A generated title card reading 'Qwen 4 vs Gemini 4 Argon: the 4.3% that decides more than the seven points', with the subtitle 'Gemini 4 Argon hallucinates on 4.3% of the questions it gets wrong; Qwen4-Max on 28.4%'. Three numbered rows read 'index score 42.9 against 50.4, Artificial Analysis v5.0.1', 'cost per completed index task $3.27 against $4.97' and 'hallucination rate 28.4% against 4.3%'. A footer reads 'Scores and per-task cost per Artificial Analysis Intelligence Index v5.0.1; prices per each vendor's published rate card.' Flat vector styling on white with blue-and-cyan accents, the OrcaRouter mark composited bottom right.
Guides & Insights

Qwen 4 vs Gemini 4 Argon: The 4.3% That Decides More Than the Seven Points

Author

Gideon Frost

Date Published

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

The scoreboard everyone reaches for in this matchup is the headline index: 42.9 for Qwen4-Max against 50.4 for G​emini 4 Argon, seven and a half points in G​oogle's favour. But the number that should stop a deployment decision sits three columns to the right of the one everybody quotes. G​emini 4 Argon hallucinates on 4.3% of the questions it gets wrong. Qwen4-Max hallucinates on 28.4% of them. That is a factor of 6.6, and for anyone shipping output that a human reads, it is worth more than seven index points.

The two figures come from the same suite, Artificial Analysis v5.0.1, and neither is a vendor claim. G​oogle has not published a hallucination rate for Argon, and A​libaba has not published one for Qwen4-Max. The independent measurement is the only one that exists.

Two models sold into the same slot

Both are million-token, image-capable, text-out flagships from companies that build their own silicon. Qwen4-Max went metered on 16 November 2026 in six A​libaba Cloud regions; G​emini 4 Argon has been on sale since 30 September 2026 through G​oogle's API. The cards are close enough that price is not the differentiator people assume.

• Input, per million tokens — G​emini 4 Argon $2.00; Qwen4-Max $1.50 in Singapore and $1.413 in the other five regions. G​oogle is 33% more expensive on the Singapore comparison and 41% more against the Global rate.

• Output, per million — Argon $10.00; Qwen4-Max $4.50 and $4.24. This is the larger gap: 2.2× on the headline number.

• Cached input — Argon $0.10 per million; Qwen4-Max $0.075 on a read, plus published write rates of $1.875 (five minutes) and $3.00 (one hour) that G​oogle does not break out the same way.

• Cost per completed index task — Argon $4.97; Qwen4-Max $3.27. Despite a 2.2× lower output rate, Qwen4-Max is only 34% cheaper per task, because most of both bills is input.

• Effective reasoning effort — Argon is measured at high; Qwen4-Max at xhigh. Both models expose a ladder, and neither number is comparable to the other model's default.

• Context — 1,000,000 tokens on Argon; Qwen4-Max is the same million on the vendor card, with the tracker recording 983,616 because thinking-mode input is capped there.

• Hosts — Argon is served by G​oogle; Qwen4-Max by A​libaba Cloud. Neither is open-weight.

A generated two-column scoreboard titled 'Qwen4-Max vs Gemini 4 Argon - the scoreboard'. The Qwen4-Max column reads index score 42.9, cost per index task $3.27, cents per index point 7.6, input per million $1.50, output per million $4.50, hallucination rate 28.4%. The Gemini 4 Argon column reads index score 50.4, cost per index task $4.97, cents per index point 9.9, input per million $2.00, output per million $10.00, hallucination rate 4.3%. A footer reads 'Scores and per-task cost per Artificial Analysis Intelligence Index v5.0.1; prices per each vendor's published rate card.' The OrcaRouter mark is composited bottom right.

The per-task split does not favour the cheaper card

Decompose the $4.97 and the $3.27 and the surprise is which one spends more on thinking. Argon emits 116,368 output tokens per index task, of which 68,301 are reasoning; Qwen4-Max emits 104,940, of which 68,914 are reasoning. The reasoning volume is nearly identical. The difference is the answer: Argon writes 48,067 tokens of final answer per task against Qwen4-Max's 36,025. Argon is more verbose where it is visible and equally verbose where it is not.

That is a cheerful anomaly in G​emini's favour, because it means the 2.2× output rate is buying 33% more answer text. The per-task premium of 52% is therefore smaller than the rate card suggests, and for workloads where output dominates — long-form generation, code the model writes itself — the gap narrows further. For workloads where input dominates, which is most RAG and most agent loops, Qwen4-Max's input rate does the work instead.

Worth stating plainly: these are modelled costs from a benchmark harness, not invoices. Your token mix will differ, and the ratio between the two models will move with it.

Where Qwen4-Max is genuinely ahead

The case against Qwen4-Max in this matchup is strong but it is not unanimous, and two rows go the other way in a way that matters.

