Google Is Back in the AI Race: What Gemini 4 Argon's Benchmarks Really Say
Gemini 4 Argon leads 13 of 18 benchmarks, ships a 1-million-token context and undercuts rivals on price. Here is where it wins, where it still loses, and what it changes.
On 13 October 2026, Google DeepMind made Gemini 4 Argon available worldwide: the first model of the Gemini 4 generation, and the first time in a while that Google has walked into a benchmark table and come out holding most of the top spots. It had been in limited release since 30 September, when a small group of vetted cybersecurity teams got early access. Now it is in the Gemini API, AI Studio, the Gemini app and Google's Antigravity coding environment.
The question everyone asks about a new frontier model is simple: is it actually better, or is it better at the benchmarks the vendor chose? Below is what Google published, what independent coverage adds, and where Argon still loses.
What Argon is
Google describes Argon as a large sparse mixture-of-experts transformer, natively multimodal, and explicitly not a fine-tune of an earlier Gemini. It was trained with heavy use of reinforcement learning, and its headline engineering change is a single 1 million token context window shared across input, thinking and output. Output alone can run to a million tokens, up from the 64,000-token ceiling of the previous generation. In plain terms: it can read a whole codebase, reason for a long time, and write a very long answer, all inside one call.
The knowledge cutoff is March 2026, and Google's own model card notes the usual limitations, including hallucination.
The numbers, side by side
Google's model card compares Argon against OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 across 18 benchmarks. Argon leads or ties on 13 of them. These are the ones Google highlights, with the scores exactly as published:
| Category | Benchmark | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Knowledge work | Vals Index | 68.9% | 63.1% | 67.0% |
| Knowledge work | AutomationBench | 51.3% | 41.4% | 42.5% |
| Knowledge work | Vals Finance Agent v2 | 65.4% | 53.5% | 58.6% |
| Knowledge work | Harvey's Legal Agent | 19.6% | 5.4% | 3.8% |
| Agentic coding | DeepSWE v1.1 | 77.9% | 74.1% | 74.2% |
| Agentic coding | Vibe Code Bench | 91.9% | 89.6% | 90.3% |
| Science / maths | LABBench 2 | 88.8% | 85.4% | 73.1% |
| Science / maths | RiemannBench | 76.0% | 72.0% | 69.6% |
| Long context | GraphWalks (256k–1M) | 84.2% | 71.8% | 66.8% |
| Multimodal | LVBench (long video) | 91.7% | 87.5% | 83.7% |
| Cybersecurity | CWE-bench v1 | 68.0% | 68.0% | 67.0% |
Where Argon clearly leads
Three results stand out. On Harvey's Legal Agent benchmark, a test of real legal work rather than trivia, Argon scores 19.6% against 5.4% and 3.8% for its rivals. That is a different tier, not a small lead. On GraphWalks, which tests reasoning over material spread across hundreds of thousands of tokens, Argon's 84.2% versus 71.8% and 66.8% is the clearest evidence that the unified million-token context is doing real work. And on AutomationBench, a measure of multi-step business tasks, Argon is roughly ten points ahead of both.
In agentic coding, the field Google most wants to win, the lead is real but narrow: 77.9% on DeepSWE v1.1 against 74.1% and 74.2%. Google also reports internal results, including replacing 32,000 lines of hand-written SIMD code in its libgav1 video decoder with a memory-safe version that runs 2.7 times faster.
Where it still trails
Google's own phrase is that Argon “does not need a clean sweep”, which is a candid way of saying it did not get one. On FrontierSWE v2, Argon scores 55.0% against GPT-6 Astra's 65.5%. On Terminal-Bench Science it trails Astra 57.6% to 68.1%, and on Terminal-bench 4.0 it trails Claude Opus 5.5 57.4% to 66.4%. On PostTrainBench, Opus 5.5 leads 49.3% to 45.3%. The pattern is consistent: the longer and more terminal-heavy the engineering task, the more the other two keep their edge.
So the honest summary is not “Google wins”. It is: Argon is the best model today for long-document, legal, finance and automation work, competitive at everyday coding, and still behind on the hardest hands-on software tasks.
Argon does not need a clean sweep. Model choice is still workload-dependent.
— Google DeepMind, on the Gemini 4 Argon launch
Price: the part that will move the market
Launch pricing is $2 per million input tokens and $10 per million output tokens, with cached input at $0.10. That is one-fifth of GPT-6 Astra's listed price ($10 / $50) and half of Claude Opus 5.5 ($4 / $20). Google has said the rate doubles to $4 / $20 after 30 November, which still leaves it at parity with Opus and well under Astra.
For anyone running agents at scale, that pricing matters more than a two-point benchmark lead. A model that is roughly as good and five times cheaper changes which workloads are economical to automate at all.
Is Google back in the race?
Yes, and for the first time since the Gemini 2.5 era it is back as a leader rather than a fast follower. Thirteen leads out of eighteen, a context window nobody else matches, and aggressive pricing make Argon the default choice for a large class of enterprise work. It is not the automatic choice for deep software engineering, where OpenAI and Anthropic still hold the hardest benchmarks.
Our own practice at Alokknight Technologies is unchanged by any single launch: we route each task to the model that measurably does it best and cheapest, and we re-run that comparison when a release like this lands. Argon just moved a lot of document, legal, finance and automation work onto Google's side of that table.