Gemini 4 Argon Benchmarks: What the Scores Show and Hide

Gemini 4 Argon benchmarks look strong on Google’s model card: Argon posts the top score in 13 of 19 rows, ties one and trails in five. The picture changes once you ask who ran each number. Google computed Argon’s own score in 10 of the 19 rows, and the independent Artificial Analysis index puts Argon level with GPT-6 Astra at 53, not ahead of it.

This guide reproduces the full table from Google’s Gemini 4 Argon model card, translates each row into plain language, and separates vendor-reported numbers from independent ones. Figures are as of October 16, 2026, three days after Argon became broadly available.

It is written for readers who have seen “ranked first” headlines and want to know which ones hold up. For a general introduction to the model, start with our Gemini 4 Argon overview.

Did Gemini 4 Argon really rank first on the benchmarks?

Not on everything. Counting the model card’s 19 rows ourselves, Argon has the highest score in 13, ties GPT-6 Astra on CWE-bench v1 and trails a rival in five. Against GPT-6 Astra alone it wins 14, ties 1 and loses 4. “Ranked first” is true of specific rows, not of the model overall.

The five losses are FrontierSWE v2, Terminal-bench 4.0, PostTrainBench, Terminal-Bench Science 0.1 and the OSWorld-2.0 offline subset. Each is a test where an agent works for a long stretch in a terminal or on a desktop. Argon is 10.5 points behind on FrontierSWE v2 and on Terminal-Bench Science 0.1, and 9.0 points behind on Terminal-bench 4.0.

The wins are uneven too. Argon leads by 12.9 points on Harvey’s legal benchmark and 12.4 on long-context GraphWalks, but five of its 13 wins are by less than two points. Headlines quote 12, 13 or 14 wins because outlets count different row sets, and some say “of 18”. We count the 19 rows as published.

The card also compares Argon with only three rivals: GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Claude Sonnet 5.5 is absent, yet Vals shows it within two points of Argon on the Vals Index (67.04% against 68.90% as of October 13, 2026).

What do the Gemini 4 Argon benchmarks measure in plain language?

Each row is a different kind of test. Some score whole business workflows, while others check one narrow skill, such as reading charts or tracing links across long text. The table below reproduces Google’s 19 rows and adds a plain-language reading of each. Bold marks the top score, and n/a means the card prints no number.

Which rows matter most for learning and development teams?

Four rows map to common L&D work. Harvey’s legal benchmark resembles policy and compliance document review, GraphWalks resembles tracing references across a large document set, Chartography resembles reading LMS and survey dashboards, and LVBench resembles searching recorded training video. Each comes with a limit on how far the score can be trusted for that job.

• Compliance and policy content (Harvey). Argon’s 19.6% against Astra’s 5.4% is the widest gap in the table, but the best score is still under 20%, so the test is hard for every model. It does not measure whether a model can write compliant training, and legal or regulatory content still needs an expert reviewer.
• Large document sets (GraphWalks). The 256k to 1M row is Argon’s strongest long-context result, 84.2% against Astra’s 71.8%. The graphs are synthetic and the score is F1, not the share of tasks solved, so treat it as evidence of retrieval across long text, not proof for messy policy libraries.
• Dashboards and reports (Chartography). Argon’s 71.6% to Astra’s 71.0% is effectively a tie on a set of about 100 prompts.
• Training video (LVBench). Argon scores 91.7% to Astra’s 87.5%, but Google gave each model a different frame budget, which the next section explains.

Who ran each Gemini 4 Argon number: Google, Vals or a public leaderboard?

The model card table does not say. It states no thinking settings and does not mark which rows Google computed itself. That information sits in Google’s separate evaluation methodology document. Reading it, Google ran Argon’s own score in 10 of 19 rows, and nine rows come entirely from outside evaluators or public leaderboards.

Self-run does not mean inflated. Argon loses four of the 10 rows Google computed and wins seven of the nine outside rows. The caution is about the middle group, where Argon’s number and its rivals’ numbers come from different harnesses and budgets. The methodology notes that Argon’s OSWorld-2.0 score is the best of three runs, that its Terminal-Bench Science run used a 6x verifier timeout, and that safety filters can blank responses in Agent’s Last Exam. Competitor numbers are mostly vendor-reported, at the maximum reasoning setting each vendor publishes.

LVBench shows the budget problem best. Google gave Argon 1 frame per second, which is about 3,600 frames for a one-hour video, against 800 frames for Astra, 300 for Fable 5.1 and 600 for Opus 5.5, because of API limits. Some of Argon’s lead may simply reflect seeing more of the video.

To be plain about it: the card itself gives no thinking setting for any row and no marker for self-computed rows. You have to open the methodology document to find them.

How do you read a vendor benchmark claim before you trust it?

Ask eight questions of any vendor benchmark claim: who ran it, at what setting, against whom, under what budget, on how many items, what the scale means, whether outsiders agree, and whether it resembles your work. The checklist below pairs each question with an example from Argon’s own table.

