VisualForge Bench

Home › Videos › Gemini 4 Argon Tested: Almost Opus 5.5 for a Quarter of the Price

14 Oct 2026 · 4 runs · 4 models

Gemini 4 Argon Tested: Almost Opus 5.5 for a Quarter of the Price

Google opened Gemini 4 Argon to everyone, so I ran it through the Gemini API at high effort. On the four tests it finished it scored 7.22, third of 17 models, behind Opus 5.5 (7.45) and GPT-6 Astra (7.36). It wrote about half as many output tokens as Opus, so each test cost about $1.43 against Opus's $6.30. Two of its six tests did not finish: Google's stream cut off on Alex Optics, and the wildcard hit my API spending cap.

Watch on YouTubeCompare models from this videoEvery score (CSV)

The results

The same scores as the leaderboard: a test that failed is left out of its run's score, not counted as zero. Time, tokens and cost cover each whole run, wildcards included.

See the whole board

Ranking

Quality v1 over measured non-wildcard pages. Different workloads: browse the outputs and coverage before comparing models.

1
gpt-6-astraeffort: highOpenAI · 8/8 scored pages
7.5
2
claude-opus-5-5effort: highAnthropic · 8/8 scored pages
7.4
3
claude-sonnet-5-5effort: highAnthropic · 8/8 scored pages
7.3
4
gemini-4-argoneffort: highGoogle · 4/5 scored pages
7.2
0510

Score against output tokens

Quality v1 against full-run tokens. Different workloads and hardware affect this comparison.

Claude OpusClaude SonnetGemini ArgonGPT Astra
6.577.588.50500k1M1.5M2M2.5M Output tokens across all tests, thinking included Average score (out of 10) gpt-6-astra (OpenAI): 7.5 / 10, 322k output tokensgemini-4-argon (Google): 7.2 / 10, at least 443k output tokensclaude-opus-5-5 (Anthropic): 7.4 / 10, 1.30M output tokensclaude-sonnet-5-5 (Anthropic): 7.3 / 10, at least 1.99M output tokensgpt-6-astra (OpenAI): 7.5 / 10, 322k output tokensgpt-6-astra · highgemini-4-argon (Google): 7.2 / 10, at least 443k output tokensgemini-4-argon · highclaude-opus-5-5 (Anthropic): 7.4 / 10, 1.30M output tokensclaude-opus-5-5 · highclaude-sonnet-5-5 (Anthropic): 7.3 / 10, at least 1.99M output tokensclaude-sonnet-5-5 · high

Output tokens include the model's thinking. Like the time, tokens and cost table, it counts every test the model took, wildcards included. A hollow dot is not an exact count: some tests reported none, or the count was worked out from the length of the output. The score axis is zoomed to the data.

Score against time

Quality v1 against full-run time. Different workloads and hardware affect this comparison.

Claude OpusClaude SonnetGemini ArgonGPT Astra
6.577.588.501h2h3h4h5h6h Time taken across all tests, thinking included Average score (out of 10) gemini-4-argon (Google): 7.2 / 10, 2h 26mclaude-opus-5-5 (Anthropic): 7.4 / 10, 3h 29mgpt-6-astra (OpenAI): 7.5 / 10, 3h 31mclaude-sonnet-5-5 (Anthropic): 7.3 / 10, 4h 49mgemini-4-argon (Google): 7.2 / 10, 2h 26mgemini-4-argon · highclaude-opus-5-5 (Anthropic): 7.4 / 10, 3h 29mclaude-opus-5-5 · highgpt-6-astra (OpenAI): 7.5 / 10, 3h 31mgpt-6-astra · highclaude-sonnet-5-5 (Anthropic): 7.3 / 10, 4h 49mclaude-sonnet-5-5 · high

Time is how long the model took to answer, added up over its tests; judging is not counted. Like the time, tokens and cost table, it counts every test the model took, wildcards included. The score axis is zoomed to the data.

Where each model is strong and weak

One row per model, one column per test and per scoring criterion.

