Gemini 4 Argon

Gemini 4 Argon is Google\u2019s new frontier flagship, announced on 30 September 2026. It sits at the top of the Gemini line and is positioned for four kinds of work: coding, knowledge work, cybersecurity defense, and creative writing.

Two things make the announcement notable beyond the specs. First, the naming: Argon is the first model in the Gemini 4 generation and the first to carry an element name under Google\u2019s new naming scheme — expect the generation to grow along the periodic table. Second, what it replaces: Gemini 3.5 Pro was cancelled before shipping, so Argon is not an upgrade to a model you could already buy; it is the ship that finally left the dock.

The headline spec change is output length. Argon raises the output token ceiling from 64K to 1M tokens — a 16× jump that changes what a single call can produce: full-repo refactors, complete documents, and long agent trajectories without intermediate stitching.

Announced 30 September 2026. On 14 October 2026 the same Google blog says the API is globally available for developers in AI Studio and the Gemini platform. The original rollout order was:

• 01Fairwind Program — trusted cyber defenders get access first.
• 02Google AI Ultra subscribers and paid API customers — the first commercial wave.
• 03Full availability — everyone else, including API resellers, after that.

Google\u2019s own API is the path that opened on 14 October 2026. A reseller still has to have a channel. llmrelay does not: gemini-4-argon returned 503 on our existing Gemini key the same day.

Three announced paths, in the order they open:

• +Fairwind Program — Google\u2019s program for trusted cyber defenders. First in line, which tells you how much of the launch story is defensive capability.
• +Google AI Ultra subscription — the consumer-side route; Ultra subscribers get Argon in the first commercial wave.
• +Paid Google API account — the developer route, at the intro prices below, also in the first commercial wave.

Developers can call it on Google\u2019s API in AI Studio and the Gemini platform. Ultra subscribers also get it in Antigravity; Ultra and Pro get it in the Gemini app and AI Mode in Search. That is Google\u2019s path, not an llmrelay endpoint.

Google\u2019s 14 October 2026 post does not announce a free API tier. The access it names is the paid API, Google AI Ultra, and Ultra plus Pro in the Gemini app and Search.

This is worth stating plainly because the search results already contain pages promising a \u201cfree Gemini 4 Argon trial\u201d. The 14 October post does not describe a free API tier. Paid Google API access is what opened.

We will update this section when that changes — either Google announces a free tier or the model reaches general availability.

Google announced API pricing of $2 per million input tokens and $10 per million output tokens during the introductory period, moving to $4 / $20 after intro ends.

The other pricing line that matters for heavy users: cached input is billed at 95% off the input price — effectively $0.10 per million tokens during intro. On repetitive workloads (fixed system prompt, stable codebase context) that is the difference between paying for the same context once and paying for it every call.

Prices are Google\u2019s published list, fetched 14 October 2026 from the Argon blog. Intro $2/$10 runs through 30 November 2026, then $4/$20. llmrelay is not serving Argon, so no relay price is quoted. Models we do sell are on the current model list.

The benchmark Google led with: on DeepSWE v1.1, Gemini 4 Argon scores 77.9% against 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra.

Read those numbers with their caveat: they are Google\u2019s own published results for a model. Vendors lead with benchmarks their model wins. The 3.7-point gap over Opus 5.5 is Google\u2019s table, not an independent rerun.

Beyond DeepSWE, Google\u2019s announcement claims leads on four more fronts:

• +CWE-bench v1: 68% — tied for first on this cybersecurity exploit benchmark.
• +AutomationBench: 51.3% — first place.
• +LVBench: 91.7% — state of the art on this long-video benchmark.
• +Finance and legal agents — leading results on Vals Finance Agent v2 and the Harvey Legal Agent Benchmark (Google published no specific scores for these two).
• +Prompt-injection robustness — leading on Gray Swan\u2019s IPI benchmark, which matters given the guardrail controversy below.

The pattern across the suite: software engineering, cybersecurity, automation, and long-video understanding. That is a defense-and-agents-heavy scorecard, which matches both the Fairwind first-wave and the positioning lines Google chose.

The 64K → 1M output jump is the spec most likely to change how work gets structured, and Google backed it with internal case studies rather than abstractions:

• +Fuchsia Zircon kernel: 800K+ lines migrated from C/C++ to Rust — the kind of multi-week, repo-scale effort a long-output model can hold in a single trajectory.
• +libgav1: 32K lines of SIMD code rewritten in Rust, with the decoder coming out 2.7× faster — a rewrite measured by its runtime result, not just its line count.
• +Data center memory optimization: 300+ TiB freed — infrastructure-scale savings attributed to Argon-driven work.

These are Google\u2019s internal numbers from the announcement, not third-party reproductions. But they sketch the intended use: long-horizon engineering tasks where the model plans, edits, and verifies across a codebase far larger than a single context window used to allow.

The most debated part of the announcement: the Fairwind Program version of Argon ships without cyber guardrails — the safety layers that normally sit on a frontier model are removed for trusted defenders.

Google\u2019s framing is that defensive security research needs the unfiltered model: probing real exploits with guardrails in the way is like testing a firewall through a proxy. The counterargument writes itself — \u201ctrusted\u201d is a judgment call, and an unguarded frontier model is exactly the artifact offensive actors want.

The announced mitigations: Argon\u2019s alignment mechanism monitors the model\u2019s chain-of-thought and can abort execution, and Google points to Argon\u2019s leading result on Gray Swan\u2019s IPI benchmark for prompt-injection robustness as evidence the underlying model resists misuse even where guardrails are lifted. Whether that holds in practice is one of the more consequential open questions of this launch.

These two are not in the same tier, and today only one of them can actually run your workload:

• +Gemini 3.8 Flash — Google\u2019s fast, cheap tier. Call it when latency and cost dominate: classification, extraction, high-volume summarization, RAG. It is available via API providers today.
• +Gemini 4 Argon — Google\u2019s frontier flagship, live on Google\u2019s API at $2/$10 through 30 November 2026. Not sold on llmrelay.

Practical answer: Google\u2019s API can run Argon. llmrelay cannot, as of the 14 October probe. If you need a model this site actually sells, use the live list.