Every frontier launch now arrives with a table where the new model wins. Gemini 4 Argon is more interesting than most, because you can check Google's table against independent scoring, and the two tell different stories.
The headline version is that Google caught up. The accurate version is narrower: Argon leads on knowledge work, long context and price against one rival, and trails on the hardest coding. That is a different kind of win than "dethroned."
The Problem It's Solving
Google spent most of this year losing the coding-agent conversation. It cancelled Gemini 3.5 Pro, pushed a 3.8 Flash release, and then reset the naming with Gemini 4. Its output ceiling was also a real constraint: 64K tokens per response is short for migrations, long research runs and multi-file refactors.
Argon is Google's answer for long-horizon work. The pitch is a model that can think for hundreds of thousands of tokens in a single trajectory rather than being chunked into many short calls. A Google spokesperson told Engadget the company sees Argon as comparable to OpenAI's GPT-6 Astra and Anthropic's Opus on key benchmarks. "Comparable" is the claim to hold onto.
How Argon Actually Works
The headline spec change is output. The Gemini API docs list a 1,048,576-token context window, a default synchronous max output of 65,536 tokens, and up to 1,048,576 with long decode. So the 1M output figure is real, but it is an opt-in mode, not the default.
Inputs cover text, image, video, audio and PDF, but output is text only. There is no image generation, no audio generation and no Live API. Thinking comes in low, medium and high levels, and the Google Cloud guide notes that the old MINIMAL setting is unsupported. Caching is implicit only. Explicit caching isn't available, which matters if your cost model depends on it.
On safety, Google says it monitors the model's internal activations for misuse, watches chain-of-thought and actions for misalignment, and hardens its sandboxes before high-risk evaluations. It also released a cyber-permissive variant, gemini-4-argon-cyber-permissive, only to approved participants in the Fairwind Program.
What Developers Are Actually Using It For
Google's own examples are the most concrete. Argon agents analyzed fleet-wide profiling telemetry and found memory optimizations Google says free up over 300 TiB once rolled out. Argon is also driving C/C++ to Rust migrations, scaling from libraries like re2 and libgav1 up to 800K+ lines in the Fuchsia Zircon kernel. In libgav1, agents replaced 32K lines of SIMD code and produced a memory-safe decoder that runs 2.7x faster than the existing Rust port, per 9to5Google's rundown. Those migrations are still under audit and review before production, so treat them as in-progress, not shipped.
Externally, Argon is in Google Antigravity for /goal runs and multi-agent orchestration, in Google Cloud, and in GitHub Copilot. Google says agents working overnight under automated test guardrails cut in-memory session prep time by 40% in long sessions in google/adk-python.
Why This Is a Bigger Deal Than It Looks
Start with where Google's table is strong. Argon scores 77.9% on DeepSWE v1.1 against 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra. On Zapier's AutomationBench it hits 51.3% against 42.5% and 41.4%. On the Vals Index it reaches 68.9%, ahead of Opus 5.5 at 67.0% and Astra at 63.1%, according to this benchmark roundup. Its biggest leads are in work shaped like an actual job: finance, legal, long documents and long video.
Now the other half. Argon scores 55.0% on FrontierSWE v2 against 65.5% for Astra and 62.3% for Opus. On Terminal-Bench 4.0, Opus leads at 66.4% and Argon sits at 57.4%. On Artificial Analysis, the independent evaluator, Opus 5.5 at max effort scores about 58 while Argon and Astra both land at 53. The Harvey legal benchmark is a lead worth reading carefully: Argon's 19.6% is the best score, and it is still under 20%.
The price story is also less clean than it looks. Argon launches at $2 per million input tokens and $10 per million output, against Astra at $10 and $50. That introductory rate doubles to $4 and $20 later. In Sandbase's cost-per-task comparison, Argon at high effort scores 52.6 for $1.99 a task, while Opus 5.5 at high effort scores 53.6 for $1.82. Argon is clearly cheaper than Astra. Against Opus at comparable effort, the advantage largely disappears.
One number in Argon's favor deserves its caveat. Its hallucination rate on AA-Omniscience is 15% versus 51% for Astra, but its accuracy is lower, 50% versus 63%, because it declines more often. That is a real tradeoff: fewer confident wrong answers, more "I don't know." For legal and finance agents, many teams will take that deal. For coding agents that need to attempt something, maybe not.
So no one got dethroned. What changed is that "best model" now depends on the job. Opus leads on composite independent score and terminal work, Astra on the hardest engineering and computer use, and Argon on long-document knowledge work and cost against Astra. For anyone routing workloads across models, that is the more useful outcome than a single winner.
Availability and Access
Google announced Argon on September 30 and made it available to developers on October 13. It runs on the paid tier of the Gemini API and in AI Studio, with Antigravity access for Google AI Ultra subscribers and eligible Gemini Enterprise customers. It is excluded from the Gemini API free tier and, for the API, from the Google AI Pro plan, per The Rundown's access summary. Pro and Ultra subscribers get it in the Gemini app, and the cyber-permissive variant stays with Fairwind-approved projects.
Google did not take the crown, but it did end the idea that one lab owns every leaderboard. The next competitive question is who is best at the specific job you are paying for.
Follow for more coverage on MCP, agentic AI, and AI infrastructure.
