Google’s Gemini 4 Argon matches GPT-6 Astra in intelligence benchmark

Google DeepMind’s Gemini 4 Argon has reached the top tier of major AI models in a new independent benchmark, matching OpenAI’s GPT-6 Astra on overall intelligence. Artificial Analysis reports that Argon scores 53 out of 100 on the organization’s Intelligence Index when used with its highest reasoning setting.

The result places Google among the three leading AI labs in the ranking, according to Artificial Analysis. Gemini 4 Argon matches GPT-6 Astra and leads GPT-6.1 Sol by one point. It also marks a sharp improvement over Google’s earlier Gemini 3.1 Pro Preview, which scored 30.

Lower error rate, stronger agent performance

Argon’s results are helped by a comparatively low hallucination rate. Artificial Analysis measures hallucinations at 15 percent on its AA-Omniscience test, compared with 51 percent for GPT-6 Astra and 54 percent for GPT-6.1 Sol. The benchmark suggests Argon is more likely to say it cannot answer than to produce an unsupported answer.

However, Argon’s factual accuracy score reaches 50 percent, below GPT-6 Astra’s 63 percent. Its overall score on the test remains close to both competing models because of the lower hallucination rate.

The model also improves Google’s standing in AI agents, which can carry out multistep tasks. It leads the AutomationBench-AA evaluation with a 78 percent score and reaches 57 percent on Terminal Bench 4. That is a 53 point gain over Gemini 3.1 Pro Preview on the latter test.

Introductory pricing changes the comparison

Google currently charges $2 per million input tokens and $10 per million output tokens during a 50 percent launch promotion. Artificial Analysis estimates a cost of $1.99 per Intelligence Index task, about 60 percent of GPT-6 Astra’s estimated $3.26.

Once the discount ends, Argon’s estimated cost per task rises to $3.98. The model supports a one million token context window and accepts text, images, video, and speech as input. It is currently available only to selected users.