Gemini 4

Google Gemini 4 Matches GPT-6 Astra but Trails Anthropic’s Opus 5.5

co-produced by AI01. October 2026, 08:25

Gemini 4 Argon. © GPT-6

Google did not hold back on self-praise. Gemini 4 Argon, it said, marks “our next era of frontier intelligence,” is fundamentally changing how Google works and delivers frontier performance in coding, knowledge work and cybersecurity defense. The company’s first true flagship model in almost a year can generate up to one million tokens in a single response, and in Google’s own comparison table it came out on top alone in 13 of 19 benchmarks, ahead of Anthropic and OpenAI.

The independent benchmarking firm Artificial Analysis has now reached a far more sober verdict. On its Intelligence Index, Gemini 4 Argon scores 53 points. That puts Google’s new flagship level with OpenAI’s GPT-6 Astra, but clearly behind Anthropic’s Claude Opus 5.5. Bloomberg had already reported internal doubts: according to employees, the model performs worse in practical use, for example on certain coding tasks, than on benchmarks. For Google, the result is still a big step: according to Artificial Analysis, the company is once again one of the three leading A.I. labs.

Level With OpenAI, Five Points Behind Anthropic

Artificial Analysis’s Intelligence Index combines ten individual tests, from coding tasks in the terminal to scientific reasoning and agentic office work. Gemini 4 Argon scores 52.6 points at “High,” its highest available compute setting. GPT-6 Astra, at its highest setting, is practically level at 52.7 points, and the cheaper GPT-6.1 Sol reaches 52. Claude Opus 5.5 scores 57.6 points at its highest setting, staying about five points ahead. Anthropic’s smaller model Sonnet 5.5 also beats Argon with 56 points.

Compared with Google’s previous models, though, the leap is huge. Its last Pro model, Gemini 3.1 Pro, scored 30 points; Argon is 23 points higher. It also beats Gemini 3.8 Flash, Google’s best model until now, by 12 points.

The individual tests paint a mixed picture:

Agentic tasks: Argon makes big gains in this area, historically a weakness for Gemini models. On AutomationBench, which tests automated workflows in business software, it ranks first with 78 percent, seven points ahead of Claude Sonnet 5.5.
Terminal-Bench 4.0: On coding in the terminal, Argon reaches 57 percent, 53 points more than Gemini 3.1 Pro. Sonnet 5.5 (64 percent), Opus 5.5 (60 percent) and GPT-6 Astra (59 percent) stay ahead, however.
Knowledge work: On the AA-Briefcase test, Argon meets 65 percent of the grading criteria, more than any model before it. It falls behind on analytical quality and presentation, though, for an overall score of 1,494 Elo.
Hallucinations: Argon makes things up less often than its rivals. Its hallucination rate is 15 percent, the lowest among the leading models, compared with 51 percent for GPT-6 Astra and 54 percent for GPT-6.1 Sol. Argon is therefore more likely to admit when it does not know something. On raw accuracy, however, its 50 percent trails Astra by 13 points.

Terminal-Bench is notable: the independent result almost exactly matches Google’s own figure of 57.4 percent. Google’s comparison table, however, used the 66.4 percent that Anthropic reported for Opus 5.5 itself. Artificial Analysis measures only about 60 percent for Opus 5.5, which narrows the gap to Argon considerably.

Cheap Only Thanks to a Discount

On price, Google has a trick up its sleeve. Argon’s standard price is $4 (about 3.42 euros) per million input tokens and $20 (about 17.09 euros) per million output tokens, the same as Claude Opus 5.5. For the launch, however, Google is offering a 50 percent discount for at least a month, so Argon currently costs just $2 and $10 (about 1.71 and 8.55 euros). GPT-6 Astra costs $10 and $50 (about 8.55 and 42.74 euros). At the discounted price, Argon costs half as much per token as Opus 5.5 and only a fifth as much as GPT-6 Astra.

Even more telling is what a typical job actually costs, since models use different numbers of tokens depending on how hard they think. Artificial Analysis therefore calculates the cost of an average task from its Intelligence Index:

Gemini 4 Argon (High), discounted price: $1.99 (about 1.70 euros) per task
Gemini 4 Argon (High), standard price: $3.98 (about 3.40 euros)
GPT-6 Astra (Max): $3.26 (about 2.79 euros)
Claude Opus 5.5 (Max): $5.98 (about 5.11 euros)

As long as the discount applies, Argon costs only about 60 percent of GPT-6 Astra per task for comparable performance, and about a third of Opus 5.5. Once the discount ends, Argon will be about 22 percent more expensive per task than Astra, though still about a third cheaper than Opus 5.5. The reason: Argon is not frugal, only cheap per token. On average, it generates 62,000 output tokens per task, compared with 27,000 for GPT-6 Astra. Google has not yet said when the discount will end.

The Low-Cost Competition

Google is not alone in this price range. OpenAI’s GPT-6.1 Sol costs the same per token as Argon at its discounted price and trails by just one point. Per task, however, Sol is far cheaper at $0.72 (about 0.62 euros); even with the discount, Argon costs 2.7 times as much. Claude Sonnet 5.5 also has the same list prices as Argon’s discounted rate and scores higher at its top setting, but it needs so many tokens that a task costs $7.62.

Gemini 4 Argon is not broadly available yet in any case. Google is initially giving the model only to selected users, mainly cybersecurity teams, and plans to open it to paying API customers and Google AI Ultra subscribers later. According to Artificial Analysis, Argon understands images, video and speech as well as text and has a context window of one million tokens.