By testBy what the judges scored
ModelAero LabJet OdysseyBlast IslandLaunch DayAlex OpticsAfterburnerHangar OpsHangar PlannerRotor RampageWorking behaviourVisual qualityEngineering qualityUsability / accessibility
gpt-6-astraeffort: high7.87.67.07.17.87.8—7.37.88.28.26.76.3
claude-opus-5-5effort: high7.77.07.57.76.08.37.4—7.98.08.06.96.2
claude-sonnet-5-5effort: high7.8best of 27.84.17.7best of 28.47.7—7.27.77.68.06.96.1
gemini-4-argoneffort: high7.67.76.96.8failed————7.97.67.05.9
Violet is strong, orange is weak, on the same 0–10 scale in every cell.Different prompts and protocols: no best-in-column verdict.

Time, tokens and cost

What each model used across all its tests, wildcards included.

ModelTestsTimeOutput tokensSpeedCostSubscription used
gpt-6-astraeffort: highOpenAI 9 3h 31m 322k75k of it thinking 25.4 tok/s ≈$31.69API price, est. —
claude-opus-5-5effort: highAnthropic 8 3h 29m 1.30M753k of it thinking 103.0 tok/s $36.60API price 5-hour window not comparable7-day window +2 pts
claude-sonnet-5-5effort: highAnthropic 9 4h 49m ≥1.99M1.29M of it thinking 130.5 tok/s ≥$33.72API price —
gemini-4-argoneffort: highGoogle 62 failed 2h 26m ≥443k86k of it thinking 80.9 tok/s ≥≈$5.74API price, est. —
≈ An estimate: the provider reported no count, so it is worked out from the length of the output (about 4 characters a token). — Not measured. Cost says how it was worked out: charged by the provider, at API prices (for a subscription), or no API charge on local hardware. ≥ means some tests had no price.

Read from the account's usage window in whole percentage points, so a short run can show 0 points. It counts everything the account did in that time, not only this run.

Runs in this video

Every run behind the numbers. Scores name their comparison basis.

Model
Score / basis
Time
Output tokens
Cost
gpt-6-astraeffort: highOpenAI · 3 Oct 2026
7.5 / 10Quality v1 · 8/8 pages
Time3h 31m
Output tokens322k
Cost≈$31.69API price, est.
OpenCompare
claude-opus-5-5effort: highAnthropic · 3 Oct 2026
7.4 / 10Quality v1 · 8/8 pages
Time3h 29m
Output tokens1.30M
Cost$36.60API price
OpenCompare
claude-sonnet-5-5effort: highAnthropic · 5 Oct 2026
7.3 / 10Quality v1 · 8/8 pages
Time4h 49m
Output tokens≥1.99M
Cost≥$33.72API price
OpenCompare
gemini-4-argoneffort: highpartial quality coverageGoogle · 13 Oct 2026
7.2 / 10Quality v1 · 4/5 pages
Time2h 26m
Output tokens≥443k
Cost≥≈$5.74API price, est.
OpenCompare

The pages each model made

Choose a model to see its pages.

gpt-6-astraeffort: high

OpenAI · 7.5 / 10 · Quality v1 · 8/8 pages · 3h 31m · ≈$31.69

metrics.csvrun.jsonOpen the run pageCompare

Screenshot of the page gpt-6-astra made for Aero Lab7.8
Aero Lab gpt-6-astra · effort: high 24m 35s · 36k tokens · ≈$4.30 Quality v1 · Working behaviour 8.5 · Visual quality 8.5 · Engineering quality 7.0 · Usability / accessibility 6.5scored by claude-code/claude-opus-5-5 (quality-1 rubric)prompt 1e8e809e8caeb2ba
Earlier evaluations

6.0/10 · engineering-1 · claude-code/claude-opus-5-5 (engineering-1 rubric)

Different rubrics; these values are historical, not a score change.

Open the pageDetails
Screenshot of the page gpt-6-astra made for Jet Odyssey7.6
Jet Odyssey gpt-6-astra · effort: high 32m 29s · 48k tokens · ≈$6.16 Quality v1 · Working behaviour 8.0 · Visual quality 8.5 · Engineering quality 6.5 · Usability / accessibility 6.5scored by claude-code/claude-opus-5-5 (quality-1 rubric)prompt eaf6eace42027b11
Earlier evaluations

6.6/10 · engineering-1 · claude-code/claude-opus-5-5 (engineering-1 rubric)

Different rubrics; these values are historical, not a score change.

Open the pageDetails
Screenshot of the page gpt-6-astra made for Blast Island7.0
Blast Island gpt-6-astra · effort: high 23m 42s · 35k tokens · ≈$4.04 Quality v1 · Working behaviour 8.0 · Visual quality 8.0 · Engineering quality 5.5 · Usability / accessibility 5.5scored by claude-code/claude-opus-5-5 (quality-1 rubric)prompt 28ba22cd0f56ca1e
Earlier evaluations

6.7/10 · engineering-1 · claude-code/claude-opus-5-5 (engineering-1 rubric)

Different rubrics; these values are historical, not a score change.

Open the pageDetails
Screenshot of the page gpt-6-astra made for Launch Day7.1
Launch Day gpt-6-astra · effort: high 38m 12s · 56k tokens · ≈$7.01 Quality v1 · Working behaviour 8.0 · Visual quality 8.0 · Engineering quality 6.0 · Usability / accessibility 5.5scored by claude-code/claude-opus-5-5 (quality-1 rubric)prompt d1e76d48a7ac8eb6
Earlier evaluations

5.8/10 · engineering-1 · claude-code/claude-opus-5-5 (engineering-1 rubric)

Different rubrics; these values are historical, not a score change.

Open the pageDetails
Screenshot of the page gpt-6-astra made for Alex Optics7.8
Alex Optics gpt-6-astra · effort: high 23m 34s · 35k tokens · ≈$2.33 Quality v1 · Working behaviour 8.0 · Visual quality 8.3 · Engineering quality 7.4 · Usability / accessibility 7.0scored by claude-code/claude-opus-5-5 (quality-1 rubric)prompt de7c304e41dd3505
Earlier evaluations

7.5/10 · engineering-1 · claude-code/claude-opus-5-5 (engineering-1 rubric)

Different rubrics; these values are historical, not a score change.

Open the pageDetails
Screenshot of the page gpt-6-astra made for Afterburner7.8
Afterburner gpt-6-astra · effort: high 29m 39s · 34k tokens · ≈$2.39 Quality v1 · Working behaviour 8.5 · Visual quality 8.5 · Engineering quality 7.0 · Usability / accessibility 6.5scored by claude-code/claude-opus-5-5 (quality-1 rubric)prompt e08d0617005c33f9
Earlier evaluations

8.2/10 · engineering-1 · claude-code/claude-opus-5-5 (engineering-1 rubric)

Different rubrics; these values are historical, not a score change.

Open the pageDetails
Screenshot of the page gpt-6-astra made for Hangar Planner7.3
Hangar Planner gpt-6-astra · effort: high 8m 07s · 23k tokens · ≈$1.65 Quality v1 · Working behaviour 8.0 · Visual quality 7.0 · Engineering quality 8.0 · Usability / accessibility 6.0Fixed 16 of 16 bugs · 14 of 14 features still work · 545 lines changedscored by claude-code/claude-opus-5-5 (quality-1 rubric)prompt 14deca81c8b3a6c9
Earlier evaluations

9.0/10 · engineering-1 · claude-code/claude-opus-5-5 (engineering-1 rubric)

Different rubrics; these values are historical, not a score change.

Open the pageDetails
Screenshot of the page gpt-6-astra made for Rotor Rampage7.8
Rotor Rampage gpt-6-astra · effort: high 22m 30s · 33k tokens · ≈$2.28 Quality v1 · Working behaviour 8.5 · Visual quality 8.7 · Engineering quality 6.5 · Usability / accessibility 6.5scored by claude-code/claude-opus-5-5 (quality-1 rubric)prompt d196c18e5127f818
Earlier evaluations

7.2/10 · engineering-1 · claude-code/claude-opus-5-5 (engineering-1 rubric)

Different rubrics; these values are historical, not a score change.

Open the pageDetails
Screenshot of the page gpt-6-astra made for Unzip the Desert7.4
Unzip the Desertwildcard gpt-6-astra · effort: high 8m 53s · 24k tokens · ≈$1.55 Quality v1 · Working behaviour 8.0 · Visual quality 8.3 · Engineering quality 6.3 · Usability / accessibility 6.0scored by claude-code/claude-opus-5-5 (quality-1 rubric)prompt 8a3d9f5098d4145f
Earlier evaluations

6.9/10 · engineering-1 · claude-code/claude-opus-5-5 (engineering-1 rubric)

Different rubrics; these values are historical, not a score change.

Open the pageDetails