Gemini 4 Argon: 1M Output, Price Doubles Dec 1

Output: 1M tokens, but not from a normal request

The model page lists a 1,048,576-token context window, 65,536 output tokens by default (up to 1,048,576 with long decode), text, image, video, audio and PDF input, and text output, according to KalFoundry. The 1,048,576-token figure is the unified context (what it reads); the 1M output limit (default 65,536, up to 1,048,576 with long decode) is what it writes, according to BougainWell. The output limit is the story, not the leaderboard, according to BougainWell: Argon writes up to 1M tokens per response, up from 64K — about a 16x jump — which is what turns a code-completion model into one meant to finish the whole task, according to BougainWell.

Google's announcement says Argon can write up to 1M output tokens, according to ComputingForGeeks. You will not get that from a normal request, according to ComputingForGeeks. The API's own model entry reports a much smaller number, according to ComputingForGeeks, and Google's docs give 65,536 output tokens per synchronous request, according to ComputingForGeeks. Per Google, long decodes via GenerateContent currently work only through API version v1beta1, according to InnFactory, and no tuning, no Live API, no RAG Engine, per the model page, according to InnFactory.

So the honest summary is: Argon can write up to 1M tokens, but only with long decode, only via v1beta1, and only if you configure it for long decode — a normal API request defaults to 65,536 output tokens (or 32,768 in one API listing, according to ComputingForGeeks — the exact default varies by API version, so check the model page for your API version). For most coding tasks, 65K output is already a lot — a large file, a multi-file refactor, a long report. For "finish the whole task" — a complete app, a full migration, a long research report — you need long decode, and you need to budget for it: 1M output tokens at $10-$20 per million is $10-$20 per response, plus input and thinking tokens (see below).

Context: 1,048,576-token unified context window shared by input, thinking and output, according to InnFactory. Long decodes allow up to 1,048,576 output tokens (default per request: 65,536), according to InnFactory. That unified context means input + thinking + output share the same 1M window — so a 1M output leaves little room for input or thinking, and vice versa. In practice, long outputs work best with moderate inputs, not with a 1M-token codebase plus a 1M-token output — the window is shared, not separate.

Input modalities: text, image, video, audio and PDF input, and text output, according to KalFoundry. Input: text, image, video, audio, PDF; thinking levels low/medium (default)/high, computer use (preview), function calling, code execution, grounding, Provisioned Throughput, according to InnFactory. Knowledge cutoff March 2026 per the model card, according to InnFactory. So Argon is multimodal in, text out — no image or audio output, unlike some other models. For coding, enterprise workflows, and cyber defense, text output is what you need — but if you need image generation, Argon is not the model.

Price: $2/$10 intro, $4/$20 from December 1, no free tier

Paid-tier pricing is $2 per million input tokens and $10 per million output tokens through November 30, 2026, rising to $4 and $20 from December 1, 2026; there is no free tier, and only implicit caching is supported, according to KalFoundry. Argon launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input rate, according to BougainWell. Google's pricing page frames the schedule by date rather than by a vague "intro" label, according to BougainWell:

Argon costs $2 per million input tokens and $10 per million output tokens until November 30, 2026, according to ComputingForGeeks. From December 1 the price doubles to $4 and $20, according to ComputingForGeeks. Google calls the lower figure an introductory price, and it's printed that way on the Gemini API pricing page, according to ComputingForGeeks. The full mode table, from ComputingForGeeks:

There is one rate across the whole 1M-token window, according to ComputingForGeeks. Gemini 3.1 Pro charged more above 200K tokens, and Argon drops that tier, according to ComputingForGeeks. Grounding with Google Search gives 5,000 free requests a month across Gemini 3 and newer models, then costs $14 per 1,000, according to ComputingForGeeks. There is no free tier, according to ComputingForGeeks. Google says so on the pricing page, and our free-tier key hit a quota limit of 0, according to ComputingForGeeks.

So budget at $4/$20 beyond November, not $2/$10 — the same headline pair as Claude Opus 5.5, according to ExplainX. Batch/Flex at half price are for non-urgent work; Priority ($3.60/$18 intro, $7.20/$36 from Dec 1) is for urgent work only. No free tier — you need a paid API key or Google AI Ultra. Availability: paid API + AI Studio now, Ultra in Antigravity + Gemini app + AI Mode, Ultra/Pro in Gemini app + AI Mode, Fairwind expanding to hundreds of partners, according to BougainWell and ExplainX.

Benchmarks: strong, but domain-targeted, not universal

Google reports 77.9% on DeepSWE v1.1, 68% tie on CWE-bench, 51.3% #1 on AutomationBench, and 91.7% on LVBench, according to KalFoundry and BougainWell — domain-targeted, not universal, with 51.3% the honest ceiling on unattended automation, according to BougainWell.

