๐ Spot: Gemini 4 Argon's introductory price — $2 input and $10 output per million tokens — closes on November 30 and doubles on December 1; stepping off a fifth-of-class entry price, the post-doubling rate lands level with Claude Opus 5.5 and at half of GPT-6 Astra.
⏱️ 7 min read
๐ In This Article
- ๐ The price ladder: which table does $2/$10 sit at
- ๐ The December 1 doubling: what the introductory price says
- ๐ One price list, two wallets: the Argon–Carbon split
- ๐ The 1M shared pool: the agentic math of long decode
- ๐ The table is not one-sided: where it leads, where it trails
- ๐ The launch backdrop: the calendar and what Reuters reported
- ๐ Murat Karakaya Academy verdict
- ๐ References
๐ What actually defines Gemini 4 Argon, which Google opened to developers on October 13, is its price: at the introductory rate of $2 input and $10 output per million tokens, the entry ticket to the reasoning class it plays in costs a fifth of the going rate; and the doubling scheduled for December 1 puts a sharper question on the table — will buyers carry the habit they built at the introductory table into the new price sheet? Argon is the probe at the end of that experiment; the subject of the story is the price band itself. ๐ Google's official documentation calls the model an "efficient frontier reasoning model"; the "strongest" label sits with its sibling Carbon, released November 10, and the "flagship" phrasing belongs to Reuters' October 13 story [1][4][11]. ๐️ At the time of writing (November 14) the introductory window is open with 16 days left to close; the new price sheet takes effect on the morning of December 1 [6].
1. The price ladder: which table does $2/$10 sit at
๐ต The numbers on the official pricing page are plain: $2.00 per million input tokens, $10.00 per million output (thinking tokens included), $0.10 for cached input — all of it through November 30; from December 1 it is $4.00 input, $20.00 output and $0.20 cache [6]. The "what's new" page repeats the same figures under the "introductory price" label; the model has no free tier — access starts at the paid plan [6][10]. Set the official competitor pages side by side and the rungs of the ladder come into focus:
| Model (top-tier reasoning) | Input /1M | Output /1M | Notes |
|---|---|---|---|
| Gemini 4 Argon (introductory, through Nov 30) | $2.00 | $10.00 | cached $0.10; no free tier [6] |
| Gemini 4 Argon (from Dec 1) | $4.00 | $20.00 | [6] |
| Gemini 4 Carbon (introductory) | $2.00 | $10.00 | same price as Argon [6][11] |
| OpenAI GPT-6 Astra | $10.00 | $50.00 | cached $1; long-context 2x; Batch/Flex $5/$25 [13] |
| Anthropic Claude Fable 5.1 | $10.00 | $50.00 | cache-read $0.25; Batch 50% off [14] |
| Anthropic Claude Fable 5.2 | $8.00 | $40.00 | successor to Fable 5.1 [14] |
| Anthropic Claude Opus 5.5 | $4.00 | $20.00 | [14] |
๐ช Here is the read: inside the introductory window, Argon sells for a fifth of the $10/$50 input-output class occupied by GPT-6 Astra and Claude Fable 5.1 [6][13][14]. After December 1 it drops to half of the ladder's top tier and lands on exactly the same rung as Anthropic's Claude Opus 5.5. The ladder also has a fixed rung: OpenAI's GPT-6.1 Sol, a permanent $2/$10 model carrying no introductory label — meaning the "cheap reasoning" market now competes on two tiers at once, temporary introductory price versus permanent low price [13]. The center of gravity in the price argument has shifted too: in the Fable 5.1 announcement Anthropic argues cost through cache-read discounts (75%, $0.25), because in agentic workloads the dominant line item is cache reads [14][15].
2. The December 1 doubling: what the introductory price says
๐ The scope of the increase is drawn clearly on the official page: only the price sheet — input, output, cache, all three doubling; the documentation says "introductory pricing through November 30, 2026" and offers no rationale [6][10]. Reuters reports the same thing in the same flat way — "rates set to double after November 30" — without commentary [4]. No rationale published by Google could be found behind the increase; this is an absence statement about the official pages scanned.
⚖️ One pattern observation, labeled as commentary: the same "introductory price → 2x on expiry" structure also exists in the family's previous member, Gemini 3.8 Flash — $0.75/$3.75 until December 31, 2026, then $1.50/$7.50 [2][6]. If you ask me, that suggests the introductory price in this family is not a temporary gesture but standard launch apparatus: enter the market at a low anchor, hold the target band when the clock runs out. It is an inference; there is no official statement.
