---
title: "Gemini 4 Argon Is Now Public: Benchmarks, Cost, and the Real Test for Coding Agents"
description: "Gemini 4 Argon is now available to developers, enterprises, and Ultra and Pro subscribers. It scores 77.9% on DeepSWE, but the useful question is whether it delivers verified changes in your repository."
canonical: "https://zainahmed.net/blog/gemini-4-argon-long-horizon-engineering-agents"
published: "2026-10-13"
---
# Gemini 4 Argon Is Now Public: Benchmarks, Cost, and the Real Test for Coding Agents
> Author: Zain Ahmed
> Published: 2026-10-13
> Read Time: 10 min read
> Category: AI & Tech
> Tags: Gemini 4 Argon, Coding Agents, Software Engineering, AI Security, Google DeepMind, Antigravity
> Canonical URL: https://zainahmed.net/blog/gemini-4-argon-long-horizon-engineering-agents
## Excerpt
Gemini 4 Argon is now available to developers, enterprises, and Ultra and Pro subscribers. It scores 77.9% on DeepSWE, but the useful question is whether it delivers verified changes in your repository.
---
Gemini 4 Argon is now open beyond trusted testers. On October 13, Google made it available through the [Gemini API and AI Studio](https://ai.google.dev/aistudio), to Google AI Ultra subscribers in Antigravity, and to Ultra and Pro subscribers in the Gemini app and AI Mode in Search. [Google's launch post has the full rollout]().
That changes what this article is for. Last time I could only list what to test once access opened. Now you can run the test. The published evidence still sets the bar: Argon scores 77.9% on DeepSWE, and Google's own examples show agents that measure, edit, and check their work. I have not run it in my own codebase yet, so everything below is Google's or a third party's claim.
Gemini 4 Argon is here.
Sundar Pichai (@sundarpichai) · October 13, 2026
Pichai’s announcement confirms the public rollout through the paid Gemini API, Antigravity for Ultra subscribers, and the Gemini app and Search for Ultra and Pro subscribers.
View original post on X
## What changed at public launch
Three things are new since the preview.
**Access and price.** Paid API users get Argon at an introductory **$2 per million input tokens and $10 per million output tokens**. After **November 30, 2026**, the price becomes **$4 and $20**. That deadline matters if you plan a long evaluation. [Google gives the dates here]().
Argon is a huge step forward for Gemini, and a labor of love.
Logan Kilpatrick (@OfficialLoganK) · October 13, 2026
Kilpatrick’s launch post covers access through the Gemini API, Gemini app, and Antigravity, and describes testing by more than 100,000 people at Google before release.
View original post on X
**Production examples.** Google says Argon in Antigravity helped its teams raise agentic code submissions by more than 35% while cutting code rollbacks by nearly a third. It names specific work:
- **Chrome:** separating core webpage handling from a roughly 6 million line folder of browser features inside Chromium's 40 million lines, across hundreds of files.
- **Agent Development Kit:** an overnight run that cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%. Google links the change [in google/adk-python](https://github.com/google/adk-python/commit/1754b3eeb6cdb2c16e3b30619a12066623df132e).
- **xTB:** up to 2.3 times faster molecular structure prediction, with the changes contributed to [the open-source repository](https://github.com/grimme-lab/xtb/pulls?q=is%3Apr+author%3Ag8j39+%28created%3A%3E2026-10-05+created%3A%3C2026-10-10%29).
- **Gemma 4 E2B on vLLM:** a 2.85 times end-to-end speedup on an A100 with 100% token parity, with [upstream pull requests](https://github.com/vllm-project/vllm/pulls?q=is%3Apr+author%3Alucianommartins+%28created%3A%3E2026-10-05+created%3A%3C2026-10-10%29) you can read.
The linked pull requests are the part worth your time. A reviewer can inspect a diff. A percentage in a blog post cannot be inspected.