1. Who ran the number? A Vals or leaderboard score stands differently from one a vendor computed. For Argon, 10 of 19 scores were Google-run.
2. At what settings? Look for thinking level, number of attempts and pass@1 versus best of several. Google reports pass@1 at the highest thinking settings. Artificial Analysis ran Argon at “high” and the GPT-6 models at “max”.
3. Against whom? Check for missing rivals. Sonnet 5.5 and GPT-6.1 Sol do not appear on the card.
4. Under what budget? Compare time, tools and inputs. LVBench frame counts differ by model.
5. On how many items? Small sets make small gaps meaningless. The two GraphWalks rows use 650 and 200 problems, Chartography about 100 prompts and RiemannBench 25 problems. Chartography’s 0.6-point gap is smaller than one prompt’s weight.
6. What does the scale mean? F1 is not the share solved, and Harvey’s top score is 19.6%. A low number can describe a hard test, not a weak model.
7. Do outsiders agree on the order? Compare rankings, not raw points, as the next section shows.
8. Does it resemble your work? Pick the two or three rows closest to your tasks and ignore the rest.

Where do independent Gemini 4 Argon benchmarks agree with Google’s table?

Partly. Independent data confirms Argon is near the top but not clearly first. Artificial Analysis scores Argon 53 on its Intelligence Index, level with GPT-6 Astra and one point above GPT-6.1 Sol. Individual rows agree on ordering more often than on raw scores.

The Terminal-Bench Science gap is the clearest lesson. Google’s Argon figure is self-computed with an extended verifier timeout, while Vals reports strict pass or fail across 70 tasks in a mini-SWE-agent harness. That does not make either wrong. It shows that one benchmark name can hide different setups, which is why the order of models is the safer thing to compare.

The Vals Index also deserves a closer look. Vals weights its sectors by share of U.S. GDP: finance 8.0, coding 5.6, legal 1.2 and tax 0.5 out of 15.3. By our arithmetic, finance carries about 52% of the score and legal about 8%. A first place there says most about finance and coding agents.

Why does Argon show a 15% hallucination rate but only 50% accuracy?

The two figures answer different questions. Artificial Analysis defines hallucination rate as wrong answers as a share of everything that was not correct, including declined answers. Argon’s 15% rate with 50% accuracy means it often says it does not know. Astra scores 51% and 63%: it answers more and is right more often, but guesses wrong more.

Our arithmetic from those definitions: per 100 questions, Argon answers about 50 correctly, about 7 or 8 wrongly and declines or partly answers about 42. Astra answers about 63 correctly, about 19 wrongly and declines about 18. Artificial Analysis’s overall AA-Omniscience scores land almost level, 42 for Argon and 43 for Astra.

Which profile is better depends on cost of error. For a compliance FAQ assistant, where a wrong answer is expensive, declining is a feature. For a first draft of every item in a question bank, it means more gaps to fill by hand.

Why do tokens per task matter more than price per token?

A lower token price does not guarantee a cheaper task if the model uses more tokens. Artificial Analysis measured Argon at about 62,000 output tokens per task against about 27,000 for GPT-6 Astra, roughly 2.3 times as many. Its cost per task was $1.99 at the introductory price, $3.98 at standard price and $3.26 for Astra.

Artificial Analysis says the saving comes from lower token prices, not fewer tokens. When the introductory price ends, the per-task cost doubles to $3.98, about 1.2 times Astra’s. Our Gemini 4 Argon pricing guide has the dates and tiers. Remember too that Argon ran at “high” and the GPT-6 models at “max”, so these are not matched effort levels.

What do the benchmarks not show about Gemini 4 Argon?

They do not show how the model behaves on your tasks, at your default settings, or across a long working day. Google’s model card lists hallucinations, occasional slowness or timeouts and a March 2026 knowledge cutoff as limitations. The card also prints no confidence intervals, so small gaps cannot be judged from it.

Press reports add a dispute. Bloomberg, as relayed by The Next Web and Android Headlines, reported that some Google staff found Argon uneven at real-world coding and weak at front-end design, and two used the word “benchmaxxing”. Google said it would be inaccurate to say the model underperforms in areas such as coding. We cannot verify either side. The table is at least consistent with a mixed picture: Argon wins DeepSWE v1.1 and Vibe Code Bench but loses FrontierSWE v2 and Terminal-bench 4.0.

Is Gemini 4 Argon better than GPT-6 Astra?

It depends on the job. Argon beats GPT-6 Astra in 14 of 19 card rows and hallucinates less, but it ties Astra at 53 on the Artificial Analysis index, uses about 2.3 times the tokens per task and trails on FrontierSWE v2, Terminal-bench 4.0, Terminal-Bench Science 0.1 and OSWorld-2.0. No source we found shows a clear overall winner.

Use these as starting points, then confirm them in a pilot:

• Long documents and video: start with Argon, where its card leads are widest, then verify on your own files.
• Long terminal or desktop agent work: start with Astra or Claude Opus 5.5, because Argon trails on the terminal and computer-use rows.
• Questions where a wrong answer is costly: Argon’s lower hallucination rate helps, at the price of more declined answers.
• High-volume work: compare cost per task, not cost per token.

Our Gemini 4 Argon vs GPT-6 Astra vs Claude Opus 5.5 comparison covers the three models side by side. For the other half of the head-to-head, see our GPT-6 Astra benchmarks review and our guide to Claude Opus 5.5.

What should you do next?

Pick the two or three card rows closest to your work and run the checklist above on each. If a row is Google-run, mixed-source or small, treat it as a hint, not evidence.

Then decide how you will test. Our guide on how to access Gemini 4 Argon shows which plan or API route fits, and our guide on how to prompt Gemini 4 Argon helps you set up a fair pilot.

If your team builds learning content, our review of GPT-6 Sol and Claude Opus 5.5 for learning and development shows how to frame the same test for training work.