3. One price list, two wallets: the Argon–Carbon split
๐ Gemini 4 Carbon, opened to developers on November 10, carries a per-token price identical to Argon's introductory rate — the documentation says "same per-token price" in plain words [9][11]. The split is not in the price row but in consumption: the latest-model page shows Argon at "Moderate" speed and Carbon at "Deliberate"; Argon is described as offering "faster responses and lower token spend per turn," while Carbon, wearing the "strongest reasoning model" title, spends more thinking tokens on hypothesis exploration for hard problems [2][11][12]. So the same price list does not mean the same bill — tokens burned per turn decide the cost per task. Argon is the model of volume and speed, Carbon the model of depth, and the choice between them will, I think, cut deeper on the invoice side after December 1 [1][11].
4. The 1M shared pool: the agentic math of long decode
๐งฎ Argon's technical showcase fits in a single number: a combined context window of 1,048,576 tokens — input, thinking and output share one pool; default synchronous output is 65,536 tokens, and with long decode a single response can stretch to "1M minus input" [1][7]. Generation past 64K splits into continuity "hops" (up to 64K per background/streaming hop, up to 32K non-streaming), and server-side continuity can run for as long as 24 hours [7]. The bill follows the mechanism: because every hop resends everything produced so far as input, total_input_tokens inflates; in Google's own worked example, roughly 397K billed input tokens correspond to 201K of real context [7]. The consolation is this: the repeated input is charged at a 95% discount through the implicit cache, which during the introductory period works out at $0.10 per million [6][7].
๐ฅ️ The documentation itself lists where you shouldn't use it: short multi-turn tool-calling agents, prompts near the input limit, and latency-sensitive work — every turn fills the shared pool, and the process can stretch into the minutes range [7]. The migration notes are equally worth recording: explicit context caching has been removed (implicit only, prompts of at least 4096 tokens), temperature/top_p/top_k are ignored, and a client timeout of at least 15 minutes is recommended [10]. A point of conceptual clarity as well: when the max_output_tokens ceiling is reached the answer is cut off — in the official guide's phrasing, lowering the budget doesn't make the model think less, it cuts it off earlier; for thinking level, the recommendation is to pick, across three tiers (low/medium/high), the lowest one that clears the quality bar [7][10]. The capability table has its borders drawn too: input is multimodal (text, image, video, audio, PDF), output is text only; computer use, code execution and function calling are there, image and audio generation are not [1].
5. The table is not one-sided: where it leads, where it trails
๐ Google's own model card carries an October table with competitor columns for GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5:
| Category | Benchmark | Argon | Astra | Fable 5.1 | Opus 5.5 |
|---|---|---|---|---|---|
| Knowledge work | Vals Index | 68.9% | 63.1% | 65.8% | 67.0% |
| Knowledge work | AutomationBench | 51.3% | 41.4% | 31.4% | 42.5% |
| Knowledge work | Vals Finance Agent v2 | 65.4% | 53.5% | 58.9% | 58.6% |
| Knowledge work | Harvey Legal Agent | 19.6% | 5.4% | 6.7% | 3.8% |
| Agentic coding | DeepSWE v1.1 | 77.9% | 74.1% | 67.4% | 74.2% |
| Agentic coding | FrontierSWE v2 | 55.0% | 65.5% | 56.3% | 62.3% |
| Agentic coding | Terminal-bench 4.0 | 57.4% | 58.2% | 57.9% | 66.4% |
| Science/math | Terminal-Bench Science 0.1 | 57.6% | 68.1% | 52.6% | 63.3% |
| Science/math | RiemannBench | 76.0% | 72.0% | 65.6% | 69.6% |
| Long context | GraphWalks 256k–1M (F1) | 84.2% | 71.8% | 65.0% | 66.8% |
| Computer use | Agent's Last Exam | 39.5% | 34.2% | — | 38.2% |
| Cybersecurity | CWE-bench v1 | 68.0% | 68.0% | 58.0% | 67.0% |
๐ The honest read of the table fits in two lines: Argon leads on knowledge work, legal and finance agents, long context and deep research; Astra and Opus lead on FrontierSWE v2, Terminal-bench 4.0 and science [5]. The independently relayed numbers also refuse to turn this table into a monologue — note: these measurements are second-hand from technical blogs; the primary leaderboard pages were not pulled live in this cycle. Artificial Analysis had it 8th of 223 models at Index 52.56 in launch week; in its October 15 snapshot, about $1.99 per task, a hallucination rate around 15% (against Astra max at 51%) and roughly 62K output tokens per task (Astra max: 27K — the relayers log that gap as a talkativeness tax); on October 28, Argon (high) tied Fable 5.1 xhigh at 50 points, behind Astra max's 57 and Opus max's 58 [16][19][20]. On the Epoch Capabilities Index it sits 5th at 165 points among 260 tracked models — Opus 5.5 at 167, Astra at 166 [18]. CursorBench shows 52.8% at default effort, "on par with Opus 5.5"; Scale AI's Humanity's Sixth Sense has it first at 55.4; LMArena's text side has it top at 1,525 — preliminary though, 4,942 votes — while Code Arena WebDev ranks it 8th at 1,679; there is no backing for a "first in everything" narrative [16][17].