**A different token limit.** My earlier draft treated one million tokens as an output-only limit. Google's launch post now describes **a 1M unified input-output token limit** in the Gemini API, plus a continuation protocol for long runs. Read [the long-decode documentation](https://aistudio.google.com/docs/long-decode) before you size a job.
## A bigger token limit gives an agent room to work
Google describes Argon's limit as **1M unified input and output tokens** through the Gemini API, with a [continuation protocol](https://aistudio.google.com/docs/long-decode) for long responses. Earlier coverage, including [Artificial Analysis](https://artificialanalysis.ai/articles/gemini-4-argon-google-top-three-labs), discussed the context window and output ceiling separately. Treat one number as a budget shared across everything the request reads and writes until you test your own workload.
For a coding agent, the extra output matters when work takes several passes. Imagine a contained C++ to Rust migration. The agent needs to choose a module boundary, translate a small part, compile it, compare behavior, read the failures, and revise the patch. A run that stops halfway through that sequence leaves the engineer to reconstruct what happened.
The continuation protocol keeps a long trajectory running across follow-up calls. It does not tell us whether the agent chose the right test, interpreted a failure correctly, or knew when to abandon a bad approach.
That extra room helps when a failed check leads to a better next experiment. If the agent keeps editing without learning from the result, the larger limit simply lets a bad run last longer.
## The benchmarks show strength, with meaningful gaps
In [Google DeepMind's comparison table](https://deepmind.google/models/gemini/), Argon scores **77.9% on DeepSWE v1.1**, ahead of GPT-6 Astra at 74.1% and Claude Opus 5.5 at 74.2%. It also reaches **51.3% on AutomationBench**, compared with 41.4% for Astra and 42.5% for Opus. These are good results for a model presented as a long-running agent.
Built to solve your hardest problems from start to finish
Google DeepMind (@GoogleDeepMind) · October 13, 2026
DeepMind positions Argon for coding, enterprise workflows, and cyber defense. The benchmark comparisons below examine the evidence behind that positioning.
View original post on X
*DeepSWE v1.1 chart: [Google DeepMind / Google Blog](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/). [Google's evaluation methodology](https://deepmind.google/models/evals-methodology/gemini-4-argon) explains the different harnesses and reporting sources.*
The rest of the table is less tidy:
| Evaluation | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
| --- | ---: | ---: | ---: |
| DeepSWE v1.1, agentic coding | **77.9%** | 74.1% | 74.2% |
| FrontierSWE v2, agentic coding | 55.0% | **65.5%** | 62.3% |
| Terminal-bench 4.0, agentic coding | 57.4% | 58.2% | **66.4%** |
| AutomationBench, knowledge work | **51.3%** | 41.4% | 42.5% |
| GraphWalks, 256K to 1M context, F1 | **84.2%** | 71.8% | 66.8% |
| CWE-bench v1, vulnerability remediation | **68.0%** | **68.0%** | 67.0% |
*Source: [Google DeepMind's published results](https://deepmind.google/models/gemini/). Its [methodology document](https://deepmind.google/models/evals-methodology/gemini-4-argon) describes the setups; these rows are not one controlled comparison using the same agent harness.*
Argon's lead on DeepSWE is meaningful within that evaluation. On FrontierSWE v2 and Terminal-bench 4.0, Astra or Opus leads. I would not turn those differences into a story about Argon's architecture. The tasks, harnesses, tool budgets, and reporting sources vary. Google's notes say its Argon results are generally single-attempt runs at the highest thinking setting, while many competitor figures come from providers or public leaderboards. [The methodology](https://deepmind.google/models/evals-methodology/gemini-4-argon) belongs next to any head-to-head claim.
### Which other scores matter to a software team?
Argon scores **84.2% F1 on GraphWalks** across problems spanning 256,000 to one million input tokens. That is evidence on a structured long-context task. It is not a test of whether an agent will preserve a repository's design constraints after dozens of tool calls.
