Gemini 4 Argon is GA: $2 in, $10 out, and Google's proof is its own agent fleet
Twelve days after Google announced it for cyber defenders only, Gemini 4 Argon is open to everyone who pays. As of 16:00 UTC on October 13 it is in the paid tier of the Gemini API, in Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode for Ultra and Pro. Introductory price is $2 per million input tokens and $10 per million output, with cached input at 95 percent off. The output ceiling went from 64K to 1M tokens, which Google frames as letting one trajectory think and write for hundreds of thousands of tokens without a handoff. For context, Meta priced Muse Wonder at $4/$20 the day before.
The benchmark list is the usual vendor pick: DeepSWE v1.1 at 77.9 percent, first on the Vals Index, 51.3 percent on Zapier's AutomationBench, 91.7 on LVBench, a tie for first on CWE-bench v1 at 68 percent, top of Gray Swan's indirect prompt injection board, 87 percent on the ERQA embodied reasoning benchmark. There is no SWE-Marathon number and no head to head with Opus 5.5 or Muse Wonder on a shared harness, so treat the ranking claims as unverified until Artificial Analysis and Arena post their runs.
What is more interesting than the table is how Google argues the case. The launch post reads like an agent-ops report. More than 100,000 people across Alphabet use Argon inside the internal Antigravity each month. Over four weeks, agentic code submissions rose 35 percent and post-submit rollbacks fell about 30 percent. The Chrome team used it to split core page handling out of a 6 million line legacy folder inside a 40 million line codebase, a refactor Google says went from months to weeks. Given the ADK team's .agents/skills rulebook and a one paragraph goal, it ran overnight and shipped optimizations that cut long-session prep time 40 percent. A bug queue of 134 open reports was triaged in under three hours, 97 closed with commit proof. Agents are migrating C and C++ to Rust up to the 800K line Fuchsia Zircon kernel, and on libgav1 they replaced 32K lines of SIMD with safe Rust that runs 2.7 times faster than the previous Rust port. The robotics post goes further and has a Boston Dynamics Spot practicing a foot-pedal trash bin in a MuJoCo twin before touching the real one, with the prompt literally asking for a two-agent verification swarm.
The rollback number is the one worth watching. Every lab now claims more code per day. Fewer reverts per merge is the first time a frontier launch led with a quality metric rather than a volume metric, and it is exactly the number nobody outside Google can check. Sundar Pichai's launch post passed 1.1 million impressions in a day and the HN thread hit 815 points, mostly people asking for the independent runs.
Links: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/ , https://antigravity.google/blog/gemini-4-argon-in-google-antigravity , https://developers.googleblog.com/gemini-4-argon-for-robotics/ , https://x.com/sundarpichai/status/2110038898009407584
← Back to all articles
The benchmark list is the usual vendor pick: DeepSWE v1.1 at 77.9 percent, first on the Vals Index, 51.3 percent on Zapier's AutomationBench, 91.7 on LVBench, a tie for first on CWE-bench v1 at 68 percent, top of Gray Swan's indirect prompt injection board, 87 percent on the ERQA embodied reasoning benchmark. There is no SWE-Marathon number and no head to head with Opus 5.5 or Muse Wonder on a shared harness, so treat the ranking claims as unverified until Artificial Analysis and Arena post their runs.
What is more interesting than the table is how Google argues the case. The launch post reads like an agent-ops report. More than 100,000 people across Alphabet use Argon inside the internal Antigravity each month. Over four weeks, agentic code submissions rose 35 percent and post-submit rollbacks fell about 30 percent. The Chrome team used it to split core page handling out of a 6 million line legacy folder inside a 40 million line codebase, a refactor Google says went from months to weeks. Given the ADK team's .agents/skills rulebook and a one paragraph goal, it ran overnight and shipped optimizations that cut long-session prep time 40 percent. A bug queue of 134 open reports was triaged in under three hours, 97 closed with commit proof. Agents are migrating C and C++ to Rust up to the 800K line Fuchsia Zircon kernel, and on libgav1 they replaced 32K lines of SIMD with safe Rust that runs 2.7 times faster than the previous Rust port. The robotics post goes further and has a Boston Dynamics Spot practicing a foot-pedal trash bin in a MuJoCo twin before touching the real one, with the prompt literally asking for a two-agent verification swarm.
The rollback number is the one worth watching. Every lab now claims more code per day. Fewer reverts per merge is the first time a frontier launch led with a quality metric rather than a volume metric, and it is exactly the number nobody outside Google can check. Sundar Pichai's launch post passed 1.1 million impressions in a day and the HN thread hit 815 points, mostly people asking for the independent runs.
Links: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/ , https://antigravity.google/blog/gemini-4-argon-in-google-antigravity , https://developers.googleblog.com/gemini-4-argon-for-robotics/ , https://x.com/sundarpichai/status/2110038898009407584
Comments