Google put a clock on its cheapest frontier pricing. Gemini 4 Argon went
generally available on October 13, 2026 at $2/$10 per million tokens - and
on December 1 that becomes $4/$20. You have roughly seven weeks of cheap
runway. The teams that win are not the ones who read the pricing page.
They are the ones who prototype now, measure their real per-task cost,
and decide whether Argon earns a place in their stack before the price
doubles. Here is the checklist to do exactly that.
The short answer
Gemini 4 Argon (gemini-4-argon) is Google's new frontier model for deep
multi-step work, generally available since October 13, 2026 at $2 input
and $10 output per million tokens. On December 1, 2026, both rates double.
If you are building anything on it, prototype before November 30, measure
tokens per task with thinking tokens counted as output, and re-price your
workload at $4/$20 before you commit to production.
What actually changed on October 13
Google unveiled Gemini 4 Argon on September 30, 2026, but kept public
access restricted to cybersecurity partners in its Fairwind Program. On
October 13, Koray Kavukcuoglu, Google DeepMind's senior vice-president
and Google's chief AI architect, announced the model was "rolling out
today globally" through the paid Gemini API, Google AI Studio, Antigravity
(for Ultra subscribers), the Gemini app and AI Mode in Search (for Pro
and Ultra subscribers), Android Studio, Firebase AI Logic, and Gemini
Enterprise. It is not available on the API's free tier.
Attribution and honesty
The benchmark figures in this post come from two buckets. Google's
launch claims (Vals Index, ERQA, prompt-injection resistance) are
vendor-reported and have not been independently reproduced. The
Artificial Analysis and Terminal Bench 4 scores come from independent
testing reported by outlets covering the launch. Pricing was
cross-checked across VentureBeat, Reuters, the rundown.ai, and datanorth.ai
on October 14-15, 2026. Google's own posts do not prominently state the
December 1 increase; the outlets report it as confirmed.
Argon is the first model in the Gemini 4 generation. Google positions it
for long-horizon, multi-step tasks in coding, finance, legal work,
cybersecurity, and science, with a unified 1-million-token context window
shared across input, thinking, and output, long decode up to 1 million
output tokens, and tunable thinking levels (low, medium, high).
The pricing, verified
Every number below is USD per million tokens, for the Gemini API:
| Tier | Input | Output | Cached input | From December 1 (in / out / cached) |
|---|---|---|---|---|
| Standard | $2.00 | $10.00 | $0.10 | $4.00 / $20.00 / $0.20 |
| Batch and Flex | $1.00 | $5.00 | $0.05 | $2.00 / $10.00 / $0.10 |
| Priority | $3.60 | $18.00 | $0.18 | $7.20 / $36.00 / $0.36 |
Three things matter here. First, output costs five times input at every
tier, and output billing includes thinking tokens - the model's internal
reasoning. A reasoning-heavy agent run is billed almost entirely on the
output side. Second, cached input is a 95% discount: repeated system
prompts, retrieved context, and conversation history can cost $0.10 per
million instead of $2. Third, Batch and Flex halve every rate, at the cost
of latency flexibility.
- 2x: price change on December 1, 2026 (Every token class, every tier)
- $40 -> $80: a 10M-input / 2M-output workload (Illustrative example; pricing sources confirm the math)
- 95%: discount on cached input (Repeated context billed at $0.10/$1M until Nov 30)
The math in that middle stat is worth doing once, because it is the shape
of your December decision. Ten million input tokens plus two million
output tokens costs $40 at the launch price (10 x $2 + 2 x $10) and $80
at the December price (10 x $4 + 2 x $20). A fixed token example of 20,000
uncached input tokens plus 5,000 billed output tokens totals $0.09 before
the hike and $0.18 after. Now scale that to your request volume and you
have your personal answer to whether Argon fits your stack.
What your workload actually costs at 100K requests a month
Per-token rates tell you nothing until you know your token profile. Here
is an illustrative model with its assumptions stated, for a customer
support agent running on Argon. Assumption: 2,500 input tokens (ticket,
retrieved docs, system prompt) and 400 output tokens per ticket, 100,000
tickets a month.
- Per ticket at launch: 2,500 x $2/$1M + 400 x $10/$1M = $0.005 + $0.004 = $0.009
- Per ticket from December 1: 2,500 x $4/$1M + 400 x $20/$1M = $0.01 + $0.008 = $0.018
- Monthly bill at 100K tickets: $900 -> $1,800
That is the honest shape of the hike: nothing about your code changed, and
your bill doubled. Two things pull that number down in practice. Cached
input at 95% off collapses the 2,500-token figure if most of it is
repeated context, and Batch/Flex pricing halves everything for work that
does not need instant answers. Two things push it up: thinking tokens
bill as output, so a heavy-reasoning agent multiplies the 400-token
assumption fast, and retries, tool calls, and guardrail loops bill every
extra request.