• GDPval — Qwen4-Max 1,808.69 against Argon's 1,633.51. On economically grounded professional tasks, Qwen4-Max is 10.7% ahead of a model that beats it by 7.5 index points. This is the clearest single inversion in the comparison.

• FrontierCode 2.0 — 26.5% against 25.7%. Within a point; effectively a tie, on the suite where Argon's headline lead would predict a gap.

• AutomationBench — 72.6% against 77.5%. Argon wins, but by five points on a suite that is closer to office automation than to software engineering.

• Terminal-Bench 4.0 — 48.5% against 57.1%. Here Argon's lead is real: eight and a half points on shell-driven work.

• HLE Diamond — 44.9% against 56.7%. Twelve points on hard science.

• Terminal-Bench Science — 23.3% against 52.9%. The widest gap on the board, and the one that most cleanly separates the two models' research capability.

• Omniscience accuracy — 27.7 against 46.1, with the hallucination caveat from the top of this piece attached to both.

Read as a set, the pattern is that Qwen4-Max holds its own on professional-judgement tasks and loses badly on anything requiring sustained technical reasoning. That is a narrower and more useful statement than "seven index points back", and it points at a different deployment.

A headless-browser screenshot of the Artificial Analysis page for Gemini 4 Argon, captured 17 November 2026 in English, showing Google's mark, the label Proprietary model, Released September 2026, the heading Gemini 4 Argon: Intelligence, Performance & Price Analysis, a rank strip reading Intelligence #22 of 81, Performance #72 of 79, Cost #59 of 80 and Verbosity #57 of 80, a statistics row giving a 45.598-second time to first answer token, 66 output tokens per second, $2.00 input and $10.00 output per million tokens, a 95% cache discount, 7.55 billion input tokens and an 11.5% output-to-input ratio, a $4.97 per-task cost, and the summary paragraph describing the model as above average in intelligence and reasonably priced for its band, notably slow per task and somewhat verbose, with text and image input, text output and a 1M-token context window.

Why a 4.3% hallucination rate is worth paying for

Put a number on the difference. If a reviewer can check a claim in thirty seconds and a fabricated claim costs ten minutes of rework when it reaches production, then the expected review cost per thousand answers is driven almost entirely by the fabricated share rather than the wrong share. Argon's 4.3% means roughly one answer in twenty-three of its incorrect set needs rework; Qwen4-Max's 28.4% means roughly one in three and a half.

Set against that, the 52% per-task premium disappears for any workflow that has a human reviewer, because reviewer time dwarfs token spend at almost any loaded rate. It disappears in the other direction for any workflow that does not, because a fabricated answer with no reviewer is not a cost, it is an incident.

The honest caveat: hallucination rate is measured on a fixed question set and is a property of the measurement as much as the model. G​oogle could change Argon's behaviour with a system-prompt revision tomorrow and this number would not move until the suite is re-run. Treat it as a strong prior, not a guarantee.

The routing question, answered for one side

Gemini 4 Argon is on our catalogue as google/gemini-4-argon at G​oogle's list price, $2.00 in and $10.00 out per million, with 0% markup on top — the same rates G​oogle publishes, passed through rather than repriced. Qwen4-Max is not on our catalogue: a qwen4 identifier returns not-found, its route is A​libaba's Model Studio API and several third-party platforms, and the closest Qwen model we serve is qwen/qwen3.8-max at $2.00 and $6.00.

That asymmetry is itself the answer to the deployment question for most teams. The model that wins the hallucination column is the one you can call today through the endpoint you already have, and the model that wins on GDPval costs you a second vendor contract to try. If you want to test the GDPval claim on your own data before committing, a routing DSL lets you send a fixed evaluation set down both paths from one call and compare the outputs side by side — which is a cheaper experiment than either model's per-task price suggests.

A headless-browser screenshot of the OrcaRouter model page for Gemini 4 Argon, captured 17 November 2026 in English, showing the identifier google/gemini-4-argon, the byline Google, an input price of $2.00 per million, a zero-data-retention badge, the note that zero data retention is available under agreement, and both the OpenAI-compatible and generateContent endpoint paths.

What the seven points do not tell you

Two things would change this comparison, and neither is far off. An independent hallucination measurement on a Qwen4-Max deployment with a system prompt tuned for abstention would test whether 28.4% is a property of the model or of how it was queried — no vendor has published one. And a GDPval re-run on a larger sample would tell us whether Qwen4-Max's 10.7% lead on professional tasks is real or noise, because that benchmark's intervals are wide and the two models are close enough for the ordering to flip.

Until either lands, the split is clean and it is not the one the index suggests. Sustained technical reasoning, science, shell work, or anything a customer reads unreviewed: G​emini 4 Argon, and the 52% premium is the least interesting number in the deal. Grounded professional judgement at volume, with a review step you were going to run anyway: Qwen4-Max is 10.7% better on the suite that measures exactly that, and it is 34% cheaper per task while you find out.