Gemini 4 Argon went generally available on October 13 across the paid Gemini API, AI Studio, OpenRouter, Vertex, Antigravity for Ultra subscribers, and the Gemini app for AI Pro and Ultra. The number traveling with it: Artificial Analysis scores Argon (high) and GPT-6 Astra (max) level at 53 on its Intelligence Index, while Argon costs $1.99 per index task to Astra's $3.26: roughly 60%.
That 60% is real. It also comes with two pieces of fine print that most of the sharing skipped: the price that produces it expires on November 30, and Argon burns more than twice as many tokens per task to get the same score. Here's the full math, plus the row the frontier comparison tables usually leave out.
Disclosure: Wallaby Token sells API access to Kimi K3, the open-weight model in the last row below. Benchmark figures are attributed to their sources; prices were checked October 16, 2026.
The 60% claim, with dates attached
Argon's introductory API price is $2 per million input tokens and $10 per million output, with cached input at $0.10. That introductory window ends November 30. From December 1, the price doubles to $4 / $20 ($0.20 cached).
Artificial Analysis has already run the forward math: at the December price, Argon's cost per index task rises to about $3.98, roughly 1.2x Astra, not 0.6x. Same model, same score, opposite conclusion. If you're choosing a default model for a workload that runs into December, the date matters more than the headline.
The verbosity tax
The second fine print is why the per-token and per-task rankings diverge. On Artificial Analysis's index, Argon generated about 62,000 tokens per task; Astra used 27,000. Argon's input and output rates are each a fifth of Astra's right now (the cached-input gap is tenfold: $0.10 vs $1.00), but it writes answers more than twice as long, which is how $2/$10 per million turns into $1.99 per task instead of something far lower. Reverse-engineering AA's per-task figures makes the gap starker: Argon also reads roughly 685K input tokens per task against Astra's 191K, a 3.6x difference on top of the 2.3x output gap.
For chat use this is a style preference. For agent loops, where output tokens cost five times input tokens on most rate cards and every turn resends context, verbosity is a bill multiplier. It's also a latency number: one third-party test on GA day (23 tasks, automated grading) clocked Argon at 134.6 seconds average against Astra's 200.1, so faster isn't the issue. Longer is.
Head to head
| Gemini 4 Argon | GPT-6 Astra | Kimi K3 (open weight) | |
|---|---|---|---|
| AA Intelligence Index | 53 | 53 | below the frontier pair |
| Price per 1M (in / cached / out) | $2 / $0.10 / $10 → $4 / $0.20 / $20 from Dec 1 | $10 / $1 / $50 | $3 / $0.30 / $15 list; $2.70 / $0.27 / $13.50 on Wallaby, ongoing |
| AA cost per task | $1.99 → ~$3.98 in Dec | $3.26 | n/a (different workload profile) |
| Tokens per AA task | ~62,000 | ~27,000 | n/a |
| Wins (per Google's/AA's tables) | Vals Index 68.9, AutomationBench 51.3, Harvey Legal, LVBench 91.7; 1M-token outputs | FrontierSWE v2 (+10.5), OSWorld-2.0 72.6%, FrontierMath T4 97.6% | n/a |
| Lock-in | Google only | OpenAI only | Checkpoint is public; runs anywhere |
Wallaby's Kimi K3 rate is 10% off the official list, with no expiry date on the discount.
Where that leaves the choice
Pick Argon for legal, finance, long-document and automation work — its benchmark wins are concentrated there, and through November the price is genuinely excellent. Just budget at the December rate, not the October one.
Pick Astra for long project-scale coding, computer use, and formal math, and also if you want shorter answers and predictable bills, since its price isn't scheduled to move.
Where Kimi K3 fits: below both on the frontier index, honestly so, at a flat rate that undercuts Argon's December price by a third and Astra by three-quarters, with no expiry date attached to any of it. For the high-volume workloads most teams actually run, from agents and CI bots to maintenance loops and drafting, the open-weight tier is where the cost-per-task war gets settled after the introductory discounts evaporate. We publish our cross-provider Kimi K3 pricing comparison and keep it current for exactly this reason.
Introductory pricing is a launch tactic, not a rate card. The interesting question was never "is Argon 60% of Astra" — it's what each model costs on the day the discount ends.
A note on dates: prices and availability as of October 16, 2026, with Argon's December 1 repricing per its published schedule. Benchmark figures are from Artificial Analysis and the cited third-party test; verify the current rate cards before budgeting.
Want to run the comparison on your own workload? A new Wallaby Token account comes with $0.50 in trial credit — enough to benchmark Kimi K3 on a real task against whatever frontier model you're currently paying introductory prices for. The minimum top-up is $20, and the rate stays 10% off the official list, ongoing.