Google DeepMind has introduced Gemini 4 Argon in a restricted preview, pairing a new frontier model with experimental output of up to one million tokens, according to Latent Space’s AINews roundup published on 1 October. The important distinction: that output stretches across resumed API calls, and access is not yet a general invitation to developers. That makes this a substantial model announcement with two practical questions attached: who can use it, and whether longer answers deliver enough useful work to justify their bill.
Google DeepMind Watch analysis
What happened
The reported initial access is for government users and trusted cyber defenders through the Fairwind Program. Google says it will refine safeguards before extending access to developers, enterprises and consumers. The account gives no firm date for that wider opening.
The experimental Long Decode Continuation feature pauses long responses and resumes them across calls. Artificial Analysis is cited as reaching one million output tokens this way, while Vals lists a maximum output of 262,000 tokens. Those figures should not be treated as interchangeable measurements of a single uninterrupted response.
The announced standard pricing is $4 per million input tokens and $20 per million output tokens. A 50% introductory discount brings those rates to $2 and $10, with no end date reported. Cached input receives a reported 95% discount.
Why it matters
Longer output could make room for substantial code changes, extended analysis and other work that would otherwise need splitting into smaller responses. Continuation changes how that work is delivered; it does not, by itself, establish that a million-token answer stays accurate or coherent throughout.
The pricing comparison also deserves more attention than the headline token rate. Artificial Analysis figures cited in the account put Argon at $1.99 per task at introductory pricing, against $3.26 for GPT-6 Astra. At Argon’s standard pricing, the same comparison rises to $3.98. Its measured average output was 62,000 tokens per task, against Astra’s 27,000: the reported saving comes from the discount, not thriftier generation.
Our read
Argon looks like a meaningful expansion of Google’s frontier offering, not merely a larger text box. The reported Artificial Analysis Intelligence Index score of 53 matches Astra, giving the announcement a concrete competitive signal alongside its unusual output feature.
But the buying decision is workload-shaped. A lower advertised token price can lose its advantage when a model produces considerably more tokens, and restricted access means most teams cannot yet put the proposition through their own tests. The calculator deserves a seat beside the leaderboard.
For teams planning evaluations, compare completed-task quality, total token use and elapsed time. Test continuation on work where length is genuinely useful, rather than rewarding a model for discovering that the scroll bar has no moral objections.
What to watch
- A firm timetable and eligibility rules for wider developer and enterprise access.
- How Long Decode Continuation handles coherence, recovery and completion across resumed calls.
- Whether the introductory discount ends, and how that changes cost per completed task.
- Evaluations across ordinary coding and knowledge-work workloads, beyond headline benchmark scores.
Discussion spark: Should teams favour a model that can produce much longer outputs, or demand better completed-task efficiency before treating output length as an advantage?
Sources and evidence
- [AINews] Gemini 4 Argon: GDM’s answer to Astra/Fable, with 1M output (1 October 2026, 06:45 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind