HomeAIAI ModelsGemini 4 Argon: Google's Benchmark Claims Meet Insider Coding Doubts

Gemini 4 Argon: Google’s Benchmark Claims Meet Insider Coding Doubts

On 30 September 2026, Google announced Gemini 4 Argon, which it calls “our next era of frontier intelligence” [1]. The announcement came with record benchmark scores, a far bigger output limit and a low starting price. But the first people to get it were a small group of cybersecurity defenders, not the public. And a few days later, Bloomberg reported that some Google staff, speaking anonymously, doubt it codes as well in real work as it does on tests [2]. Google disputes that.

So there are two stories here: what Google says Argon can do, and how much of that anyone outside Google could check.

Update, 12 November 2026: Google opened Gemini 4 Argon to the public on 13 October. Its release post says the model is available through the paid Gemini API, AI Studio and the Gemini platform, in Google Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode in Search for AI Ultra and Pro subscribers. The introductory price now ends on 30 November 2026. This article has been updated to reflect that; TSN covered the worldwide release, pricing and access separately.

What did Google announce?

The announcement was a blog post by Koray Kavukcuoglu, SVP of Google DeepMind and Google’s Chief AI Architect [1]. In it, Google says:

  • Argon is a new frontier model trained with a focus on defensive cybersecurity.
  • It’s rolling out first to trusted cyber defenders through Google’s Fairwind Program.
  • Wider access will come “as soon as possible,” starting with paid API customers and Google AI Ultra subscribers, after US voluntary pre-release engagement and more guardrail work.

Google didn’t give a date for that wider rollout in the announcement [1]. It came 13 days later, on 13 October [4].

There’s one detail worth slowing down on. Google says Fairwind defenders and its own internal teams get Argon without its cyber guardrails [1]. The idea is that people defending systems need a model that will dig into security flaws without refusing. It also means the most capable version is going to a small, vetted group first.

What does it cost?

These are Google’s announced API prices [1]:

  • Introductory price: $2 per million input tokens and $10 per million output tokens.
  • Cached input: 95% off the input price.
  • After the introductory period: $4 per million input and $20 per million output.

A token is a small chunk of text, often part of a word. The 30 September announcement gave no end date for the introductory price; Google’s 13 October release post and its Google Cloud announcement say it runs until 30 November 2026, with the standard price applying from 1 December [4][5].

For comparison, $2 in and $10 out is the same headline rate OpenAI lists for its new GPT-6.1 Sol model [3]. That’s a comparison of price lists, not of quality. Argon’s price is set to double once the introductory period ends.

What does Google say it can do?

Google’s post lists benchmark scores. All of these are Google’s own claims, and we didn’t open any third-party checks of them [1]:

  • DeepSWE v1.1: 77.9%, which Google calls state of the art
  • AutomationBench: 51.3%, which Google says ranks first
  • LVBench: 91.7%, which Google calls state of the art
  • CWE-bench v1: 68%, which Google says ties for first
  • Vals Index: Google says Argon is “leading”

Google also says Argon’s output limit has grown to 1 million tokens, up from 64,000 [1]. Put simply, the model can write a reply that’s book-length or longer in one go, rather than stopping after a long chapter.

The post also has internal success stories, which again are Google’s claims [1]. Google says Argon helped optimise a quantum computing subroutine that “beat published baseline by 40%,” found memory savings across Google’s computer fleet “freeing up over 300 TiB,” and helped move C and C++ code to Rust, including a rewrite of part of the libgav1 video decoder that Google says runs 2.7 times faster than an earlier Rust version. Google also names Wiz’s “Scan for Good” project as an early user and says it turned up a critical flaw in healthcare software.

Why are some people at Google skeptical?

