Two weeks ago the honest read on Gemini 4 Argon was that the announcement was not the access. Fairwind trusted cyber defenders could touch it. Most marketing, growth, and agency builders could not.
That changed on October 13.
Koray Kavukcuoglu’s Google Blog post says Gemini 4 Argon is now globally available for developers and enterprises via the API in AI Studio and the Gemini platform, for Google AI Ultra subscribers in Google Antigravity, and for Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search. Google Cloud’s same-day post adds Gemini for business on web, desktop, and mobile, and names the API model ID: gemini-4-argon. The Verge and Ars Technica both corroborate the paid-only consumer roll and the Ultra gate on Antigravity.
The gate swung open. The operator job did not get easier. It got more specific: which door you use, what you pay after November 30, whether your stack actually continues past 64K tokens, and whether you still need Fairwind.
A note on evidence. This is a primary-source review. I have not run Argon in AI Studio, the API, Antigravity, or the Gemini app for this piece. Everything below comes from Google’s October 13 launch posts, the Gemini API long-decode and “what’s new” docs, the Cloud developer guide, the DeepMind model card, and same-day trade coverage from Ars Technica and The Verge, all read on October 14.
What it is
Gemini 4 Argon is Google’s frontier model in the Gemini 4 family: a sparse mixture-of-experts transformer with native multimodal inputs, built for long-horizon work in coding, finance, legal, science, and cyber defense. The model card and launch posts frame it for multi-step agentic tasks, deep document synthesis, and reasoning that can run longer than earlier Gemini generations.
On the API surface that marketers and agency builders will actually wire, three product facts matter more than any benchmark table.
First, one model ID for public access: gemini-4-argon. Cloud docs also name gemini-4-argon-cyber-permissive for Fairwind-approved cybersecurity workloads. Those are not the same product for the same buyer.
Second, a unified 1,048,576-token window shared across input, thinking, function calls, and output. There is no separate million-token output bucket sitting beside a separate input budget. Fill the window with a fat brief and you have less room left to think and answer.
Third, long decode (Google’s docs also call the client-side path a continuation protocol). A single request still tops out around 64K output tokens on streaming and SDK calls, or 32K on non-streaming REST, unless you turn on continuation. Past that limit the model returns finish_reason: CONTINUATION and a continuation_token. You (or a new enough SDK) resend the original request with that token. You do not paste the partial answer back into contents and ask it to keep going.
What changed on October 13
On September 30 Argon was announced with intro pricing and Fairwind-first access. Vault covered that as announcement-is-not-the-access: live for trusted cyber defenders, Ultra and paid API named as a later path without a date.
October 13 is the dated flip.
Public Argon is buyable on the paid Gemini API and in AI Studio. Enterprises get it in Gemini for business and on the Gemini platform. Consumers on Pro or Ultra get it in the Gemini app and AI Mode. Ultra subscribers get Antigravity. Scientists can apply for six months of Antigravity’s highest usage tier free (10,000 seats, rolling, subject to eligibility and capacity). Fairwind begins expanding to hundreds of additional partners for Argon without cyber guardrails.
The pricing footnote is now a calendar event, not a vague “introductory period.” After November 30, 2026, public Argon moves from $2 / $10 per million input / output tokens to $4 / $20. Cached input stays at 95% off the then-current input rate. Google is explicit that the intro schedule applies to the public-access Argon model only.
Ars Technica’s Ryan Whitwam adds the consumer fine print operators will hit in practice: Gemini app access for AI Pro and Ultra, Antigravity and AI Studio coding surfaces Ultra-only for now, and no free-tier Argon. The Verge’s Robert Hart rests the same point: unclear when or if Plus or free users get it. Treat that as access architecture, not as a complaint about Google’s marketing.
What works
You can finally staff a real eval. Two weeks ago a ticket that said “bench Argon on Ultra” was fiction for most teams. Now a paid API key or a Pro/Ultra seat is enough to put Argon next to your current frontier stack on your own briefs, not Google’s demos.
Intro pricing is aggressive while it lasts. At $2 / $10, Argon undercuts many frontier rivals on list price. Cached input at 95% off matters for agent loops that re-send the same system prompt and tools. Google’s long-decode billing notes make the cache discount do real work across continuation hops, because each hop re-prefills prior tokens as input.
The continuation path is documented. Interactions API background execution can run hops server-side when you set background: true and raise max_output_tokens past 65,536. GenerateContent can auto-continue in current Google GenAI SDKs (Python google-genai 2.30.1+, JavaScript @google/genai 2.29.1+, and matching floors in Go, Java, .NET, Kotlin), or you can own the continuation_token loop yourself. That is the difference between “1M context on the slide” and “1M context in production.”
Thinking level is an operator dial. low, medium (default on public Argon), and high replace the old thinking_budget / MINIMAL knobs. Cloud docs say thinking_budget and MINIMAL return HTTP 400 on Argon. For GTM and agency builders, that means you stop treating temperature as the quality lever and start A/B testing thinking level against latency and spend.
Defensive cyber is in the public SKU. Google Cloud says public Argon blocks offensive tasks such as pentesting and exploitation, while still supporting defensive work through CodeMender or your own harness. CWE-bench v1 is reported at 68%, tied for first in Google’s framing. Teams that only need remediation help do not have to wait on Fairwind.
