> ## Content Index > Fetch the complete content index at: https://theainewsfiles.com/llms.txt > Use this file to discover other available public pages before exploring further. # Google's Gemini 4 Argon is out: where its own benchmark table has it ahead and behind - URL: https://theainewsfiles.com/gemini-4-argon-launch-benchmarks-google-model-card/ - Published: 2026-10-15T07:22:43.000Z - Updated: 2026-10-15T07:22:43.000Z - Description: Google made Gemini 4 Argon generally available on 13 October. Its own model card has it ahead on knowledge-work benchmarks and behind rivals on some coding and science tests. - Author: J E Cowie - Tags: News Google made Gemini 4 Argon generally available on 13 October 2026, two weeks after announcing it on 30 September and opening it first to cyber defenders in its Fairwind Program. The company's launch post, written by Koray Kavukcuoglu, describes a model that outperforms other frontier models on benchmarks across domains. The model card published the same day carries a benchmark table comparing Argon with rival models. The table is Google's own, and it is more mixed than that claim suggests. ## What Google says it shipped Google's launch post says that Gemini 4 Argon "is now globally available today for developers and enterprises via the API in AI Studio and Gemini platform, for Google AI Ultra subscribers in Google Antigravity, and Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search." The model card lists six distribution channels: the Gemini App, Gemini API, Google AI Studio, Gemini for business, Google Antigravity and Google AI Mode. Google DeepMind's model card describes Argon as a large-scale sparse mixture-of-experts transformer with native multimodal support for text, vision and audio inputs. It outputs text, with a combined input and output window of up to 1M tokens. The model card says it "is not a modification or a fine-tune of a prior model" and gives a knowledge cutoff of March 2026, adding that in some domains the model's knowledge may be less up to date. The launch post gives an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off. ## The benchmark table The model card compares Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 across a range of categories. These are Google's figures, reported "as of October, 2026". The card points to a separate methodology page for details. **Knowledge work.** Argon is ahead of all three rivals on all four listed benchmarks: - Vals Index: Argon 68.9%, against 63.1% (GPT-6 Astra), 65.8% (Claude Fable 5.1) and 67.0% (Claude Opus 5.5). - AutomationBench: 51.3%, against 41.4%, 31.4% and 42.5%. - Vals Finance Agent v2: 65.4%, against 53.5%, 58.9% and 58.6%. - Harvey's Legal Agent Benchmark: 19.6%, against 5.4%, 6.7% and 3.8%. The launch post says Argon ranks "#1 on the Vals Index". **Agentic coding.** Here the table splits. On DeepSWE v1.1 Argon scores 77.9%, against 74.1%, 67.4% and 74.2%. On Vibe Code Bench it scores 91.9%, against 89.6%, 90.3% and 90.3%. On the other two it does not lead: - FrontierSWE v2: Argon 55.0%, against 65.5% for GPT-6 Astra, 56.3% for Claude Fable 5.1 and 62.3% for Claude Opus 5.5. - Terminal-bench 4.0: Argon 57.4%, against 58.2%, 57.9% and 66.4% for Claude Opus 5.5, which is the highest score in that row. **Other categories.** - Science and math: Argon scores 88.8% on LABBench 2 and 76.0% on RiemannBench, ahead of all three rivals on both. On Terminal-Bench Science 0.1 it scores 57.6%, behind GPT-6 Astra (68.1%) and Claude Opus 5.5 (63.3%). - ML engineering: on PostTrainBench Argon scores 45.3% and Claude Opus 5.5 scores 49.3%. - Long context: on GraphWalks, Argon scores 99.7% up to 128k and 84.2% from 256k to 1M, ahead of all three rivals in both rows. - Computer use: on Agent's Last Exam Argon scores 39.5%, ahead of GPT-6 Astra (34.2%) and Claude Opus 5.5 (38.2%). On the offline subset of OSWorld-2.0 it scores 69.2% to GPT-6 Astra's 72.6%; the other two models have no figure listed. - Multimodal understanding: Argon scores 71.6% on Chartography and 91.7% on LVBench, ahead of all three rivals on both. - Cybersecurity: on CWE-bench v1 Argon and GPT-6 Astra both score 68.0%. Claude Opus 5.5 scores 67.0% and Claude Fable 5.1 scores 58.0%. The launch post says Argon's depth across multi-step tasks "leads Argon to outperform other frontier models on benchmarks across domains." My reading is that the table backs the launch post's claim most clearly for knowledge work, long context and multimodal understanding. It is a mixed picture for coding: Argon leads on DeepSWE v1.1 and Vibe Code Bench, but all three rivals score above it on FrontierSWE v2 and Terminal-bench 4.0, with GPT-6 Astra top on the first and Claude Opus 5.5 top on the second. ## Google's own usage claims The launch post also offers results from Google's own use of the model. Google says Argon in Antigravity helped its teams increase agentic code submissions by more than 35% while reducing code rollbacks by nearly a third. These are Google's internal figures. The post describes further projects, including a Chrome refactor, a Gemma inference speedup and a quantum chemistry package. It reports their outcomes in Google's own words, with no independent testing cited. ## Cybersecurity access and safety Google frames [Argon as a cyber-defence tool](https://theainewsfiles.com/gemini-4-argon-fairwind-bloomberg-skepticism/). The launch post says: "Starting today, we're beginning expansion to hundreds of additional Fairwind Program partners." It says Wiz has been using Argon for its "Scan for Good" initiative and describes two cases involving a federal agency's systems and healthcare infrastructure. The model card is direct about risk. For the cybersecurity critical capability level in Google's Frontier Safety Framework it says: "We cannot rule out that Gemini 4 Argon has reached this CCL". It adds that Google developed a safety case for the model's cyberoffensive capabilities that it assesses lowers risk to below that level. It gives the same "cannot rule out" wording for the biological critical capability level. On defences, the launch post says external bio red teaming "could only unlock less than 1% of bio sub-tasks we test for". It also says that on an adversarial set meant to stress-test the model, violations "reduced nearly 2-fold". ## What to take from it The release is a shipped product: Google's launch post says it is available now, and the model card lists the channels it is distributed in. The capability claims come from the company that built the model, and the benchmark table is Google's selection of tests and rivals. Read in full, that table shows a model that leads on knowledge work, long context and multimodal understanding, and does not lead on everything in coding. ## Sources Google DeepMind (2026) 'Gemini 4 Argon — Model Card', Google DeepMind, 13 October 2026, available at: [https://deepmind.google/models/model-cards/gemini-4-argon/](https://deepmind.google/models/model-cards/gemini-4-argon/?ref=theainewsfiles.com) (Accessed: 15 October 2026). Kavukcuoglu, K. (2026) 'Gemini 4 Argon: our next era of frontier intelligence', Google, 30 September 2026, available at: [https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/?ref=theainewsfiles.com) (Accessed: 15 October 2026). Kavukcuoglu, K. (2026) 'Gemini 4 Argon: Built to solve your hardest problems', Google, 13 October 2026, available at: [https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/?ref=theainewsfiles.com) (Accessed: 15 October 2026). ## Picture credits - Top picture: Google's headquarters. Image: Googleplex, Mountain View: Austin McKinley, [CC BY 3.0](https://creativecommons.org/licenses/by/3.0/?ref=theainewsfiles.com), via [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Google%5FCampus,%5FMountain%5FView,%5FCA.jpg?ref=theainewsfiles.com)