Gemini 4 Argon API: Benchmarks, Pricing and Setup
Gemini 4 Argon costs $2/$10 per 1M to 30 Nov 2026, then $4/$20. Google's benchmarks and limits, plus abliterated models for cyber security work.
Model releases·Published ·Checked ·By RouterPlex·13 min
gemini-4-argon is live on RouterPlex at vendor list price, 0% markup, $0 top-up fee.

On this page
Google announced Gemini 4 Argon on 30 September 2026. It is Google DeepMind's new frontier model, and Google's launch table puts it ahead of GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 on 14 of 19 benchmark rows. The introductory API price is $2 per 1M input tokens and $10 per 1M output tokens through 30 November 2026. From 1 December 2026 it becomes $4 / $20.
Gemini 4 Argon does not answer on RouterPlex right now. The pause on Gemini routes that began on 27 October 2026 was lifted on 7 November 2026, but our upstream still has no working channel for gemini-4-argon, so requests fail or time out. Google's own paid Gemini API still serves gemini-4-argon. For authorized security work on RouterPlex today, use the abliterated models below. They answer the pentest, CVE and red-team questions that stock models often refuse.
Updated 10 November 2026: the RouterPlex pause on Gemini routes is lifted, but Argon still does not answer on our upstream. Abliterated prices are updated. Google announced Gemini 4 Carbon, its next model, on 28 October 2026.

Sources: Google's Gemini 4 Argon announcement by Koray Kavukcuoglu (30 September 2026), and Google DeepMind's evaluation methodology. Both checked 30 September 2026. Pricing, limits and the GA date are from Google's Gemini API pricing page, Argon model page and changelog, checked 14 October 2026. Every score on this page is Google's. The images are Google's launch graphics. RouterPlex has not re-run these evals.
Gemini 4 Argon at a glance
| Gemini 4 Argon | |
|---|---|
| Announced | 30 September 2026 |
| Generally available (Gemini API) | 13 October 2026 |
| Maker | Google DeepMind |
| Model ID | gemini-4-argon |
| Intro API price (to 30 Nov 2026) | $2 input / $10 output per 1M tokens |
| Price from 1 Dec 2026 | $4 input / $20 output per 1M tokens |
| Cached input (Google) | 95% off input ($0.10 intro, $0.20 standard) |
| Context window | 1,048,576 tokens, shared by input, thinking and output |
| Output per request | 65,536 synchronous; up to 1M with Google's long decode |
| Thinking levels | low, medium (default), high |
| Inputs | Text, image, video, audio, PDF |
| On RouterPlex | Listed, but not answering on our upstream (10 November 2026) |
Gemini 4 Argon benchmarks
This is Google's full launch table. Blue cells are where Google marks Argon as the best score. Grey cells mark another model's lead.

The same numbers as text, so you can search and copy them:
| Category | Benchmark | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 | Claude Opus 5.5 |
|---|---|---|---|---|---|
| Knowledge work | Vals Index | 68.9% | 63.1% | 65.8% | 67.0% |
| Knowledge work | AutomationBench | 51.3% | 41.4% | 31.4% | 42.5% |
| Knowledge work | Vals Finance Agent v2 | 65.4% | 53.5% | 58.9% | 58.6% |
| Knowledge work | Harvey's Legal Agent Benchmark | 19.6% | 5.4% | 6.7% | 3.8% |
| Agentic coding | DeepSWE v1.1 | 77.9% | 74.1% | 67.4% | 74.2% |
| Agentic coding | FrontierSWE v2 | 55.0% | 65.5% | 56.3% | 62.3% |
| Agentic coding | Vibe Code Bench | 91.9% | 89.6% | 90.3% | 90.3% |
| Agentic coding | Terminal-bench 4.0 | 57.4% | 58.2% | 57.9% | 66.4% |
| ML engineering | PostTrainBench | 45.3% | 44.3% | 40.2% | 49.3% |
| Science and math | Terminal-Bench Science 0.1 | 57.6% | 68.1% | 52.6% | 63.3% |
| Science and math | LABBench 2 | 88.8% | 85.4% | 68.6% | 73.1% |
| Science and math | RiemannBench | 76.0% | 72.0% | 65.6% | 69.6% |
| Long context | GraphWalks, up to 128K (BFS F1) | 99.7% | 98.7% | 91.4% | 90.6% |
| Long context | GraphWalks, 256K to 1M (BFS F1) | 84.2% | 71.8% | 65.0% | 66.8% |
| Computer use | Agent's Last Exam (pass rate) | 39.5% | 34.2% | n/a | 38.2% |
| Computer use | OSWorld-2.0 (offline subset, partial score) | 69.2% | 72.6% | n/a | n/a |
| Multimodal | Chartography | 71.6% | 71.0% | 46.2% | 66.3% |
| Multimodal | LVBench | 91.7% | 87.5% | 79.7% | 83.7% |
| Cybersecurity | CWE-bench v1 | 68.0% | 68.0% | 58.0% | 67.0% |
Where Argon leads
- Knowledge work is the clearest win. Argon leads all four rows. The gap is widest on Harvey's Legal Agent Benchmark: 19.6%, against 6.7% for the next model. Google also calls Argon the leading model on the Vals Index, which weights finance, coding, legal and tax work by each sector's share of U.S. GDP.
- Long-horizon coding. DeepSWE v1.1 is 77.9%, which Google calls a new state of the art. That is 3.7 points above Opus 5.5 and 3.8 above GPT-6 Astra.
- Very long prompts. On GraphWalks between 256K and 1M tokens, Argon scores 84.2%. The next best is 71.8%.
- Video and charts. LVBench, a long-video test, is 91.7%. Google calls that state of the art.

