AI
Google launches Gemini 4 Argon with a one-million-token working window
Google says Gemini 4 Argon is globally available across its consumer, developer and enterprise products, with a one-million-token window shared between input and output. The model leads some long-context and professional-task evaluations but trails competing systems on specified coding-agent tests. Google reports case studies in Chromium refactoring, inference optimization, cybersecurity and science, while acknowledging that outcomes depend on tools, tests and inference budgets. API limits, latency targets and data-residency terms remain unspecified in the supplied launch material.
Editorial summary · Publisher article
What matters
- Introductory API pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after November 30.
- Argon leads the cited Vals and DeepSWE evaluations but trails other models on FrontierSWE and Terminal-Bench.
- Google reports a 2.85-times end-to-end inference improvement for Gemma on an A100 and faster internal development workflows.
- The model is also being offered through research and cybersecurity programs with eligibility and capacity conditions.
A long-context model with uneven benchmark leadership
Argon’s central product claim is sustained, multi-step work over very large inputs. The supplied comparisons show strength in long-context and professional knowledge tasks, but no universal lead: coding-agent results vary by benchmark, tool access, reasoning budget and retry policy.
Google’s deployment evidence
Google cites a Chromium refactor, a faster Gemma inference path, vulnerability work and scientific applications. These examples illustrate how Google and early partners use Argon, but they do not establish performance across unrelated repositories, hardware or operational settings.
Context & caveats
- The benchmark and case-study figures are reported by Google or its partners and were produced under particular evaluation conditions.
- The supplied material does not specify a stable API identifier, rate limits, latency targets, maximum response behavior or data-residency terms.
Why this made the edition
The report contains substantial launch details, comparative benchmark results, deployment caveats and concrete examples, while attributing company case studies.
Source material