After being announced on September 30, Gemini 4 Argon is now widely available. The new model puts Google back in the conversation about the very best AI models. Unlike at the end of last year, the landscape is complex, and opinions are divided regarding this new challenger to Anthropic and OpenAI.
Argon is the first Gemini model since 3.1 Pro that hasn’t been released as an efficient “Flash” variant. Access was initially limited to participants in Google’s security-focused Fairwind Program. Anyone currently using the Gemini API, app, AI Mode in search results, or Antigravity for developers can now use Argon. The API pricing includes a temporary 50 percent discount through December 1, resulting in a rate of $2 per million input tokens and $10 per million output tokens.
Based on initial experiences, we’ve noticed that Google is back in the AI race, that Argon is a strong writer, but its coding style is divisive and it’s a bit slow to deliver.
At the top of various rankings
The Vals Index, one of many frequently cited AI benchmarks, ranks Gemini 4 Argon in first place. Interestingly, it takes the top spot not from Opus 5.5, but from the smaller Sonnet 5.5. Argon also leads in LMArena’s Text Arena blind taste test. The Artificial Analysis Index, which has recently been somewhat difficult to reconcile with the actual perceived quality of AI models, gives Argon the same score as GPT-6 Astra: 53. That’s slightly behind Anthropic’s Opus 5.5 (58) and Sonnet 5.5 (56).
Artificial Analysis also measures a hallucination rate of 15 percent, the lowest among models with an index score of 45 or higher. However, there is a caveat here. Argon answers only 50 percent of difficult factual questions correctly, compared to 63 percent for Astra. The model simply refrains from answering more often. On the other hand, according to several users on X, it dredges up rather obscure facts, undoubtedly a result of Google’s extensive database based on search indexing of the public internet.
Business testers see significant progress compared to Gemini 3.1 Pro, the previous flagship model. This is in line with expectations, given that a successor has been so long in coming. According to Box CEO Aaron Levie, Argon scored 14 percentage points higher on Box’s evaluation of complex work. Google also explicitly trained the model on creative writing. It also ranks first in LMArena’s creative writing category. Users report that, unlike many open-weight models, Argon does not exhibit Claude- or ChatGPT-like speech patterns, often a sign of
Opinions are divided on coding
When it comes to programming, opinions are most divided. Even in Google’s own comparison, Argon wins two of the four agentic coding benchmarks (DeepSWE v1.1 and Vibe Code Bench), while it finishes last on FrontierSWE v2 and Terminal-Bench 4.0. Typically, a model’s own report offers the most optimistic portrayal of its capabilities; in the real world, an AI model’s performance rarely exceeds that.
Independent measurements immediately show that this is once again the case. The aforementioned Vals also conducts an AI test, which ranks Argon second on Vibe Code Bench and Code Migration, in each case just behind Sonnet 5.5. Google’s table listed Argon as the winner on Vibe Code Bench, but Sonnet 5.5 had not been tested there. On LMArena’s WebDev ranking, the model finishes ninth; on the Arena agent ranking, it finishes fifth. CursorBench scores 52.8 percent, comparable to Opus 5.5.
Bloomberg reported on the day of the announcement that some Google employees found Argon to perform worse on real-world programming tasks and front-end design than the benchmarks suggest. Critics referred to “benchmaxxing,” or training specifically for the test. This was also the accusation leveled at poorly received models in the past, such as Llama 4. Google calls this characterization inaccurate and points to internal figures showing a 35 percent increase in agentic code pull requests and nearly a third fewer rollbacks.
Back in action
For Google, the launch itself is a sign of improvement. Sundar Pichai announced Gemini 3.5 Pro in May at Google I/O, but the model didn’t appear in June after all. According to SemiAnalysis, Google quietly scrapped it in August. Google did not confirm this. Comparisons with Fable and later Astra likely doomed this model.
The major advantage for Google now is that the AI labs have explicitly committed to developing their models at a slower pace. The gap that Google now appears to have largely closed is therefore unlikely to re-emerge anytime soon. Nevertheless, the internal LLMs from Anthropic and OpenAI are still presumably far more capable than what the outside world gets to experience.
OpenAI’s Astra 6.1, for example, is sitting on the shelf but has not been released due to concerns about the deceptive and dangerous tendencies it exhibited during evaluations. A highly capable AI model that lies about its own actions is certainly not what OpenAI needs after what feels like a never-ending series of security incidents involving “escaped” agent “swarms.”
Read also: Has OpenAI learned anything from the Hugging Face hack?