini-4-argon, and it does not yet carry Gemini 4 Neon or Gemini 4 Hydrogen. So this is not a recommendation to run Neon through us today. It is the shape of the test to run when it lands, and the reason the per-rung comparison matters more than the per-model one: a policy that fixes the model and ignores the rung leaves most of the bill unmanaged.

What would change this reading
Three things, all checkable. The first is a published per-task cost for Neon at high rather than xhigh; the tracker currently gives $6.44 for xhigh only, and that single row would settle whether Google's 32%-cheaper-per-task claim survives independent measurement. The second is whether anyone runs the pelican at xhigh, since the only setting with a documented failure mode on short prompts is the one absent from the grid. The third is 1 December, which changes the arithmetic of every comparison in this piece without either model changing at all.
Questions the numbers keep inviting
Does a higher effort level cost more per token?
No. The per-token rate is fixed per model: $3 and $15 for Neon, $2 and $10 for Carbon today, $1.50 and $7.50 for Hydrogen. What changes is the number of output tokens the model emits before it stops, which is why the same 40-token prompt produced nine bills between 2.25 and 20.31 cents.
Is the cheapest rung always the cheapest answer?
Not necessarily, because a rung that under-thinks a hard task produces an answer you discard and re-run. On the pelican every rung produced something usable; on the Intelligence Index the gap between low and high is sixteen points, so a low-rung answer to a hard task is a different product rather than a cheaper one.
Which of Neon and Carbon is cheaper?
It depends on the rung and the date. Today Neon is cheaper at low and at high and dearer at medium; from 1 December it is cheaper at every rung on this prompt, because Carbon's output rate doubles while Neon's does not.