Google also reports **68.9% on the Vals Index**, **65.4% on Vals Finance Agent v2**, and **19.6% on Harvey's Legal Agent Benchmark**. The legal result leads Google's table, though the absolute score leaves considerable room for improvement. For teams whose agents work from video or charts, Argon reaches **91.7% on LVBench** and **71.6% on Chartography**. Those numbers are useful when the corresponding inputs appear in your workflow; otherwise they are breadth, not a reason to change coding tools. [Google publishes the full set here](https://deepmind.google/models/gemini/).
An outside measurement adds another constraint. [Artificial Analysis gives Argon a 53 on its Intelligence Index](https://artificialanalysis.ai/articles/gemini-4-argon-google-top-three-labs), level with Astra in that test. On AA-Omniscience it reports **50% accuracy** and a **15% hallucination rate**. The lower hallucination rate appears to come partly from declining to answer more often. I would not carry that figure over to code review or security work as a general reliability guarantee.
## Google's engineering examples show the work behind the score
The decoder project is the clearest case study. Google says Argon agents took an existing Rust port of **libgav1**, its AV1 decoder, and replaced **32,000 lines of SIMD code** through profile-guided experiments. They examined compiler output and wrote safe Rust that the compiler could vectorize. Google reports identical video output and the **2.7-times speedup against the earlier Rust port**. It says the result moved closer to optimized C++; it does not claim the new Rust version beats C++. [Read Google's account](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/).
That account has the parts I would want to inspect in an agent run: a baseline, repeated measurement, attention to generated machine code, and an output check. It is still Google's internal report, with no public reproduction in the launch post. The useful claim is narrower than "the agent wrote better code": it reportedly improved a specific implementation under a specific comparison.
Google says Argon agents are also working on C and C++ to Rust migrations involving **re2**, **libgav1**, and more than **800,000 lines** in the Fuchsia Zircon kernel. These critical rewrites are still undergoing automated and manual auditing, emulation testing, and review before production. A patch and a deployed migration are different milestones. [Google states the review status here](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/).
The company reports two other outcomes. Data-center memory optimizations are expected to free more than **300 TiB once rolled out**, with **500 TiB to 1 PiB** in estimated total savings. In quantum research, Google says Argon improved the published spacetime baseline for one subroutine by **40% in minutes**. These are company-reported results, and the public post does not give enough detail to reproduce either one independently. The memory figure is a projected rollout benefit, not a measured saving already delivered across the fleet. [Both claims appear in Google's launch post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/).
## A longer agent run needs tighter boundaries
Every extra step gives a coding agent another chance to read untrusted text from a repository, log, web page, or tool result. If that text contains instructions disguised as data, the agent may follow them instead of the user's task.
Google calls Argon its most resilient model so far against indirect prompt injection. It reports a **0.7% attack success rate at 15 attempts** on the Gray Swan IPI evaluation shown for its [Fairwind Program](https://deepmind.google/fairwind-program/). That is a result under the benchmark's conditions, not a guarantee for an agent with access to your code, credentials, and network.
*Gray Swan IPI results as presented by [Google DeepMind / Google Blog](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/). Lower is better; the figures describe this benchmark's attack conditions.*
Google describes safeguards against harmful misuse, adversarial training against prompt injection, monitoring of model actions and reasoning, and more isolated sandboxes for high-risk evaluations. It says monitors can stop execution when necessary. Findings from training-run monitors went to an incident-response team rather than straight back into training, to reduce the chance of teaching the model to evade them. Google also says shortcut attempts on coding benchmarks fell to **2%**, the lowest among its Gemini models. It describes training Argon to report negative findings honestly and to pause before irreversible steps. These are [Google's stated controls](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/), not an independent assessment of how they hold up in deployment.
Argon also scores **68% on CWE-bench v1**, tied with Astra in Google's comparison. Google says Wiz used Argon through Scan for Good to identify a critical healthcare software vulnerability that exposed sensitive information. The launch post does not establish that it was a previously unknown zero-day, so I would not describe it that way. [Google gives the cybersecurity details here](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/).