⚠️ The methodology column deserves its lines too: apidog's 19-line analysis argues that five rows of Google's comparison table are mixed-source on both sides and that three of the losses fall in those rows, and that the max-thinking comparison compares ceilings while the API default is medium; machinenotes underlines the "self-reported" warning separately [16][18]. Field observations draw the same scattered pattern: quantslant's own four coding tasks — 51/51 unit tests, $0.028, a 25-second run — earn the verdict "the cheapest model in this class and price I have tested," yet on terminal and repo agent work the independent data points to Opus 5.5; XDA measured Argon 7/16 against Claude Code's 9/16 on the same scrappy side project; kaganerkan says he would pick it "for document, legal and analysis work, not as an autonomous developer" [17][23][24]. All three are single-operator observations with no comparison rubric, which is why they stay at the "testers' eyes" level here.
๐ก️ The cyber side's independent measurements give the same two-faced picture: on Simbian's Cyber Defense Benchmark (October, 32 models) it ranks 3rd at 43.7% coverage — the first top-3 finish for a Google model and the biggest jump over its own generation at +11 points; meanwhile its 65.4% cache-hit rate is the worst of the top five, and "random refusal" behavior is reported [21]. In Averlon's measurements, CTI-REALM comes in at 0.62 (best in the field 0.72); on RealVuln it scores 55.1 with 361 matched open findings — 275 of 536 findings unmatched; the translation: thorough researcher, weak editor [22].
6. The launch backdrop: the calendar and what Reuters reported
๐ฐ️ According to Reuters, this launch is a land-recovery move against OpenAI and Anthropic after months of delay and a DeepMind leadership change (Demis Hassabis stepping aside as CEO, departures from the Gemini team); the cancellation of Gemini 3.5 Pro had been announced in September [4]. Also per Reuters, early-testing customers include Deloitte, LexisNexis, Ambience Healthcare and Isomorphic Labs — those names do not appear in Google's official announcements, which speak in generic "early access program" language with Chrome/ADK/xTB/vLLM cases and one named cybersecurity partner, Wiz; the name list therefore travels only with that attribution [4][8].
๐งพ On the program side, the blog and Reuters agree on two expansions: the top Antigravity quota given free for six months to 10,000 scientists (applications open), and the Fairwind Program — which opens guardrail-free Argon to cyber defenders — spreading to hundreds of new partners [4][8]. The security column: the model card says distribution under safeguards, with a "CCL cannot be ruled out" (corroborated uplift) assessment in biology and cybersecurity; violations in adversarial sets down roughly 2x and shortcut attempts down to 2% are on record [5][8]. An architecture note as well: Argon is no fine-tune of its predecessor — a MoE transformer, native multimodal, trained on TPUs with JAX plus ML Pathways, knowledge cutoff March 2026; the parameter count is not disclosed [5]. ๐️ The family calendar in one line: October 13 Argon → November 10 Carbon → November 30 the intro price's last day → December 1 the new sheet; the main announcement also promises "the rest of the Gemini 4 family" [3][8][12][6].
Murat Karakaya Academy verdict
๐งญ What the data says: this launch has two leaders — on the technical side, the 1M shared pool and a fifth-of-class entry ticket during the introductory window; on the pricing side, a doubling with clear scope and date and no stated reason. Independent measurements put Argon near the top of the table but never alone: 5th on ECI, 8th on AA, first in the text arena on a preliminary result, 8th in the code arena [16][17][18][19][20]. On the legal-finance-document strip, Google's table and the independent relays point the same way; on the autonomous-code side, the independent data keeps pointing at Opus 5.5 [5][17][23][24].
