A generated title card reading Qwen4-Max vs Gemini 4 Neon, with the subtitle 28% cheaper per index point, 17 points behind, and three flat line icons suggesting a cost meter, a rising score line and a stopwatch.
Guides & Insights

Qwen4-Max vs Gemini 4 Neon: 28% Cheaper per Index Point

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Qwen4-Max and G​emini 4 Neon are forty-eight hours apart and seventeen Intelligence Index points apart, and the interesting number is neither of those. It is $0.076 against $0.106 — what one point of Artificial Analysis Intelligence Index costs you on each model, once you divide the measured cost per task by the score. Qwen4-Max, generally available since 16 November 2026 at $1.50 in and $4.50 out per million tokens, buys index points 28% more cheaply than G​emini 4 Neon does at $3.00 and $15.00. Neon is the better model by a margin no one should argue about. Qwen4-Max is the better purchase per point, by a margin that is smaller than the price gap and larger than most buyers assume.

Two models, two different questions

Before the arithmetic, the framing, because these two are not competitors in the ordinary sense and treating them as one is how the comparison goes wrong.

G​emini 4 Neon is the fourth model under G​oogle's G​emini 4 heading, and it is positioned against its own siblings rather than against A​libaba. It defaults to high effort and carries four rungs — low, medium, high, xhigh — where Carbon and Argon stop at high. Qwen4-Max carries three — low, high, xhigh — and also defaults to high. Both are text, image and video in, text out. Both run a million-token context. Both are proprietary, both are closed-weight, and both are sold on the same pitch: long-horizon agentic work where the model has to hold a large working set and keep going.

What separates them is what that pitch costs. Neon is priced as the most expensive tier in its family — $3.00 and $15.00 flat, no introductory window, no free tier, against Argon at $2.00 and $10.00 through 30 November. Qwen4-Max is priced as the cheap end of the frontier, $1.50 and $4.50 in Singapore and $1.413 and $4.24 in the Global-scope regions. On the rate card that is a two-to-one gap. On the meter it is not, and the reason is the score.

What each index point costs

Artificial Analysis publishes both models on index v5.0.1 — the rebased index, so an older score from a summer leaderboard is not comparable — and publishes a measured cost per index task alongside. That second figure is the one that matters, because it folds in how many tokens each model actually burns to finish, which is where price cards lie.

• Intelligence — Qwen4-Max 43, against Neon 61 at xhigh and 60 at high. Neon's default is high, so the fair single number for Neon is 60.

• Price per million, input and output — Qwen4-Max $1.50 and $4.50 in Singapore, $1.413 and $4.24 elsewhere. Neon $3.00 and $15.00, flat everywhere.

• Cost per index task — Qwen4-Max $3.27. Neon $6.44, and that figure is measured at xhigh rather than at its default, which is the only cost the tracker headlines for Neon.

• Cost per index point — Qwen4-Max $0.0760. Neon $0.1056 at xhigh, and a lower figure at high, which nobody has published.

• Output tokens to finish the suite — Qwen4-Max 127M, Neon 123M at high and 283M at xhigh. At their defaults the two are within 3% of each other on verbosity, which is the fact that makes the rest of the arithmetic clean.

• Time per index task — Qwen4-Max 29.8 minutes, Neon 14.8 at high and 21.5 at xhigh. This is the largest functional gap in the comparison and no amount of pricing closes it.

• Output speed — Qwen4-Max 59 tokens per second, Neon 101 at high and 110 at xhigh.

• Context — Qwen4-Max 984,000 as measured, 1,000,000 as documented. Neon 1,000,000 as documented and as served, with no measured window published.

Read the cost-per-point row twice, because it does something the price row does not. The rate card says Neon costs twice as much. The meter says Neon costs 1.97 times as much per task — nearly exactly twice, which means the two models spend a similar number of tokens per task and the per-point saving is coming almost entirely from the price, not from efficiency. Qwen4-Max is not buying its 28% by being terse. It is buying it by being cheap.

And then the score row takes most of it back. Twenty-eight percent cheaper per point is a real advantage and a small one; seventeen index points is a tier change. A team that picks Qwen4-Max to save 28% per point has decided that the difference between 43 and 60 does not matter to their workload, and the price advantage is the consolation, not the argument.

A generated two-column scoreboard for Qwen4-Max and Gemini 4 Neon across six shared dimensions: index 43 against 60 at high, price $1.50 and $4.50 against $3.00 and $15.00 per 1M, cost per index task $3.27 against $6.44 at xhigh, time per task 29.8 minutes against 14.8, output speed 59 against 101 tokens per second, and a 1M context on both sides.

Speed is where the gap widens

Per-point cost is the right unit for a batch job and the wrong one for anything in front of a person, and this comparison punishes that distinction harder than most.

Neon decodes at 101 tokens per second at its default and 110 at xhigh. Qwen4-Max decodes at 59. That is a 1.7× gap at the default and a 1.9× gap at the ceiling, and it lands on top of a per-task time gap of 14.8 minutes against 29.8 — the same order of magnitude, which is what you would expect when the models emit similar token counts at different rates.