*CWE-bench v1 chart: [Google DeepMind / Google Blog](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/). The published leaderboard breaks equal pass-at-one scores using pass-at-four results.*
Trusted defenders in [Fairwind](https://deepmind.google/fairwind-program/) receive controlled access for authorized threat simulation, reverse engineering, and defensive research under identity, access, and oversight requirements. The corresponding question for a software team is concrete: which files may the agent read, which commands may it run, which networks may it reach, and where must a person approve the next step?
## Count the whole run, not just the token price
Google's introductory Argon price is **$2 per million input tokens and $10 per million output tokens**, with cached input priced 95% below the input rate. The introductory period ends **November 30, 2026**, when the rates become **$4 and $20**. [The launch post sets out those prices](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/).
Long trajectories can change the bill even when the unit price looks attractive. Artificial Analysis reports an average of **62,000 output tokens per Argon task** in its Intelligence Index suite, versus **27,000 for Astra**. It estimates **$1.99 per Argon task** at the introductory rate and **$3.98** at the later rate. Those are measurements and estimates for its workload. Your repository may require different tool calls, retries, and review time. [Artificial Analysis details its cost calculation](https://artificialanalysis.ai/articles/gemini-4-argon-google-top-three-labs).
Availability belongs in that calculation too. The public model is on the paid API tier, and Antigravity access requires Google AI Ultra. Google separately offers eligible scientists six months of its highest Antigravity tier, and cybersecurity defenders can apply to the [Fairwind Program](https://deepmind.google/fairwind-program/) for access without cyber guardrails. [See Google's rollout statement]().
## The test I would run in a real repository
I would give Argon one contained migration or performance repair with an acceptance test that can fail for the right reasons. Before the run, I would record the current behavior, the performance baseline if relevant, the files it may change, and a fixed time and token budget. A disposable branch and limited credentials make the result easier to inspect.
During the run, I would watch for the behavior Google's decoder story implies. Did the agent find the right boundary? Did it run a meaningful test after changing code? When a check failed, did it use that evidence to revise the patch, or did it keep editing around the failure? A green final build is less convincing if a reviewer has to reconstruct the investigation from scratch.
I would stop the run at the budget or permission boundary, then compare its verified result with another model or the team's current workflow on the same task. The scorecard should include elapsed time, input and output tokens, failed attempts, tool use, review time, and whether the final change meets the acceptance criteria. That is the discipline behind [verifying AI changes in a TypeScript app](/blog/essential-ai-coding-feedback-loops-for-typescript): decide what could reveal a wrong result before you celebrate a passing check.
That leaves a question I can test in a real codebase: does a longer run produce more verified work within the same budget? I want to see how much of the benchmark survives contact with messy production software.
## Primary documentation and research links
- **LinkedIn Announcement:** [Google DeepMind on LinkedIn](https://lnkd.in/p/g_EUcZfH)
- **Official Google Launch Post:** [Introducing Gemini 4 Argon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/?utm_source=x&utm_medium=social&utm_campaign=&utm_content=)
- **DeepMind Model Architecture:** [Google DeepMind Gemini Models](https://deepmind.google/models/gemini/)
## Interactive article data
```json
{
"sourceReport": {
"title": "Gemini 4 Argon: Model evaluation",
"publisher": "Google DeepMind",
"url": "https://storage.googleapis.com/deepmind-media/gemini/gemini_4_argon_model_evaluation.pdf",
"pages": 5,
"description": "Read the benchmark results, test conditions, and reporting sources behind the comparisons in this article."
}
}
```
---
## Related Endpoints & Navigation
- **All Articles**: [https://zainahmed.net/blog](https://zainahmed.net/blog)
- **Author Profile**: [https://zainahmed.net/about](https://zainahmed.net/about)
- **Machine-Readable LLM Manifest**: [https://zainahmed.net/llms.txt](https://zainahmed.net/llms.txt)
- **RSS Feed**: [https://zainahmed.net/feed.xml](https://zainahmed.net/feed.xml)