๐ง Our opinion: the pricing strategy is a habit-forming apparatus, and the decision moment is the morning of December 1. If usage keeps its introductory-table pace at the $4/$20 band, Google will have sold not a cheap model but a new price ceiling; if demand contracts at that moment, what is left on our hands is a customer-acquisition-cost experiment. The Gemini 3.8 Flash precedent (same structure, dated December 31) suggests the family trusts this apparatus; because no rationale was published, this entire reading is my interpretation — the data tells you the date and the numbers, not the intent.
๐ฏ Three concrete paths for the reader. One: legal, finance, long-document and deep-research work is the strongest strip in the table, and the 1M pool is a genuine advantage there — but long decode is budget management, not free 1M output; the input resent at every hop keeps showing up on the bill [7]. Two: for terminal- and repo-heavy autonomous coding, the independent data does not point at Argon; XDA's 7/16 is a small sample, yet the direction is consistent — if you are considering tying that work to Argon, run it on your own tasks before December 1 and read your measured cost per task against your post-December-1 $4/$20 bill [17][23]. Three: on latency-sensitive agents, don't make long decode the default; set the thinking level the way the official guide recommends — the lowest tier that clears the quality bar — and clean up the migration notes (implicit-cache threshold, ignored sampling parameters, 15-minute timeout) before porting [7][10].
References
- ๐ ai.google.dev — "Gemini 4 Argon official model page" · accessed Nov 12, 2026
- ๐ ai.google.dev — "Latest Gemini models — pricing/comparison" · accessed Nov 12, 2026
- ๐ blog.google — "Build with Gemini 4 Argon" (Oct 13, 2026) · accessed Nov 12, 2026
- ๐ reuters.com — "Google launches flagship AI model Gemini 4 Argon for developers, businesses" (Oct 13, 2026) · accessed Nov 12, 2026
- ๐ deepmind.google — "Gemini 4 Argon — DeepMind model card" · accessed Nov 12, 2026
- ๐ aistudio.google.com — "Gemini API official pricing page" · accessed Nov 12, 2026
- ๐ aistudio.google.com — "Long decode official documentation" · accessed Nov 12, 2026
- ๐ blog.google — "Introducing Gemini 4 Argon" (Oct 13, 2026) · accessed Nov 12, 2026
- ๐ aistudio.google.com — "Gemini 4 Carbon model page" · accessed Nov 12, 2026
- ๐ aistudio.google.com — "What's new in Gemini 4 Argon" · accessed Nov 12, 2026
- ๐ aistudio.google.com — "What's new in Gemini 4 Carbon (Carbon/Argon choice)" · accessed Nov 12, 2026
- ๐ blog.google — "Gemini 4 Carbon for developers" (Nov 10, 2026) · accessed Nov 12, 2026
- ๐ developers.openai.com — "OpenAI official API pricing page" · accessed Nov 12, 2026
- ๐ platform.claude.com — "Anthropic official model pricing documentation" · accessed Nov 12, 2026
- ๐ anthropic.com — "Claude Fable 5.1 / Mythos 5.1 announcement" · accessed Nov 12, 2026
- ๐ apidog.com — "Independent 19-row analysis + AA/Cursor/Scale relays" · accessed Nov 12, 2026
- ๐ quantslant.com — "Independent coding test (Oct 13 run) + Arena note" · accessed Nov 12, 2026
- ๐ machinenotes.com — "Epoch ECI 165 #5 + self-reported warning" · accessed Nov 12, 2026
- ๐ blog.sandbase.ai — "AA field breakdown (Oct 12 snapshot) + AA null fields" · accessed Nov 12, 2026
- ๐ aihustleguy.com — "Oct 28 AA snapshot comparison table + single-operator observations" · accessed Nov 12, 2026
- ๐ simbian.ai — "Cyber Defense Benchmark #3/32, cache-hit nuance" · accessed Nov 12, 2026
- ๐ averlon.ai — "CTI-REALM/RealVuln/VulnGym independent measurements" · accessed Nov 12, 2026
- ๐ xda-developers.com — "Claude Code vs Argon field test on the same project" · accessed Nov 12, 2026
- ๐ kaganerkan.com — "Independent review: strong on documents/analysis, weak on autonomous code" · accessed Nov 12, 2026
Hashtags: #MuratKarakayaAkademi #Gemini4Argon #ArtificialIntelligence #LLMPricing #ReasoningModels #GoogleDeepMind