Google launches Gemini 4 Argon worldwide, undercutting GPT-6 Astra and Claude Opus 5.5 on price

Logan Kilpatrick, Google's AI product lead, wrote in a social media post that this was only the start of the Gemini 4 era and that the model was designed to help people solve problems they previously would not have attempted. The name Argon refers to the inert gas used as a protective barrier in welding and semiconductor manufacturing, a stable presence rather than a spectacle, ifanr noted.

Google positions the model for long, multi-step, real-world work in coding, finance, law, cybersecurity and frontier science. Argon placed first on the Vals Index, which measures performance in key economic sectors, according to ifanr. The API offers a unified limit of up to 1 million input and output tokens with a new continuation protocol. Introductory pricing is $2 per million input tokens, $10 per million output tokens, and a 95 percent discount on cached input. Developers can reach the model through AI Studio, Android Studio and the Gemini API, while Google AI Pro and Ultra subscribers can use it in the Gemini app, AI Mode in Search and the Google Antigravity agent platform.

Google's own engineering teams used Argon as a production tool during the testing period, ifanr reported. Agent-generated code commits rose more than 35 percent through Antigravity while rollback rates fell by nearly a third. The model helped Chrome engineers refactor 6 million lines of legacy code inside a 40 million-line codebase, eliminating a class of browser crashes. Given only a short statement of goals and rules, it optimized Google's ADK agent framework overnight, cutting long-session memory preparation time by 40 percent, and raised the end-to-end inference speed of the open model Gemma 4 E2B by 2.85 times within days with no loss of quality.

In early evaluations cited by ifanr, Deloitte measured a 100 percent completion rate on long-horizon, cross-tool tasks; Kepler found Argon's multi-year financial statements matched human analyst models more than 99 percent of the time; and the legal technology company MinutaIA used it to surface case details human experts had missed. Google says the model accelerated quantum chemistry calculations and quantum error correction decoding, designed five new enzymes within hours, and helped find a new counterexample to a conjecture in nuclear fusion research. In an embodied-intelligence test, Argon watched people play a beanbag game and, within hours, derived joint trajectories, wrote simulation code and deployed it to an Apollo humanoid robot, which then made seven consecutive throws. Google also said it will give 10,000 scientists worldwide six months of free access to the top tier of Antigravity.

The company attached its tightest safety stack yet, including a real-time monitor on reasoning steps. Google says unauthorized shortcuts in coding tests fell to 2 percent and the success rate of indirect prompt-injection attacks to 0.7 percent. Working with the security firm Wiz, it also helped federal agencies and healthcare systems patch high-risk vulnerabilities in advance.

Independent testers reported wide speed and cost gaps against rival flagships. The team at AIHubmix ran the same prompt through the APIs of Gemini 4 Argon and GPT-6 Astra to generate a pixel-art night scene of an ancient building; GPT-6 Astra took 25 minutes and cost $11, while Argon finished in 3 minutes for $0.50 with comparable color and lighting detail, according to the report. Blogger Rubens Soto had both models produce a cinematic 3D animation explaining how a Tesla-style electric motor works; Argon's run cost $3.83 against $16.36 for Opus 5.5. Tester Pranav Reddy, using Three.js to simulate a 3D humanoid robot fight, found Argon ahead of Opus 5.5 on several measures of interface completeness and joint motion.

Image quality was not uniformly ahead. The testing firm AI/ML API compared 3D scenes of Godzilla fighting King Kong and noted that Argon cost $4.22 against $23.82 for GPT-6 Astra but still had clear room to improve in the rendering and form of complex creature models.

Access for unpaid users remains constrained. Under rules in effect with the launch, unpaid users are assigned models automatically, with most requests handled by the lightest Flash-Lite; turning off automatic selection locks them to Flash-Lite. The Gemini app has moved to a dynamic compute allowance that refreshes every five hours and is tied to a weekly cap, with higher thinking levels, Deep Research and calls to flagship models consuming quota faster.