Google DeepMind Launches Gemini 4 Argon Frontier Model

Full analysis

Google DeepMind just shipped Gemini 4 Argon, a frontier model built for long, multi-step work. The piece everyone will miss: it can now produce up to 1 million output tokens in one go, a 15x jump over the old 64K ceiling, and Google is handing it to cybersecurity defenders with the attack guardrails switched off. Pricing is $2 per million input tokens, $10 per million output, and 95% off on cached inputs.

This is easy to undo for anyone kicking the tires. It is hard to undo if you rebuild an agent pipeline around a 1M-token output budget and a cache discount only Google's infrastructure offers, then find yourself locked to one vendor's economics. Nothing here forces a decision this week. The phased Fairwind rollout means most readers can't even touch it yet.

What's actually being decided isn't "is Argon the best model." It's whether long-output agentic work (migrations, legal review, autonomous vuln patching) has crossed from demo to production, and whether Google's pricing resets what everyone else has to charge.

The Skeptic

Google is grading its own homework on every number. DeepSWE v1.1 at 77.9%, and Google controls the versioning. "300 TiB of memory freed" and "40% on quantum optimization" have no outside auditor and can't be checked. The Fuchsia kernel rewrite is the one claim with teeth, because a shipped Rust kernel is either there or it isn't. The 1M output-token limit is real as a spec, but sustained coherence across a million tokens under messy production traffic is a different animal than a clean internal run. And "without cyber guardrails for trusted defenders" is a tidy way to generate dramatic security PR while punting the dual-use policy fight to later.

The Safety Lens

Shipping a frontier model to defenders with the attack guardrails off is the biggest decision in this announcement, and it sits in one paragraph. The principle holds: defenders need the same firepower as attackers. But "trusted defenders" is a trust boundary, and trust boundaries leak. The Fairwind vetting bar is opaque. The genuinely new thing is Google watching the model's own chain-of-thought (its step-by-step reasoning) for signs of misalignment, and asking the industry to keep that reasoning readable rather than hide it. That's a quiet coordination play to set a norm before any regulator does. Announcing prompt-injection defenses without publishing how well they hold is a posture, not a result.

The Compute Pragmatist

A 1M-output model run at Google's scale burns enormous memory bandwidth, so the 300 TiB "freed" story probably pays for part of Argon's own serving bill. The 95% cache discount isn't generosity. Google built aggressive prefix-caching on its own TPU pods and needs to fill idle capacity, so it's pricing to pull output-heavy work onto hardware it owns end to end. That's the squeeze. Rivals renting NVIDIA GPUs from third parties can't match a 95%-off cached path without the same vertically owned stack. Anthropic and OpenAI can discount cache too, but the margin math is worse when you don't own the chips. Expect a pricing response within two quarters, or a visible gap on long-output jobs.

The Builder

The cache discount changes agent economics on Tuesday morning. Long sessions that reread the same big context (a codebase, a contract set, a case file) get cheap, because you pay full freight once and 95% off every turn after. That's the whole game for coding agents and document review. But a fully maxed 1M-output session runs about $10K in output alone. That's an enterprise-only ceiling. Anyone building for small business should sit this capability out until it's cheaper. The first thing to break in 90 days won't be the model. It'll be prompt injection inside multi-agent chains, where the hardened sandbox guards the outside wall but not the agents trusting each other.

The Enterprise Buyer

For a CTO, the attraction is obvious and the paperwork is missing. Autonomous code migration and long-horizon legal/finance/tax work map straight onto expensive internal headcount. But there's no indemnification language, no third-party eval, no data-residency detail, and no SLA in this announcement. "Available to trusted defenders" is not a contract. The Fuchsia rewrite is the reference customer story, and the reference customer is Google itself, which sets a prior no outside shop will hit. Procurement signs for audit logs, liability cover, and a number they can verify. Right now they'd be signing for a benchmark Google versions and controls.

Where they disagree

The Researcher and the Skeptic split on what the 1M output ceiling means. One says it's the durable contribution, because output length (not input) has been the real wall for agents that write code or long documents. The other says a spec sheet number tells you nothing about coherence under adversarial load. Both are right, and that gap is exactly what a buyer has to test.

The Compute Pragmatist and the Enterprise Buyer split on the cache discount. The Pragmatist sees genuine structural advantage from owning the chips. The Buyer sees a lock-in: build around 95%-off cached inputs and you've welded your economics to one vendor's TPU stack, with no portable equivalent on GPU clouds.

The Safety Lens stands alone on the guardrails-off decision. Nobody else priced it, and it's the one choice here that can't be walked back once a "trusted defender" turns out not to be.

What it hinges on

Three things. Does the 1M output stay coherent on real, messy work, or degrade into slop past a few hundred thousand tokens. Does the cache discount survive once idle TPU capacity fills up, or is it a launch loss-leader. And does anyone outside Google reproduce the headline gains without Google's internal infrastructure. Before committing a pipeline, run your own long-output eval on your actual codebase or document set, measure where coherence breaks, and price a version that doesn't lean on the 95% cache path so you know your floor if that discount moves.

Prediction: At least one of OpenAI or Anthropic will raise its published maximum output-token limit to 256K or higher on a frontier model by the end of Q1 2027 (by the next major model release from either lab, expected by March 2027).

Confidence: Medium. Output ceiling is now the competitive axis, and rivals move on exactly this.

Why: Gemini 4 Argon just made 1M output tokens the visible frontier spec, and output length is the actual bottleneck for the agentic coding and document work all three labs are chasing, not input context, which is already in the millions. Once Google names a number this large on the axis that matters for agents, matching it becomes a checklist item for the next OpenAI and Anthropic releases, the same way context-window one-upmanship played out in 2024. The opposite (both labs holding at today's lower output caps through Q1) would mean conceding the agent-workflow story to Google, which neither has done when a rival set a public spec. I'm not calling a full 1M match because the serving cost of long output is brutal without owned silicon, so a jump to the 256K range is the more likely first move.

Revisit by 2027-03-31: We're right if OpenAI or Anthropic ships a frontier model with a published output-token limit of 256K or more. We're wrong if both keep their top models' output limits below 256K through March 2027.