For an IDE assistant, a support agent, or any surface with a spinner, 59 tokens per second is on the slow side of usable and 101 is comfortable. The tracker describes Qwen4-Max as "notably slow per task" and its time-to-first-answer-token at 137 seconds, which is the number that decides whether a streaming UI feels alive. If your workload is interactive, the 28% per-point advantage is being purchased with a latency cost that no rate card shows and that users will notice before finance does.

The reverse case is worth stating plainly too, because it is the case Qwen4-Max is actually sold for. A long-running agent that wakes on a schedule, works for twenty minutes and writes a patch does not care whether the tokens arrive at 59 or 110 per second — it cares whether the run finishes before the next one is due. On that shape, 29.8 minutes against 14.8 is still a real cost, but it is a cost in wall-clock rather than in attention, and wall-clock is the thing you can buy more of.

Context and the ceilings nobody quotes

Both models advertise a million-token window, and both have a smaller number behind it that decides what a real long-context job can do.

• Input ceiling — Qwen4-Max documents 991,808 tokens in non-thinking mode and 983,616 in thinking mode, the difference being that reasoning and answer come out of the same bucket. Neon's documentation carries a 64K standard output ceiling with long-decode generation extending toward 1M, which is a different kind of limit: an extension rather than a flat rate.

• Output ceiling — Qwen4-Max 131,072 non-thinking, or 262,144 shared between thinking and answer. Neon 64K standard, extending under long-decode generation.

• Effort ladder — Qwen4-Max low, high, xhigh, defaulting to high, with no medium: Chat Completions maps medium to high, so asking for the middle rung costs what the top rung costs. Neon low, medium, high, xhigh, defaulting to high, with all four available.

• Caching — Qwen4-Max publishes explicit and implicit caching, with Singapore cache reads at $0.075 per million and writes at $1.875 for a five-minute window. Neon carries implicit caching only, at $0.15 per million — a 95% discount on a $3.00 input rate.

• Batch — Qwen4-Max supports batch inference in Beijing only, at $0.707 and $2.12 for Batch File and the real-time rate for Batch Chat. Neon runs batch and flex at half rate, $1.50 and $7.50, in every region.

That last row is the one that changes a real bill and gets the least attention. A batch workload on Neon costs $1.50 and $7.50 — a third of Qwen4-Max's Singapore rate on input and 60% on output — and it is available wherever Neon is. Qwen4-Max's batch path is one region, which means a team outside Beijing running asynchronous bulk work is paying the full $4.50 output rate while the comparison article they read assumed the batch price. If your workload is batch, the 28% per-point advantage becomes something much larger, and if it is real-time, it stays where it is.

A screenshot of the OrcaRouter model page for google/gemini-4-argon, captured in an English interface, showing the OrcaRouter Models navigation, the google/gemini-4-argon identifier, a time-to-first-token figure and the Pricing panel with zero-data-retention availability noted.

Where each one belongs

The honest recommendation is narrower than either vendor would like and it does not pick a winner.

Neon is the right default for anything where a wrong answer is expensive and a slow one is not — long-horizon coding, multi-step research, agentic loops that run unattended and get reviewed after the fact. Seventeen index points is the kind of gap that changes what a model can be trusted with rather than how fast it finishes, and at $6.44 a task you are paying for the ceiling, not the average. Against Carbon it is also the only G​emini 4 rate card that does not move on 1 December, which makes it the stable choice inside its own family even though it is the expensive one today.

Qwen4-Max is the right default for high-volume work where the model's job is well-specified and the cost multiplies — tagging an incoming queue, drafting the routine reply, running the cheap rung of an escalation ladder, and the long tail of agent steps that do not need a frontier model behind them. At $3.27 a task against $6.44 it is roughly half the bill for a model that scores 43, and 43 sits above the median of models in its own price band. Buying Neon for that work is buying seventeen points you will not use.

The trap is the middle, and it is where most teams land. A workload that needs 55 will not run on Qwen4-Max and does not need Neon, and the correct answer is neither of these two — it is a model one rung down in each family, priced between them, which is exactly the decision a routing layer exists to make. OrcaRouter does not serve Qwen4-Max or G​emini 4 Neon; what it does carry from the G​emini 4 line is google/gemini-4-argon and google/gemini-4-carbon, both at Google's list prices with 0% markup passed through, so the 1 December increase on those two lands here the same day it lands on G​oogle's card. Qwen4-Max is reachable through A​libaba's own Model Studio API and several third-party platforms. One API for 200-plus models means the classification tier and the frontier tier can be different models without being different contracts, and automatic failover means the day one of them has a bad hour the pipeline does not.

A screenshot of the OrcaRouter Models catalogue, captured in an English interface, showing the models count and provider count in the header, the filter rail with input modalities, context length, input price, status, access tier and inference region, and individual model rows with copyable model IDs.

The number to carry out of this

Qwen4-Max is 28% cheaper per index point and 17 points behind, decodes at 59 tokens per second against 101, and costs $3.27 a task against $6.44 at the setting the tracker headlines. Every one of those numbers is a different answer to "which is cheaper," and they are all correct.

The one that should decide it is the workload, not the model. If a task completes in one pass and the answer is checked, Qwen4-Max halves the bill. If a task completes over twenty minutes and the answer ships unchecked, Neon is the only one of the two that finishes it — and 1.7× the decode rate is what makes the difference between an agent that is useful and an agent you stop waiting for. Price the points you actually consume. Nobody has ever been glad they bought seventeen they did not.