Per Google's evaluation, from InnFactory: DeepSWE v1.1 77.9% (3.8 Flash: 73.7%), Terminal-bench 4.0 57.4% (19.1% for Flash — a huge gap, suggesting Argon is much better at terminal/agentic tasks), OSWorld 2.0 69.2% (59.0%), LVBench 91.7% (87.8%), GraphWalks 256k to 1M 84.2%, according to InnFactory. Strongest official scores, from ExplainX: Vals Index 68.9%, DeepSWE v1.1 77.9%, AutomationBench 51.3%, GraphWalks 256k–1M 84.2%, according to ExplainX. Independent long-horizon eval, from ExplainX: Vending-Bench 2 #3 — $13,718 ± $3,100 simulated year-end cash, behind Astra and GPT-6 Sol, according to ExplainX.

Argon's safeguards include refusing harmful cyber/CBRN requests, monitoring activations, adversarial training against prompt injection, and monitoring chain-of-thought, according to KalFoundry. Fairwind defenders can get Argon without cyber guardrails on a rolling basis, according to ExplainX. Bloomberg reported internal skepticism that Gemini 4 does well on benchmarks but less well in real work, according to ExplainX; Google disputed it. Treat vendor benchmarks as domain-targeted, not real-world verdicts, and check ComputingForGeeks' same-prompt test against Opus 5.5 and GPT-6.1 Sol before choosing.

The thinking-token catch: output price includes thinking

Output price includes thinking tokens, according to ComputingForGeeks. That matters more here than it did on older Gemini models, because Argon at high spent three tokens thinking for every one it showed us, according to ComputingForGeeks. Gemini 3.1 Pro defaulted to high, and Argon defaults to medium, according to ComputingForGeeks. Only high thought, using 1,100 to 1,800 tokens per call, according to ComputingForGeeks — though that figure seems low for a "high" thinking level, and may refer to a specific test, not all high-thinking calls, so treat it as a test-specific figure, not a general rule.

The catch is: you pay for thinking tokens at output price ($10-$20 per million), and Argon at high thinking spends multiple thinking tokens per output token — in ComputingForGeeks' test, three thinking tokens for every one shown token, at high. That means a 10,000-token output at high thinking could cost you 30,000 thinking tokens + 10,000 output tokens = 40,000 tokens at output price, or $0.40 at $10/M, $0.80 at $20/M, before input. At medium (default), thinking is lower, but still counts. So budget for thinking, not just output — and use low/medium thinking for simple tasks, high only for complex reasoning where the extra thinking is worth the cost.

Temperature, top_p and top_k are ignored, according to InnFactory. Thinking_budget, 'MINIMAL', candidate_count, frequency_penalty and presence_penalty return HTTP 400, according to InnFactory. So Argon has a restricted API surface — you can't tune temperature or penalties the way you can on other models, and some parameters error out. Check the model page for supported parameters before integrating, and don't assume your existing Gemini 3 code will work unchanged on Argon.

EU: No EU single region (e.g. Frankfurt) — in the EU only via the EU multi-region; single regions only in London (UK) and Singapore, according to InnFactory. EU multi-region endpoint with documented DRZ and MLP per the model page, from the first GA day, according to InnFactory. So EU users get Argon via EU multi-region, not a Frankfurt single region — which may matter for data residency requirements, so check your compliance needs before using Argon in the EU.

Verdict: use Argon for long, complex tasks, not for everything

Argon is Google's flagship reasoning model for complex software engineering, deep math/science, autonomous multi-step agents, and long decode up to 1M tokens, according to BougainWell. Use it for: long-horizon coding (multi-file refactors, complete apps, migrations), enterprise workflows, cyber defense, and science/math problem solving, where 1M output and 1M context matter, and where $4/$20 (from Dec 1) is justified by the task's value.

Don't use Argon for: simple coding, chat, or short tasks — cheaper models like Gemini 3.8 Flash (73.7% on DeepSWE, vs Argon's 77.9%, at a fraction of the price) or GPT-6.1 Terra ($1/$5) are better value for simple work. Don't use Argon on free tier — there is none. Don't budget at $2/$10 beyond November — budget at $4/$20. And don't assume 1M output from a normal request — you need long decode via v1beta1, and you need to budget for thinking tokens at output price.

For most developers, the smart pattern is: Flash for simple tasks, Argon for long, complex tasks, Batch/Flex for non-urgent Argon work to halve the price, and high thinking only when the task justifies the thinking-token cost. Argon is a powerful, expensive, long-output model — use it where length and complexity justify the price, not as a default for everything.