Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens on Google's standard paid tier through 30 November 2026, then $4 and $20 from 1 December 2026, according to Google's pricing page as of 14 October 2026. Cached input is $0.10 per million (then $0.20), batch and flex tiers are half the standard rate, and priority is $3.60/$18 (then $7.20/$36). For a typical blog draft that works out to under five cents at introductory prices, so the budget risk is not content drafting; it is long agent runs and reasoning tokens.
Key Takeaways
- Standard: $2 in / $10 out through 30 Nov 2026, then $4 / $20, per Google's pricing page. Output includes thinking tokens.
- Batch and Flex: $1 / $5 through 30 Nov, then $2 / $10. Priority: $3.60 / $18, then $7.20 / $36.
- Context caching is implicit only (explicit caching is not supported), so cache savings are not guaranteed on any given request.
- Argon is not available on the Free tier or the Google AI Pro plan for the API; it is a paid-tier model.
- A 3,000-token draft costs about $0.05 now and about $0.09 after the promotion; a 30-step agent run can cost 30 to 70 times more.
- Plan for the December price doubling now, not after the invoice arrives.
The price sheet
These figures are from Google's Gemini API pricing page, paid tier, USD per million tokens, as read on 14 October 2026. The page is long and changes; recheck before budgeting.
| Tier | Input (to 30 Nov) | Output (to 30 Nov) | Input (from 1 Dec) | Output (from 1 Dec) |
|---|---|---|---|---|
| Standard | $2.00 | $10.00 | $4.00 | $20.00 |
| Batch | $1.00 | $5.00 | $2.00 | $10.00 |
| Flex | $1.00 | $5.00 | $2.00 | $10.00 |
| Priority | $3.60 | $18.00 | $7.20 | $36.00 |
Cached input: $0.10 (standard) and $0.05 (batch/flex) to 30 November; $0.20 and $0.10 from 1 December; priority is $0.18 then $0.36. The page lists no storage price per hour for Argon. Search grounding gets 5,000 free requests per month shared across Gemini 3 and newer models, then $14 per 1,000 requests, as does Maps grounding.
Three independent write-ups confirm the headline rates and the 30 November date: DealnTech, 9to5Google and Tech Insider. One timing wrinkle: Artificial Analysis, writing on launch day, described the 50% discount as lasting "at least one month" with the end date unconfirmed, per its Argon report. The 30 November date came later, so use the later date and the pricing page.
What "output" really includes
Google's table labels output as including thinking. A reasoning model bills the hidden thinking tokens at the output rate, so your invoice can exceed what you see on screen. Artificial Analysis reports that Argon averages 62K output tokens per task on its Intelligence Index compared with 27K for GPT-6 Astra, and says its lower cost per task comes from lower token prices, not from using fewer tokens. If you plan from visible word counts, you will under-forecast.
Google's launch post also describes a 1M output-token limit, up from 64K. DealnTech notes it is unclear how much of that budget a single response can use. A larger ceiling means a larger possible bill from one runaway call, so set a maximum output limit in your code.
Worked examples
All scenarios below are my own illustrations with assumed token counts, not measurements. Replace the token numbers with your logs. Arithmetic uses the standard tier.
1. One blog draft
Assume 8,000 input tokens (brief, style guide, outline) and 3,000 output tokens.
- Until 30 Nov: 8,000 x $2/M = $0.016; 3,000 x $10/M = $0.030. Total $0.046.
- From 1 Dec: $0.092.
- Batch tier: $0.023 now, $0.046 after.
If thinking adds 6,000 hidden tokens, output becomes 9,000: $0.016 + $0.090 = $0.106 now. That is the part people miss.
2. 200 posts a month
200 drafts at the scenario-1 figure: $9.20 now, $18.40 from December. Still small next to editor salary.
3. 500 meta descriptions in bulk
Assume 1,500 input and 100 output tokens per page, so 750K input and 50K output in total.
- Standard now: $1.50 + $0.50 = $2.00; December $4.00.
- Batch now: $1.00; December $2.00.
For anything that can wait, batch halves the bill. Whether Argon is the right model for a rule-bound job is a separate question; my Haiku vs Sonnet routing post argues smaller models often are enough.
4. A style guide sent 1,000 times
A 4,000-token guide across 1,000 requests is 4M input tokens.
- Uncached, standard now: $8.00; December $16.00.
- If every call hit the cache: $0.10/M gives $0.40; December $0.80.
Caveat: Google says caching is implicit only, so you cannot force a hit. Treat the cached price as a best case and measure your real hit rate from usage metadata. For comparison, Claude Sonnet 5.5's cache read is $0.20 per million per my Sonnet 5.5 cost post.
5. A 30-step research agent
Agents resend their growing context at every step. Assume 20,000 input tokens per step and 1,500 output tokens per step over 30 steps: 600K input and 45K output.