On 4 October 2026, NWA Online ran a reprint of a Bloomberg story by Julia Love and Davey Alba [2]. Most of its critical claims come from people who asked not to be named. Here’s what they said, according to Bloomberg:

  • Gemini 4 has done well on benchmarks but does less well when employees put it to work, and struggles with certain coding tasks.
  • One person called its front-end design work uneven.
  • Two people said it seems affected by “benchmaxxing,” meaning it’s been tuned to score well on tests rather than to do real jobs well.
  • Google had planned to release Gemini 3.5 Pro in June, after announcing it at its I/O event in May, but dropped that effort.

The same report shows disagreement inside Google. Some employees think Anthropic’s Fable and OpenAI’s Astra are improving faster. Others think the coming version has caught up. One Google employee familiar with the work said there’s “large consensus” internally that Gemini 4 is at the frontier and denied that it struggles with messy real-world coding [2].

Google’s official answer, as reported: “Google said it would be inaccurate to say that Gemini 4 is underperforming in areas such as coding.” The company pointed Bloomberg to Kavukcuoglu’s earlier comments that he was encouraged by the model’s performance [2].

One more thing to be careful about: the report talks about “Gemini 4.” From the reprint, it isn’t clear which internal version the employees were describing, or whether it was exactly the Argon model announced on 30 September.

Why the stakes are high

A frontier model is a costly bet. A Bloomberg Intelligence analyst, Mandeep Singh, estimated that a single training run can cost as much as $400 million [2]. That’s an analyst’s estimate, not a Google figure.

Google also has a lot riding on Gemini across its products. According to Bloomberg, Google says its consumer Gemini app and AI Mode in Search have each passed 1 billion users [2]. That’s a company claim.

When this much rides on a model, benchmark scores turn into a marketing battle, and that’s why the gap between test results and day-to-day use matters.

Gemini 4 Argon’s public release moves the benchmark test to outsiders

When Argon was announced, almost nobody outside Google and its vetted cyber defenders could use it, so almost nobody could check Google’s benchmark claims. That has changed. Since 13 October, paying developers and AI Pro and Ultra subscribers have been able to put Argon to work on their own code, which is the test the anonymous staff quoted by Bloomberg said mattered most. Their complaint was not that the scores are false but that tuning a model to score well on tests can flatter it on messy, real-world jobs. Google denies that, and wide access is the only way the argument gets settled.

The evidence so far is still thin in both directions. The scores TSN reported remain Google’s own, the doubts still come from unnamed people, and the Bloomberg report referred to “Gemini 4” without making clear which internal version was meant. TSN also read a reprint rather than the original story. What developers can now see for themselves is the trade-off in Google’s launch plan: the strongest version, without cyber guardrails, stays with Fairwind defenders, while the public model ships with them. Google’s own safety report says it cannot rule out Argon reaching critical cyber and bio risk levels, which explains the caution.

Price is the other test. At $2 in and $10 out, Argon matched the headline rate of OpenAI’s GPT-6.1 Sol; from 1 December it will cost twice that. Teams evaluating it now, including through GitHub Copilot, should judge it at the price they will actually pay.

Sources

  1. Koray Kavukcuoglu, “Gemini 4 Argon: our next era of frontier intelligence,” Google blog, 30 September 2026. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
  2. Julia Love and Davey Alba, Bloomberg, reprinted as “Google grapples with Gemini skepticism,” NWA Online, 4 October 2026. Critical claims are from unnamed sources. https://www.nwaonline.com/news/2026/oct/04/google-grapples-with-gemini-skepticism/
  3. OpenAI API documentation, GPT-6.1 Sol model page (pricing, used only for the price comparison). https://developers.openai.com/api/docs/models/gpt-6.1-sol
  4. Koray Kavukcuoglu, “Gemini 4 Argon: Built to solve your hardest problems,” Google blog, 13 October 2026. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/
  5. Michael Gerstenhaber, “Gemini 4 Argon available on Google Cloud,” Google Cloud blog, 13 October 2026. https://cloud.google.com/blog/products/ai-machine-learning/gemini-4-argon-available-on-google-cloud

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

More in this category

Latest stories

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.