Google releases Gemini 4 Argon, its new frontier model, with 1-million-token context and focus on cybersecurity, software engineering, and legal work

Google announced Gemini 4 Argon on September 30, 2026, and made it broadly available on October 13–14, 2026. The model represents Google DeepMind's next frontier-capability release, arriving eight weeks after Mistral Large 4 and five months after the Anthropic Opus 5.5 and xAI Grok models. Argon is accessible via the Gemini API, Google AI Studio, and Google Cloud's AI infrastructure, initially to paid tier customers and Google AI Pro/Ultra subscribers. Google is also providing pre-release access to trusted cyber defenders through its Fairwind Program and working with US government agencies on government-specific access. The model is technically a successor to Gemini 3.5 Sonnet, which handled most production inference until Argon's release.

The defining technical change is the context window: Argon supports 1 million tokens of input, output, and thinking combined, compared to 64K on prior versions. The increase matters for two use cases. First, it allows developers to load entire codebases, legal documents, or financial reports at once, eliminating the need to chunk long contexts into sequential API calls. Second, Google introduced thinking-tokens during inference, allowing the model to perform hidden reasoning steps before generating an answer—a capability borrowed from OpenAI's o1-mini. Google does not publish the proportion of its token budget allocated to thinking versus output, but notes that users can control the thinking depth trade-off for latency-critical applications. These two changes (massive context plus thinking time) position Argon for the "AI-as-solver" tier—autonomous agents that work on complex, multi-day problems with access to full context.

Google published benchmarks emphasizing specialized, high-value domains: software engineering, long-form understanding, and business process automation. On DeepSWE v1.1 (autonomous code bug fixing), Argon scored 77.9%, which Google calls state-of-the-art. On LVBench (understanding video content over minutes-long clips), it achieved 91.7%, also state-of-the-art per Google's claim. On AutomationBench (Zapier's end-to-end business-automation benchmark), Argon ranked #1 with 51.3%. On CWE-bench v1 (security vulnerability patching), it tied for first at 68%. On Vals Index (a composite of finance, coding, legal, and tax work), Argon ranks highest. On Gray Swan's Indirect Prompt Injection benchmark, Argon reduced attack success rates to 0.7% across 15 adaptive attempts, Google reports. These are preliminary figures and all benchmarks are either Google's own internal measurements or single-source vendor reports—independent evaluation has not yet appeared.

Google's framing of Argon is tightly scoped: cyber defense, software engineering, and enterprise knowledge work (legal and finance research). The company is not claiming general superiority across all tasks. In fact, Google notes internal data suggesting Argon performs best when given explicit permission to think longer, use code execution, and iterate autonomously—making it well-suited for asynchronous, deadline-flexible tasks (code review, vulnerability analysis, long-form document research) but not for latency-sensitive endpoints. This positioning resembles OpenAI's framing of o1-mini: a reasoning-first model for hard problems, not a replacement for fast general-purpose models. Google's Gemini 3.5 Sonnet will continue to handle most user-facing production inference. Argon is the specialist tier.

The Fairwind Program and guardrail strategy signal Google's approach to frontier-model risk. Argon will initially be available without the content safety guardrails that constrain other Google models—meaning it can generate code for defensive cybersecurity purposes that a guardrail-constrained model would refuse. Google is betting that trusted defenders will use this capability responsibly and will report findings. In parallel, Google is deploying multiple layers of safeguard: monitoring of chain-of-thought reasoning, sandboxed execution for code generation, prompt-injection detection, and CBRN (chemical, biological, radiological, nuclear) attack mitigations. The model will later ship to broader audiences with guardrails enabled, Google says.

Pricing and cost-competitiveness matter for frontier-model adoption. Argon's introductory rate ($2/$10 input/output) is half the standard rate ($4/$20), making it cheaper than Anthropic's Claude Opus 5.5 during the promotional window. Cached tokens—a feature where frequently re-used inputs are stored and charged at 95% discount—become valuable for long-context workloads where the same document (a codebase, a legal corpus) is queried repeatedly. For a financial researcher building a hedge-fund investment strategy with Argon analyzing a 500K-token annual report, the cache discount compounds. However, the pricing is quoted per-million tokens, and Argon's higher usage of tokens for thinking (compared to fast-inference models) means actual per-task cost will depend on the thinking depth users request.

Argon's arrival intensifies the frontier-model release cadence. In the eight weeks since Mistral Large 4 (October 6) and in the five months since Anthropic Opus 5.5 (May 2026), the frontier has produced five major releases: DeepSeek V4-Pro (August 13), Mistral Large 4 (October 6), and now Gemini 4 Argon (October 13–14). Google, Anthropic, OpenAI, xAI, and Mistral are each on roughly 4–8 week release cycles. This density of capability increases suggests the frontier is in an acceleration phase: models are improving faster, specialized capabilities (reasoning, code, vision, long-context) are maturing, and the differentiation between labs is shifting from raw scale to specific capability gaps. For enterprises and researchers, the short release cycle is both an opportunity (pick the best model for your task) and a moving target (last month's frontier is dated by this month's release).