- Standard now, no caching: $1.20 + $0.45 = $1.65; December $3.30.
- With 80% of input cached (assumed): 480K x $0.10 = $0.048; 120K x $2 = $0.24; output $0.45. Total $0.74 now; $1.48 from December.
If that agent runs 40 times a week, the no-cache scenario is $66 a week now and $132 after. Add grounding charges on top if the agent searches heavily: past 5,000 free requests a month, $14 per 1,000. At 8,000 grounded requests a month that is 3,000 x $14/1,000 = $42.
6. Priority for a live tool
Say a customer-facing assistant answers 10,000 queries a month at 3,000 input and 500 output tokens: 30M input, 5M output.
- Standard now: $60 + $50 = $110.
- Priority now: $108 + $90 = $198; December $396.
Pay for priority only where latency is a measured problem.
Comparing with the other two models
Using the same scenario-1 job (8,000 in, 3,000 out) and list prices:
| Model | Input / output per 1M | Cost per draft |
|---|---|---|
| Argon (to 30 Nov) | $2 / $10 | $0.046 |
| Argon (from 1 Dec) | $4 / $20 | $0.092 |
| GPT-6.1 Sol | $2 / $10 (per ProPakistani and LLM Stats) | $0.046 |
| Claude Opus 5.5 | $4 / $20 (per Anthropic) | $0.092 |
Per-token parity is not per-task parity: if one model thinks twice as long, its real cost per draft doubles. The framework for this is in my model comparison.
Who actually pays these rates
- Free tier: Argon is not available, per the pricing page.
- Google AI Pro plan: The pricing page says Argon is unavailable on that plan for API use, though Pro subscribers can use it in the Gemini app per Google's blog. App access and API billing are separate.
- Paid API tier: Yes, pay-as-you-go.
- Enterprise: Google says enterprise customers can use it through Gemini for business and Antigravity; I have no published enterprise price.
A budgeting template
Copy this into a sheet.
| Workflow | Runs/month | Avg input tok | Avg output tok (incl. thinking) | Cache hit % | Tier | Cost now | Cost from 1 Dec |
|---|
Formula per row, standard tier: runs x [(input x (1 - cache hit) x in-rate) + (input x cache hit x cache-rate) + (output x out-rate)] / 1,000,000.
Guardrails that save money
- Set a hard
max outputper call. - Log input, output and thinking tokens for every run.
- Cap agent steps and ask for a human check after step 10.
- Use batch for anything not time-sensitive.
- Alert when daily spend passes twice its 7-day average.
- Re-run the sheet on 1 December with the new rates.
When not to use Argon
- Bulk rules-based tasks where a small model passes your test.
- Anything with sensitive customer data you have not cleared with your own data policy.
- Workflows that need guaranteed explicit caching, since Argon supports implicit only.
- Regions or accounts where access is not confirmed. I have not verified API availability in India; check your console.
FAQ
How much does Gemini 4 Argon cost per million tokens?
On the standard paid tier it is $2 input and $10 output through 30 November 2026, then $4 and $20 from 1 December. Cached input is $0.10 then $0.20. Source: Google's pricing page as read on 14 October 2026.
Is there a cheaper way to run Argon?
Batch and Flex tiers are half the standard price: $1/$5 now and $2/$10 from December. They suit jobs that can wait. Priority costs more, at $3.60/$18 now.
Does Argon's price include thinking tokens?
Google's table labels output as including thinking. Reasoning tokens therefore bill at the output rate. Log them separately so you can see them.
How does context caching work for Argon?
Google's page says caching is implicit only and explicit caching is not supported. You cannot guarantee a cache hit, so measure your hit rate before counting on the $0.10 rate.
Can I use Argon on the free API tier?
No. The pricing page marks it unavailable on the Free tier and on the Google AI Pro plan for the API. It is a paid-tier model.
When does the introductory price end?
On 30 November 2026, per Google's pricing page and several trade reports. Launch-day coverage said the end date was unconfirmed, so rely on the pricing page.
Is Argon cheaper than Claude Opus 5.5?
Until 30 November, yes: $2/$10 against $4/$20 per million tokens. From December the list prices match. Cost per task still depends on how many tokens each model uses.
What does grounding with Google Search cost?
The pricing page lists 5,000 free requests per month shared across Gemini 3 and newer models, then $14 per 1,000 requests. Maps grounding has the same structure.
What is the biggest hidden cost?
Hidden reasoning tokens and long agent loops. A single draft costs cents, but a 30-step agent with a growing context can cost dollars per run. Cap steps and output.
Need a cost-aware AI workflow?
I'm Younus Fardeen, an organic growth, SEO and AEO strategist with 4+ years of marketing experience in edtech and startups. If you want help building content and agent workflows that are measured by cost per accepted asset rather than hype, see my work and contact me through the form at younusfardeen.in. Prices are as of 14 October 2026 and will change; always confirm against Google's current pricing page.