What breaks or stays gated
Fairwind is still a separate door. Public Argon is not “Argon without cyber guardrails.” The launch post says Google is expanding Fairwind to hundreds more partners and that cyber defenders should apply for rolling access to the unmitigated track. Cloud docs name gemini-4-argon-cyber-permissive with a default HIGH thinking level for Fairwind-approved projects. If your agency pitch depends on offensive simulation, you are still in an apply queue, not a checkout flow.
Consumer access is paid, and coding agents are Ultra-shaped. Pro gets the Gemini app and AI Mode. Antigravity’s multi-threaded agent workflows, per Ars, stay on Ultra. Free users are not in this launch. Do not brief a client that “Gemini has Argon now” without naming the seat.
Long decode is easy to misconfigure. Without raising max_output_tokens and enabling background or continuation, output still stops at 64K (or 32K on unary REST). Older SDKs will not continue for you. The OpenAI-compatibility endpoint does not support long decode. Agents that compact context off the summed prompt_token_count across hops will trim too early; docs tell you to use initial_prompt_token_count (or the first hop’s prompt count in REST) instead. Thinking-only hops that return empty text with CONTINUATION are normal. Treat them as progress, not as a dead response.
Unified context punishes fat prompts. Input plus thinking plus tools plus output share one million tokens. A 200K brief leaves less ceiling for a long proof or a multi-file refactor. Prompts near the input ceiling are a documented reason not to use long decode.
Explicit caching is gone. Migration guides say to remove cachedContents / caches.create. Implicit caching covers shared prefixes (4,096 tokens on the consumer API docs, 6,144 on the Cloud guide) with durable caching on Cloud. Front-load static instructions if you want the 95% hit rate.
Sampling knobs are ignored or rejected. temperature, top_p, and top_k are ignored. frequency_penalty, presence_penalty, and candidate_count return 400 on Cloud. Prefilling model turns in chat history is unsupported. Your old Gemini 3.1 Pro request shape will not silently upgrade.
Safety remains a product feature, not a footnote. The model card says Google cannot rule out cybersecurity and biological CBRN critical-capability levels without mitigations, and that those mitigations are why public and Fairwind tracks differ. Gray Swan indirect prompt injection success is reported at 0.7% across 15 adaptive attempts. Shortcut attempts on coding benches drop to 2% with a real-time misalignment monitor. That is good for brand-safe agents. It is also why red-team fantasies of “full Argon” on a normal API key are wrong.
Knowledge cutoff is March 2026. The model card says so. Do not ask Argon to be your live news desk for October launches without retrieval.
Who it is for right now
It fits API and AI Studio builders who need a frontier model for long coding, legal/finance document work, or agentic pipelines and who can upgrade SDKs this week. It fits Gemini for business teams handing off multi-step research and drafting inside Google’s seat. It fits Pro and Ultra subscribers who want Argon in the Gemini app or AI Mode, and Ultra shops that live in Antigravity. It fits security teams that want defensive remediation on the public SKU, and Fairwind-eligible defenders who still need the cyber-permissive track.
It fits badly as a free-tier chat upgrade, as a drop-in for teams that refuse to raise timeouts past a few minutes, as a “write a 700,000-word book” toy (docs say the model ignores length targets), or as a substitute for Fairwind if you need unmitigated cyber capabilities.
For marketers and agency builders specifically: use Argon where the job is long synthesis, migration scripts, measurement pipelines, or creative systems that need deep multi-file reasoning. Keep lighter models on short copy, classification, and high-QPS chat. The who-for matrix is a SKU choice, not a brand preference.
Pricing and limits
Introductory public API price through November 30, 2026: $2 per million input tokens, $10 per million output tokens (output includes thinking). Cached input: 95% off ($0.10 per million during intro). From December 1: $4 / $20 input / output, with cached input at $0.20 per million if the 95% discount holds on the new base. Consumer API docs say Argon is not on the Free tier or the Google AI Pro plan for API access. Antigravity scientist grants are separate from API list price.
Limits operators should put on a sticky note: unified 1,048,576 tokens; default single-request output 65,536 (32,768 unary REST); long decode up to 1M minus input; HTTP read timeouts of about 20 minutes for hard jobs; background Interactions can run up to 24 hours; cache between hops lasts roughly 15 to 20 minutes before re-inputs lose the discount.
Build every business case at $4 / $20. Use the intro window to learn token shapes on your workloads, not to lock a forever COGS assumption into a client retainer.
What to do this week
The gate is open. The calendar is the constraint.
- Map seats before demos: paid API / AI Studio for builders, Gemini for business for enterprise seats, Pro for Gemini app and AI Mode, Ultra for Antigravity. Write “not free tier” on the brief.
- Price every pilot at the post-November $4 / $20 rate. Treat $2 / $10 as a temporary learning discount that ends November 30.
- Upgrade GenAI SDKs past the long-decode floors, set
max_output_tokensabove 65,536 only for jobs that need it, and turn on background execution or a realcontinuation_tokenloop. Smoke-test one hard prompt end to end before you migrate a production agent. - Separate Fairwind from public Argon in every deck. If you need cyber-permissive access, apply. If you only need defensive remediation, stay on public Argon plus CodeMender or your harness.
- Retire explicit cache calls, stop sending temperature and top-p, front-load static system prompts for implicit cache, and raise client timeouts to about 20 minutes on hard tasks.
- Run one side-by-side on your actual marketing or GTM workload (brief synthesis, measurement SQL, creative systems refactor) at
mediumandhighthinking levels. Keep the cheaper level that clears your quality bar.
Public Argon is no longer a rumor with a Fairwind velvet rope. It is a dated SKU with a November price cliff and a continuation setup you either wire this week or pretend you have. Wire it.
