The call
The model was announced on 30 September for a closed group of cyber defenders and moved to general availability on 13 October (independent, Tech AI Wire, citing the Gemini API changelog; The AI Insider). It topped Hacker News the same day at 924 points (Zeli digest).
What it is, and what it is not
What it is. Google DeepMind’s top model, generally available on the paid Gemini API, AI Studio, Antigravity, and eligible Gemini business subscriptions. One 1,048,576-token window shared by input, thinking, and output; up to about a million output tokens on background requests; function calling, code execution, search grounding, and structured outputs (vendor, via Tech AI Wire). It ranks first on the Vals Index of finance, legal, tax, and coding work at 68.9%, against 66.97% for Claude Opus 5.5 (independent, Vals AI).
What it is not. It is not the best model on every job: on Artificial Analysis’s Intelligence Index it scores 53 against Opus 5.5’s 58, and it trails Opus on Terminal-Bench 4.0, 57.4% to 66.4% (independent, Artificial Analysis, and VentureBeat’s table via Tech AI Wire). It is not on the free tier. And the launch price is not the price: output and input both double on 1 December (vendor, via Tech Insider).
The number nobody is quoting
Everyone is quoting $2 and $10, the same per token as GPT-6.1 Sol and half of Claude Opus 5.5 (independent, eesel). Per token is the wrong unit. Argon thinks long: it used about 110 million output tokens to run the Artificial Analysis index, against a median of 85 million for its price tier, 29% more (independent, Artificial Analysis; our maths). Output prices include thinking tokens.
| Cost per task | To 30 Nov | From 1 Dec | Label |
|---|---|---|---|
| Argon, Artificial Analysis index task | $1.99 | $1.99 × 2 = $3.98 | independent; our maths |
| Claude Opus 5.5, same index | $1.82 | $1.82 | independent |
| Argon against Opus 5.5 | 1.99 ÷ 1.82 = 1.09x | 3.98 ÷ 1.82 = 2.19x | our maths |
| Argon, Vals Index test | $15.68 | $15.68 × 2 = $31.36 | independent; our maths |
| Claude Opus 5.5, Vals Index test | $32.14 | $32.14 | independent |
| Argon against Opus 5.5 | 15.68 ÷ 32.14 = 0.49x | 31.36 ÷ 32.14 = 0.98x | our maths |
Opus 5.5 per-task figure from Artificial Analysis as reported by eesel; Vals figures from Vals AI. December figures assume the same token use at double the price.
Read it both ways. On Artificial Analysis’s mixed workload, Argon is already 9% dearer per task than Opus 5.5 at the launch price, and 2.2 times dearer from 1 December. On Vals’s professional workload, the launch price halves the bill, and December takes the whole saving back: 0.98x, parity. Either way, the 50% saving in the headline lasts 47 days from today.
At 10,000 agent tasks a month on the Artificial Analysis mix, that is $19,900 in November and $39,800 in December for the same work, against $18,200 on Opus 5.5 (our maths). A pilot priced in October will be wrong by half in seven weeks.
Missing
What Google has not told an enterprise buyer, as of 14 October:
- Vertex AI and Gemini Enterprise Agent Platform availability. GA coverage names the Gemini API and AI Studio; Vertex is not mentioned (Tech AI Wire, Tech Insider).
- Data residency and regions for Argon specifically.
- Rate limits and quotas at GA.
- A long-context surcharge, or confirmation there is none, across the shared million-token window.
- Provisioned throughput for steady enterprise load.
- Independent latency. Artificial Analysis lists output speed as unknown.
- Salesforce support. We found no Salesforce announcement of Argon in Agentforce or Model Builder.
Who owns what
The price schedule is Google’s job. The token count is ours.
In a Salesforce stack, Argon sits behind an Apex callout through a Named Credential, or behind whatever Google route your Agentforce or Data Cloud set-up already uses. Either way, the bill is set by how many tokens each case, summary, or agent step consumes, and the thinking level you choose (low, medium by default, or high) moves that number more than the list price does. Log input, thinking, and output tokens per request from day one, in a custom object or your observability stack, so the December invoice is a forecast, not a surprise.
Monday morning
- Step 1. Re-price every model line at $4 and $20. Any business case or forecast using Argon’s launch price gets the December rate. Owner: finance partner for AI.
- Step 2. Shadow-test on 200 real tasks. Run Argon next to your current model on documents and agent steps you already grade, on the paid API, with no customer data until residency is confirmed. Owner: head of AI platform.
- Step 3. Measure cost per finished task. Tokens in, thinking, and out, times the December price, divided by tasks that passed review. Owner: AI engineering lead.
- Step 4. Test the thinking levels. Run the same 200 at low and medium; keep high only where accuracy pays for it. Owner: AI engineering lead.
- Step 5. Ask Google for the enterprise terms in writing. Vertex availability, residency, quotas, and provisioned throughput. Owner: CIO or procurement.
We move this to WATCH if Google holds Vertex availability back past 30 November. We drop the price warning if Google extends the launch rate or publishes Argon cost per task below Opus 5.5 at $4 and $20.
Frequently asked questions
How much does Gemini 4 Argon cost?
$2 per million input tokens and $10 per million output until 30 November 2026, then $4 and $20 from 1 December. Cached input is 95% off. Output prices include thinking tokens, and Argon uses more of them than most models in its tier.
Is Gemini 4 Argon cheaper than Claude Opus 5.5?
Per token, yes, until 30 November. Per task, it depends on the workload: on the Vals Index it costs about half until then and the same afterwards; on the Artificial Analysis index it already costs 9% more, rising to 2.2 times from December.
Should we pilot Gemini 4 Argon now?
Yes, as a shadow test on work you already grade, budgeted at the December price and measured as cost per finished task. Keep customer data out until Google confirms residency and Vertex availability for Argon.
Living page. Last updated 14 Oct 2026. What would change our call: Google confirming Vertex availability and residency, or extending the launch price past 30 November.