The same bill math applies to Claude's pricing tiers - I ran it for Haiku
5.5's 100K-token price cliff when
Anthropic changed the agent-cost game. The method is the same: find your
per-task token profile, price it at both rates, and decide before the
deadline forces the decision on you.
How Argon compares once the intro ends
The December price is not arbitrary: it parks Argon exactly on top of
Claude Opus 5.5 and Meta Muse Wonder. From December 1, the choice is task
fit, not price.
| Model | Input / 1M | Output / 1M | Standing |
|---|---|---|---|
| Gemini 4 Argon (through Nov 30) | $2 | $10 | Launch price |
| Claude Sonnet 5.5 | $2 | $10 | Matches Argon's intro price |
| GPT-6.1 Sol | $2 | $10 | Matches Argon's intro price |
| Gemini 4 Argon (from Dec 1) | $4 | $20 | Standard price |
| Claude Opus 5.5 | $4 | $20 | Matches Argon's standard price |
| Meta Muse Wonder | $4 | $20 | Matches Argon's standard price |
| GPT-6 Astra | $10 | $50 | The expensive end of frontier |
At the launch price, Argon sits exactly where Sonnet 5.5 and GPT-6.1 Sol
sit. At the December price, it sits exactly where Opus 5.5 sits - and
datanorth.ai's analysis makes the blunt point: a business case built on $2
and $10 is only valid for seven weeks. Any contract, quote, or unit
economics you sign now must survive $4/$20.
Argon earns the slot
- Deep multi-step work where the 1M unified context pays off: repo-scale refactors, long legal or finance documents, multi-file agent runs
- Workloads where cached context dominates: 95% off repeated input shrinks the real bill far below the headline rate
- Batchable offline jobs: Batch and Flex halve every rate, and background agents rarely need instant answers
- Prototypes built before November 30 that cleared your acceptance bar on the cheap rate and still clear it at $4/$20
Route the work elsewhere
- Short one-shot tasks where a $0.10/M-input model suffices - Haiku-class models do classification, extraction, and summaries at a fraction of the price
- Token profiles where output dominates: five-times-output billing punishes heavy-thinking runs, so compare against models with cheaper reasoning first
- Anything still on the API free tier - Argon is not there, and there is no announced date for it
- Budgets signed on $2/$10 economics that nobody re-checked for December
The benchmark picture, honestly
Google's claims for Argon are strong: first place on the Vals Index
(finance, coding, legal, tax), 87% on the ERQA embodied-reasoning test,
prompt-injection attacks succeeding only 0.7% of the time across 15
adaptive attempts on Gray Swan's test, and the lowest benchmark-cheating
rate of any Gemini model at 2%. Google also cites internal engineering
wins - a 6-million-line Chromium refactor finished in weeks, and +35%
agentic code submissions with a third fewer rollbacks in Antigravity.
Those are vendor numbers. Treat them as directional.
The independent numbers are cooler. Argon scored 53 on the Artificial
Analysis Intelligence Index - level with GPT-6 Astra and Claude Fable 5.1,
behind Claude Opus 5.5 at 58 and Sonnet 5.5 at 56. On Terminal Bench 4 it
scored 57%, behind Sonnet 5.5 at 64%, Opus 5.5 at 60%, and Astra at 59%.
Nothing about these scores says "run away from Argon." They say the
model is a genuine frontier contender that has not yet proven a gap over
the competition - which is exactly why the intro pricing exists, and
exactly why you should test your own tasks instead of trusting any
leaderboard.
The build-and-measure checklist
This is the whole post in one framework. Do it in order, before November
30, and you will enter December knowing - not guessing - what Argon costs
you.
1. Confirm your access path and billing surface
Decide where you will run Argon: the paid Gemini API in AI Studio, or
through GitHub Copilot, which bills it at provider list pricing in AI
credits (1 credit = $0.01). Copilot's page flags the rates as
promotional through December 1 but does not say what they become after.
Do not assume Copilot follows Google's schedule - budget for a possible
increase on both surfaces.
2. Lock your baseline numbers today
Write down what you run now: the model, the price, and your best
estimate of monthly input/output tokens. Without a baseline, the intro
price feels cheap and the December price feels expensive, and neither
feeling is data.
3. Instrument per-task token logging
Log input, cached input, and output tokens per task from day one. Break
output out by thinking vs final answer if the API exposes it, because
thinking tokens bill as output and they are where agentic workloads
hide their cost. You cannot optimize what you never measured.
4. Prototype before November 30, on the intro price
This is what the seven-week window is for. Build the representative
workflow - the one with real retries, real tool calls, real context
sizes - and run it at production-shaped volume on $2/$10. A benchmark
demo is not a measurement.
