Gemini 4 Argon vs DeepSeek V4.1 Flash vs Muse Spark Open

Three launches, three strategies: benchmarks, price per million tokens and who holds the weights

Published 20 October 2026 Updated 20 October 2026

Gemini 4 Argon vs DeepSeek V4.1 Flash vs Muse Spark Open is the choice in front of anyone picking a model this month, and the three launches pull in three different directions. Independently, Google leads DeepSeek 53 to 39, yet DeepSeek charges many times less per token. So the thing that decides it is not the benchmark score, and it is not today's price either.

Three bets in one month

Google announced Gemini 4 Argon on September 30, then opened it to developers and to Google AI Pro and Ultra subscribers on October 13. It is a closed model: the only way in is Google's API and apps. DeepSeek took the opposite route, launching V4.1 Flash on September 10 and cutting its prices the same day.

Then on October 15, Meta released Muse Spark Open. Its Chief AI Officer, Alexandr Wang, called it the most capable US open-weight model to date. That is Meta's own description of its release, not a measurement by anyone outside the company. Inside the same few weeks, then, you get a closed subscription model, a cheap Chinese flash model, and an open heavyweight.

Google's flashiest proof was a robot. Argon watched one human video, wrote its own physics simulation, and taught Apptronik's Apollo humanoid to play cornhole, landing 7 consecutive throws after real-world refinement. That is a demo on Google's own stage, showing what Argon can do, not where it ranks against the other two.

Gemini 4 Argon vs DeepSeek V4.1 Flash: which one is smartest

Start with the numbers Google published itself. On Google's own benchmark table, Argon scores 77.9% on the DeepSWE coding benchmark and ties for first on CWE-Bench at 68% for fixing vulnerabilities. Those are the vendor's figures, published by the vendor, and a vendor tends to show the benchmarks it wins.

The independent check comes from Artificial Analysis, a tracker that compares models side by side. It rates Argon at 53 on its Intelligence Index against 39 for DeepSeek V4.1 Flash. On Terminal-Bench 4.0, which tests agent work in a terminal, the gap is wider: 57% versus 27%, more than double. If your workload is agents running commands, that is the difference you would feel.

Then there is the part the launch posts skip. The same tracker finds only three benchmarks where DeepSeek and Muse Spark can be compared directly, which is too few to call a winner. On that thin slice, Muse Spark 1.3 leads 69 to 55, but three rows do not crown anybody.

Meta's own materials claim Muse Spark Open beats open rivals such as Mistral Large 4 and Qwen3.8 on agentic tests like MCP-Atlas and JobBench, while admitting it trails some closed Chinese models on narrower coding tests such as Terminal-Bench 2.1. That is the whole picture: the only head-to-head with independent numbers behind it goes clearly to Argon over DeepSeek, and Muse has not been measured against Argon in the material behind this comparison.

What a million tokens costs

Argon is the smarter of the two measured models. The bill is the counterweight. Google charges an introductory $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input during the promotion, which matters a lot if your app resends the same long prompt over and over.

DeepSeek plays a different game. Artificial Analysis lists V4.1 Flash at roughly $0.30 per million input tokens and $1.20 per million output tokens: many times cheaper than Argon, on both sides of the meter. For an agent that chews through output all day, that is the difference between a rounding error and a line item.

But there is a turn on Google's side. The introductory price is a promotion that ends November 30, after which it rises to $4 in and $20 out, double what you see today. Budget on the launch price and the December invoice will surprise you.

DeepSeek has its own fine print. Its own price list is denominated in yuan, with off-peak rates that double at peak hours. And from September 14 it rerouted all traffic for the older V4 Pro model to the cheaper Flash pricing. Cheaper is good news; the catch is that the model behind your API call changed without you choosing it.

Muse Spark Open is the odd one out on price. The weights are free to download under Meta's Muse Model License, which Meta describes as carrying limited commercial terms that let many developers use it at no cost, so read those terms before you ship anything on it. But free weights are not free intelligence. Unless you host the model yourself, you pay a provider such as Groq, Fireworks AI, Together AI or OpenRouter for the inference. Which brings back the opening question. If it is not the score and not today's price, what is it? The answer is hiding in that word: yourself.

Can you run any of them?

Argon: no. It is fully closed, reachable only through Google's API and apps or a Google AI Pro or Ultra subscription. The material behind this comparison does not list what those plans cost.

Muse Spark Open is open, but open does not mean small. It is an 816-billion-parameter Mixture-of-Experts model with only 42 billion parameters active per token, and it takes text, images, audio and video. Apple's own machine learning team found it too big for a single Mac Studio. Their working setup chained four Mac Studios with M5 Ultra chips over Thunderbolt 5, and it required macOS 26.2 or later plus MLX v0.30.1 or later with "jaccl" communication support. The floor, in other words, is a stack of top-end Macs, not a laptop.

On that rig, Apple measured 36.73 tokens per second and a 4.8-second wait for the first token at a 65,536-token context. That is Apple's measurement, independent of Meta. Usable for one person; not a server for a whole team.

DeepSeek's weights are open too, published on Hugging Face under the MIT license, the most permissive of the three. The catch is that the model keeps changing shape. At launch it was described as a 552-billion-parameter Mixture-of-Experts model; the October 17 refresh on Hugging Face is listed at 763 billion parameters and now reads images as well as text. No hardware requirements for running it locally appear in the material behind this comparison, and at that size it is not safe to assume one machine will do. For most people DeepSeek means the cheap API, with the open weights as an exit door.

Who holds the weights

What decides this comparison is not the score and not today's price: it is who holds the weights. With Argon, Google holds them, so Google sets the price, and after November 30 that price doubles. With DeepSeek and Muse the weights are public, so if a provider raises prices or swaps the model underneath you, you can take your work somewhere else.

If you need the strongest model for coding or security work, take Argon, the clear winner over DeepSeek on independent tests, and budget for the post-promo price. If you run high volume and cost is your constraint, take DeepSeek V4.1 Flash instead of Argon: many times cheaper per token, with MIT weights behind it. If you want a US open-weight model that handles audio and video, and you have either a provider budget or a stack of Macs, take Muse Spark Open, after reading its license.

Don't take Muse if you expected a model that runs on a laptop. Don't take Argon if you cannot live with a price you do not control. And if one line survives this comparison, make it this one: the cheapest token is the one you can take somewhere else.

Sources