Comparisons
Gemini 4 Argon API Pricing: The Intro Rate Doubles on December 1
Verdict: Gemini 4 Argon's API pricing is real — $2 per 1M input tokens and $10 per 1M output, which I pulled straight from OpenRouter's live model API today. But it's a promo. On December 1 it doubles to $4/$20. Stacked against DeepSeek V4.1-Flash on a 50/50 input-output mix, Argon runs about 8x more expensive today and 16x after the promo dies. You're paying that multiplier for a first-tier model — Artificial Analysis has Argon High at 53 points, tied with GPT-6 Astra Max and Claude Fable 5.1 Max. Whether that's worth it depends entirely on whether your task actually needs first-tier. Most API workloads don't.
Google's Gemini 4 line finally shipped in mid-October, and Argon is the first model out. The pitch is agentic: long-horizon task execution — the model reads, calls tools, edits code, checks its own work across many steps instead of answering in one shot. The demos Google pushed hardest were multimodal (3D scene generation from a text prompt, simulation-to-real robot control via Apptronik's Apollo). Skeptical takes exist too: Ethan Mollick, who got early access, noted it still struggles with complex multi-agent tasks and is "occasionally overconfident, occasionally lazy." Google also published a 217-page safety report alongside launch, which tells you how long this thing sat in review — the Pro version slipped repeatedly before Argon went out the door.
I'm not going to relitigate the benchmarks; plenty of launch-week coverage did. What almost nobody covered is the pricing structure — specifically what happens when the intro rate expires. So here's the ledger.
The price table, verified today
These Argon numbers came from OpenRouter's models API on October 19, 2026 — not from a launch blog post:
| Input / 1M | Output / 1M | |
|---|---|---|
| Gemini 4 Argon (standard) | $2.00 | $10.00 |
| Gemini 4 Argon (:batch) | $1.00 | $5.00 |
| Argon cache read | $0.10 per 1M | |
| Argon cache write | $0.1875 per 1M | |
| DeepSeek V4.1-Flash (peak) | $0.30 | $1.20 |
| DeepSeek V4.1-Flash (off-peak) | $0.15 | $0.60 |
| DeepSeek cache hit (peak / off-peak) | $0.006 / $0.003 per 1M | |
DeepSeek's numbers are pulled live from their official pricing page today. One structural note: DeepSeek has no batch endpoint — their off-peak window (roughly 16:30–00:30 UTC) plays that role, at half price for whatever you can defer. Both models run 1M-token context. Argon reportedly supports output completions up to 1M tokens (per Google Cloud's agent platform docs), against DeepSeek's 384K max output — if you generate book-length outputs, that's a real difference.
The $4/$20 post-promo price isn't something I could verify against a Google-run page — ai.google.dev is unreachable from my network, and I won't dress that up. It's consistent across three independent reports (QuantumBit, MindStudio, aipricing.guru), all citing the December 1 revert. Treat it as strongly sourced but secondhand.
The math nobody printed
Take a 50/50 input-output mix per 1M total tokens (500K each — a reasonable proxy for chat and most agent loops):
- Argon, intro rate: $1.00 + $5.00 = $6.00
- Argon, after December 1: $2.00 + $10.00 = $12.00
- DeepSeek Flash, peak: $0.15 + $0.60 = $0.75 — that's 8x cheaper than the intro, 16x after
- DeepSeek Flash, off-peak: $0.075 + $0.30 = $0.375 — 16x today, 32x after
- Argon batch: $0.50 + $2.50 = $3.00 — still 4x DeepSeek at full peak price
That's the deal in one line: Argon's discounted price is roughly DeepSeek's undiscounted price multiplied by 8. The promo makes it tolerable; December 1 removes the anesthesia.
The gap gets worse where it matters most for agent workloads: caching. If you're running a coding agent, the same system prompt and file context get re-read thousands of times. DeepSeek charges $0.006 per 1M cached-input tokens at peak; Argon charges $0.10 — a 17x difference on the exact token class that dominates long agent sessions. QuantumBit's writeup relayed Google's own cost claim that Argon High averages $1.99 per task, about 39% cheaper than GPT-6 Astra — fine, but that's a comparison inside the first tier. The tier below is priced on a different axis entirely.
So who should actually pay for Argon?
My honest take, with the caveat that I haven't benchmarked Argon myself — this is price-table math plus third-party evals, not hands-on testing:
Argon is worth its price when the alternative is failing, not when it's merely cheaper. If your tasks sit in the band where GPT-6 Astra or Claude Opus 5.5 currently win and DeepSeek Flash genuinely drops them, the 8x is buying capability you can't get otherwise — and at $2/$10 it's the cheapest entry yet into that tier, roughly half of Opus 5.5 and a fifth of Astra per QuantumBit's relayed figures. If your workload tolerates overnight deferral, DeepSeek's off-peak at $0.375 per mixed 1M is absurd value that no first-tier model is chasing.
The trap is the middle: teams paying 8x for "better" on tasks where better was never the bottleneck. If you can't name the specific runs where Flash fell short, you're buying a reputation, not solving a problem. And if you're signing up for Argon because of $2/$10, calendar the December 1 revert now — the bill you approved in October is not the bill you'll get in December.
One more thing worth watching: if the post-promo $4/$20 sticks, Argon lands almost exactly on Gemini 3.8 Flash's old multiple — Google pricing its flagship at a premium to everything beneath it and letting the intro discount do the persuading. I'd expect a price move from the Chinese labs within weeks if first-tier demand shifts. DeepSeek's off-peak pricing has survived every competitor launch since June; December 1 is the next real test of whether that changes.
Dated caveat: prices verified October 19, 2026 via OpenRouter's live models API (Argon) and DeepSeek's official pricing page (Flash). The $4/$20 December 1 revert is triple-sourced secondhand reporting, not verified against a Google page. Benchmarks and cost-per-task figures are relayed from QuantumBit's launch coverage. Pricing changes; check the official pages before committing spend.
Sources: OpenRouter models API (google/gemini-4-argon, queried 2026-10-19); DeepSeek API docs pricing page (queried 2026-10-19); 量子位 Gemini 4 launch coverage; MindStudio Gemini 4 Argon pricing overview; aipricing.guru Google AI pricing tracker.
Related: DeepSeek Off-Peak Pricing Explained · DeepSeek's Search API: $0.75/1K and a Free Month · Gemini 3.8 Flash Pricing · DeepSeek V4 API Cost Calculator · Ling 3.1 Flash: Open Source, $0.30/$0.90