5. Run the same workload at December rates
Take the token profile from step 4 and re-price it at $4/$20. If the
December bill still beats your baseline model's bill for equal quality,
commit. If it only works at the intro price, your business case expires
on November 30 - decide now, not in January.
6. Pull the three cost levers
First, cache: get repeated system prompts and retrieved context into
the 95%-off cached tier. Second, batch: move background agents to Batch
or Flex and halve the rate. Third, compress: shrink prompts, cap
thinking levels (low/medium/high are tunable), and watch the
thinking-token line - it is output-priced.
7. Benchmark two alternatives on your tasks, not leaderboards
Run your representative workload on at least one cheaper option (Sonnet
5.5 or GPT-6.1 Sol at $2/$10) and, if quality is the question, on Opus
5.5 at $4/$20. Record errors, retries, review time, and total billed
usage. From December 1, Argon and Opus cost the same - buy the one that
does your work best.
8. Set a budget alarm and a November 29 decision date
Put a spend alert on the project now, and put a calendar block on
November 29. That block is your decision meeting: keep Argon, move the
workload, or renegotiate. The teams that get hurt by price hikes are
the ones who never scheduled the decision.
Where to build, measure, and compare
Every tool below is a real link to the official source. If you build on
Argon, these are the dashboards you will live in.
| Tool | What it does | Use it for |
|---|---|---|
| Gemini API / AI Studio | Official API docs, pricing page, and playground | Building on Argon; verifying the $2/$10 -> $4/$20 schedule |
| Anthropic API | Claude API docs and pricing | Benchmarking against Sonnet 5.5 ($2/$10) and Opus 5.5 ($4/$20) |
| OpenAI Platform | OpenAI API docs and pricing | Benchmarking against the GPT-6 family, including GPT-6 Astra ($10/$50) |
| GitHub Copilot | AI coding assistant with per-model billing | Running Argon through Copilot at provider list pricing, billed in AI credits |
| Artificial Analysis | Independent LLM benchmark and pricing tracker | The independent scores this post cites (53 index score, 57% Terminal Bench 4) |
Cut your AI bill on purpose
The pattern behind this post is a consulting product wearing a blog post's
clothes. Teams adopting agents in 2026 do not have a model problem - they
have a measurement problem: no per-task token logging, no per-decision
price model, no calendar block before the invoice changes. I help teams
and founders set up exactly that: the routing, the acceptance checks, and
the cost monitoring designed together before the bill doubles. If your
agent or API bill is growing faster than your output, that is a week of
work, not a rewrite. Details are on my about page.
The intro price is a seven-week gift, not a business plan. Measure on it,
decide against December rates, and set the decision date now.— The whole post in one line
FAQ
What is Gemini 4 Argon and when did it launch?
Gemini 4 Argon is the first model in Google's Gemini 4 generation, built
for deep multi-step work in coding, finance, legal, and science. Google
announced it September 30, 2026 (limited to Fairwind Program cybersecurity
partners) and made it generally available October 13, 2026. The API model
ID is gemini-4-argon.
How much does Gemini 4 Argon cost?
Through November 30, 2026, the Standard API rate is $2 per million input
tokens and $10 per million output tokens, with cached input 95% off at
$0.10. From December 1, 2026, the price doubles to $4/$20 with cached
input at $0.20. Output billing includes thinking tokens, and the model is
not available on the API's free tier.
Why does the price double on December 1, 2026?
The $2/$10 rate is an introductory launch price. Reuters and VentureBeat
report Google confirming the increase; from December 1, Argon matches
Claude Opus 5.5 at $4/$20 per million tokens. That gives developers a
seven-week evaluation window before the economics change.
Where can I use Gemini 4 Argon?
Developers can call it through the paid Gemini API and Google AI Studio.
Google AI Ultra subscribers get it in Antigravity and in the Gemini app
and AI Mode in Search, which also covers AI Pro subscribers. Android
Studio, Firebase AI Logic, and Gemini Enterprise carry it too. It is not
on the free tier.
Is Gemini 4 Argon better than Claude Opus 5.5 or GPT-6?
Google's vendor-reported claims include first place on the Vals Index and
87% on the ERQA reasoning test, but independent scores are cooler: 53 on
the Artificial Analysis Intelligence Index and 57% on Terminal Bench 4,
behind Sonnet 5.5, Opus 5.5, and GPT-6 Astra. From December 1, Argon and
Opus 5.5 cost exactly the same, so the choice comes down to which model
does your work best, not price.
How do I measure my real per-task cost before the hike?
Log tokens per task with input, cached input, and output billed separately
(thinking tokens count as output). Prototype before November 30 so your
baseline is measured on the intro price, then re-price the same workload
at December rates. Use Batch or Flex pricing for background jobs, lean on
the 95% cached-input discount, and set a November 29 decision date so the
invoice never surprises you.
Top comments (0)