Access, pricing, and the 1M-token window

Google’s introductory API pricing runs through November 30, 2026. Fresh input costs $2 per million tokens, output costs $10 per million tokens, and cached input receives a 95% discount. Beginning December 1, 2026, fresh input rises to $4 per million tokens and output to $20.

The Gemini API provides a one-million-token unified window shared by the prompt and generated output. A larger prompt therefore leaves less capacity for the response. The window supports unusually large repositories, document collections, and multimodal inputs, but it does not remove the need to manage retrieval quality, latency, and token costs.

The launch information summarized here does not specify a stable API model identifier, default reasoning level, maximum single-response output, rate limits, latency targets, or data-residency terms. Production teams should confirm those details in the current API documentation before migrating workloads.

Benchmarks expose a split

Argon records its strongest results on long-context analysis and professional knowledge work. It scores 68.9% on the Vals Index, 51.29% on Zapier’s automation benchmark, and a reported 15% hallucination rate, the lowest among the high-scoring models in that evaluation.

Artificial Analysis gives Argon a score of 53 at its high-reasoning setting. GPT-6.1 Astra is roughly level at maximum reasoning, while Claude Opus 5.5 scores 58 and Sonnet 5.5 scores 56. Those settings may use different inference budgets, so the scores do not represent a controlled comparison of latency or cost.

DeepSWE rewards repository-level software work, while FrontierSWE and Terminal-Bench place more emphasis on agents operating through development tools and shells. Argon leads DeepSWE but trails Astra by 10.5 points on FrontierSWE and Opus by 9 points on Terminal-Bench. Vals also ranks Argon second to Sonnet 5.5 for app building and code migration.

Benchmark harnesses, tool permissions, reasoning budgets, and retry policies can materially change these results. Teams choosing a coding model should run representative tasks against their own repositories, tests, dependency systems, and review requirements.

Long jobs are Argon’s main bet

Argon combines its shared one-million-token window with a continuation protocol that Google says preserves task state across extended runs. The design targets jobs that require repeated planning, tool use, revision, and verification rather than a single generated response.

• Chromium refactor: Argon helped separate core webpage handling from a legacy folder containing roughly six million lines within Chromium’s 40-million-line codebase. Google says the team completed months of architectural work in weeks.
• Agent Development Kit: During an overnight run against the ADK repository, Argon reduced in-memory session preparation time by 40% for long sessions and improved cold-import speed by 21%.
• vLLM on Gemma: Google reports a 2.85-times end-to-end inference speedup on an Nvidia A100. Throughput for Gemma 4 E2B rose from about 37,900 to more than 108,000 tokens per second without a measured quality loss.
• Internal development: Antigravity teams using Argon reportedly increased agent-generated code submissions by more than 35% and reduced rollbacks by nearly one-third.

These case studies show how Google uses the model, but they do not establish expected performance across other repositories or hardware. The outcomes depend on task design, tool access, tests, review processes, and the amount of inference compute assigned to each run.

Security and science move beyond chat

Argon scores 68% on CWE-bench v1, tying for the highest reported result on repairing software vulnerabilities. Cloud security company Wiz used early access to connect vulnerabilities in healthcare infrastructure and identify a file-download bypass that exposed a master key. Approved defenders can apply through Fairwind for a version without the standard cyber guardrails.

Google also reports several scientific applications. Argon translated an AlphaQubit quantum error-correction decoder into an optimized hardware kernel that nearly doubled inference throughput. A Google team working with Professor Matt Landreman published Grad counterexamples concerning a 1967 conjecture about stellarator magnetic-field solutions in nuclear fusion.

Gamow Labs used Argon to identify a chromosomal rearrangement in a rare-disease genome that conventional software had missed for years. Google also cites work on novel enzyme design, although the launch material provides no experimental method, validation data, or success rate for that example.

Safety results come with limits

Google reports lower attack and shortcut rates across three internal or commissioned evaluations. The figures describe performance under specific test conditions and should not replace application-level threat modeling, access controls, sandboxing, logging, or human review.

Google also describes an automated, defense-in-depth monitor that inspects agent reasoning during execution for signs of misalignment. The company says it avoids training penalties aimed directly at the model’s intermediate reasoning so that suspicious behavior remains visible to the monitor. Independent validation and deployment-specific testing remain necessary, particularly for agents with network, shell, or credential access.

Where Argon fits

Google describes Argon as the first model in the broader Gemini 4 family, with faster and smaller variants expected to follow. Those models may offer more practical economics for high-volume applications. Until then, Argon’s clearest advantages are its long shared context, professional-task scores, and support for sustained agent runs, balanced against weaker terminal benchmarks and unspecified production limits.