Gemini 4 正式发布，我们终于有了一个写作强于代码的前沿模型Gemini 4 Officially Released: We Finally Have a Frontier Model Better at

After a seven-and-a-half-month hiatus, Google's flagship Gemini model has finally been updated.

Just moments ago, multiple Google-affiliated accounts simultaneously announced: Gemini 4 Argon is officially released.

And this may be the only recent frontier model whose writing, knowledge, and long-context capabilities significantly outshine its coding and agent abilities.

Time flies, and even Gemini 4's naming style has changed—at first glance, it even looks like it could be an OpenAI model. Google has not officially explained the origin of the name; all the author knows is that the internal codename from the testing period was carried over to the official release.

On a literal level, it represents argon, element 18, an inert gas, which forms an amusing parallel with Chrome (chromium).

No need to rush to the Gemini App to try it out—continuing the recently popular "security" narrative, Gemini 4 Argon is currently only available to a small number of cybersecurity partners through the Fairwind program. Google has not provided a full public release timeline, only stating that "when the time comes, it will first be made available to paid API users and Google AI Ultra subscribers," meaning regular subscribers will have to wait a while longer.

Earlier, model versions allegedly identified as Gemini 4 leaked from testing grounds, but since Google never confirmed them, nobody knows whether they were truly Gemini 4 or whether they contained Fable and Astra. Those so-called comparative test results are therefore not very reliable. For now, the author can only list the data officially released by Google (which also significantly reduces the workload of this article).

At the very least, judging from official benchmarks, Gemini 4 Argon is exceptionally powerful:

You read that correctly—among the 18 listed tests, Gemini 4 achieved top scores in 13 of them.

Its biggest lead appears in the knowledge work dimension. In the Vals Finance Agent v2 and Harvey's Legal Agent Benchmark tests, which involve real financial analysis and real lawyer work, it is no exaggeration to say that Gemini 4 leads any other model by two full tiers. This is not surprising—even at its worst, no one has ever doubted Gemini's world knowledge capabilities.

Another traditional Gemini strength is long context. In the GraphWalks test, which evaluates the ability to effectively use 256k–1M ultra-long contexts, Gemini 4 also leads other flagship models by more than 10%.

A new feature can explain this margin of dominance.

Gemini 4 Argon has raised its output limit from the previous generation's 64,000 tokens to 1 million tokens. In other words, as long as your wallet can handle it and Gemini is talkative enough, it can output roughly 700,000 to 1 million Chinese characters in one go—an epic upgrade for chain-of-thought construction. Google's official explanation is:

When a model has sufficient room to engage in deep thinking and generates hundreds of thousands of tokens within a single reasoning trajectory, it brings a whole new depth to reasoning, allowing it to tackle difficult problems in a single pass.

By comparison, Gemini 4's coding performance is somewhat inconsistent. On DeepSWE v1.1, the most commonly used real-world long-cycle software building test, Gemini 4 scored 77.9%, leading GPT-6 Astra and Claude Opus 5.5 by nearly 4 percentage points.

But on FrontierSWE v2, which requires building projects from scratch to tackle difficult problems, and Terminal-Bench 4.0, which involves operating Linux terminals for complex tasks, Gemini 4 ranks at the bottom.

The author would summarize Gemini 4's characteristics roughly as follows: knowledge/writing > coding > terminal.

Third-party tests also support this assessment:

In Arena.ai scores, which are determined by blind-review voting, Gemini 4 ranks first in Text Arena thanks to high scores in handling difficult prompts, instruction following, and creative writing; but in the more engineering-oriented Code Arena: WebDev and Agent Arena, Gemini 4 only manages 8th place.

In a sense, a Gemini with weaker coding ability is a blessing for the era. Over-optimizing agentic coding capabilities seems to have made many frontier models "forget how to speak human," from the excessive line-breaking that has persisted since the GPT-5.3 era to the mysterious jargon problems that began with Opus 4.7. At a time like this, a Gemini model that retains its literary sensibility precisely because of its average coding ability can handle many writing challenges.

Another noteworthy change is the significant drop in hallucination rate.

In Artificial Analysis testing, Gemini 4's hallucination rate is only 15%, the lowest among all frontier models. This means that when facing uncertain answers, it is more inclined to say "I don't know" rather than fabricate a response.

But for ordinary users, the most critical issue is of course pricing—and Google has always been especially generous in this regard.

Gemini 4 Argon's official API pricing is $2 per million input tokens and $10 per million output tokens, with cached hits priced at 5%, i.e., $0.10. The official announcement states that pricing will double after the introductory promotional period, but based on experience, the promo period will most likely not end before the next-generation model arrives.

At the promotional price, Gemini 4 costs one-fifth as much as GPT-6 Astra and Fable 5.1, and is roughly on par with GPT-6.1 Sol and Sonnet 5.5. In other words, as long as its capabilities are not falsely advertised, Gemini 4 is currently the model with the best "intelligence-to-price ratio"—flagship intelligence at mid-tier pricing.

The release of Gemini 4 Argon also puts an end to an awkward record: for nearly half a year, the Google AI Pro plan's benefit description read "includes access to the 3 Pro model"... Considering that since the 3.5 Flash era, even Flash models can outperform Pro, this membership plan became quite comical—to the point that many people subscribed solely for the 5TB Google Drive.

At least now, Google can finally promote its membership program with a clear conscience.

Writing this, I'd also like to share a few thoughts beyond the topic.

Everyone has their own "moonlight" (cherished favorite), and in the AI world, mine is Gemini 2.5 Pro. In early last year, the 2.5 Pro's writing ability was truly breathtaking, and it was precisely after that that I formally integrated AI into my workflow. Later, when I saw Gemini 3 released, I immediately shared it to my social feed.

In that sense, no one wants Gemini 4 to succeed more than I do.

This Gemini fan's fantasy

However, the author still respects objective laws. As many have commented, large models are, in the final analysis, a form of manufacturing—just like the manufacturing of smartphones and chips. The author cannot expect a model that has not been updated for a long time to suddenly outshine all its rivals; that would be as unrealistic as a chip factory stuck at 10nm for years suddenly producing a 2nm chip.

The enormous gap between Gemini 3.8 Flash's benchmark scores and real-world performance also suggests that Google may have engaged in some "special optimizations" to game benchmarks during this period in order to cover up its embarrassment. Similarly, Gemini 4 achieving SOTA on most benchmarks is rather suspicious:

Almost at the same time as Gemini 4's announcement, Bloomberg also reported that Google's internal employees are skeptical of Gemini 4's actual performance. Citing "insiders," their report stated that Gemini 4's performance on real-world tasks leaves much to be desired, and that it "struggles with certain coding tasks."

Although Google was quick to push back, and some employees have said Gemini 4 is quite good, the author can generally believe that the Gemini 4 that eventually reaches users will not be as strong as the benchmarks suggest. Moreover, by the time Gemini 4 is fully deployed, that will be enough time for ChatGPT and Claude to iterate another version.

Regardless, looking on the bright side, the release of Gemini 4 Argon does at least represent Google's return to the race.

Its ceiling may not be extremely high, but its floor is still guaranteed. What needs to be done next is to catch up step by step, the way large-model manufacturing should be done.