Where Argon trails
Google's own table shows five rows where another model is ahead:
- Terminal-bench 4.0: Opus 5.5 leads, 66.4% to 57.4%.
- FrontierSWE v2: GPT-6 Astra leads, 65.5% to 55.0%. Opus 5.5 is also ahead at 62.3%.
- Terminal-Bench Science 0.1: Astra leads, 68.1% to 57.6%.
- OSWorld-2.0: Astra leads, 72.6% to 69.2%. Anthropic models are not listed on this row.
- PostTrainBench: Opus 5.5 leads, 49.3% to 45.3%.
If your work is terminal-heavy agent coding, the table does not make Argon the default choice. Test your own tasks first.
How Google ran the numbers
The methodology document is short, and it matters:
- Argon scores are pass@1 and ran on the Gemini API at the highest thinking setting.
- Most competitor scores are the vendors' own published numbers, at their maximum reasoning setting when available. They were not re-run under one setup.
- Some Argon scores are Google's own runs. DeepSWE used a mini-swe-agent setup. Terminal-Bench Science used a 6× verifier timeout. OSWorld-2.0 is the best of 3 runs.
- PostTrainBench, LABBench 2 and GraphWalks were run by Google for every model.
- LVBench frame counts differ by model because of API limits: 1 FPS for Gemini, 800 frames for Astra, 300 for Fable 5.1 and 600 for Opus 5.5.
That is normal for a launch table. It is also a good reason to re-test on your own prompts before you switch.
Cybersecurity: Google's headline use case
Google trained Argon for cyber defense. It says Argon can find, validate and patch serious software vulnerabilities on its own. Trusted defenders get Argon without cyber guardrails through the Fairwind Program. That is also why it is the first group with access.

- CWE-bench v1 (fixing security vulnerabilities): 68.0%, tied first with GPT-6 Astra.
- Real-world vulnerability discovery (Google internal, 20 languages): 85.8%, against 71.0% for Gemini 3.8 Flash Cyber.
- Wiz penetration test benchmark (no source code access): 70.9%, against 58.2%.
- Wiz used Argon in its Scan for Good program. Google says it found a critical vulnerability in hospital software that earlier frontier models had missed.
On prompt injection, Google reports the lowest attack success rate on Gray Swan's indirect prompt injection benchmark: 0.7% at 15 attempts. Claude Opus 5.5 and Claude Fable 5.1 are both at 1.0%. Lower is better on this chart.

Outside the Fairwind Program, Argon keeps its cyber guardrails, so some offensive security prompts still get refused. On RouterPlex, the abliterated models are the route for that work.
Gemini 4 Argon API pricing
Google's standard tier:
| Tier | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|
| Introductory, through 30 Nov 2026 | $2.00 | $0.10 | $10.00 |
| From 1 Dec 2026 | $4.00 | $0.20 | $20.00 |
Output includes thinking tokens. There is no long-context surcharge. Google lists Batch and Flex at half the standard rate ($1 / $5 intro) and Priority at $3.60 / $18 intro. Caching is implicit only: Google does not support explicit context caches for Argon.
What one request costs
A 40,000-token prompt that returns 2,000 tokens, with no cache, at Google's list price:
| Period | Input / Output per 1M | 40K in / 2K out |
|---|---|---|
| Intro, to 30 Nov 2026 | $2 / $10 | $0.100 |
| From 1 Dec 2026 | $4 / $20 | $0.200 |
The same turn costs $0.012 on glm-5.3-flash-abliterated and $0.088 on glm-5.3-abliterated, both live on RouterPlex.
Thinking tokens are the real bill
Thinking tokens bill at the output rate, and Argon thinks at medium by default. In our launch test, a one-sentence answer at reasoning_effort: "high" returned 516 output tokens for about 30 visible words: $0.005 for one short reply. Use low for extraction, routing and short answers. Keep high for the long-horizon coding and analysis work the benchmarks measure.
The 1M output limit needs long decode
Google's headline 1M-token output is real, but it is not the default. A normal synchronous request returns at most 65,536 output tokens. Above that, Google requires its long-decode mode: a background job on the Interactions API that can run for up to 24 hours, with every hop billed. A full 64K response costs about $0.66 at the intro rate. If you run agents on Argon, set max_tokens and a spend limit.
Availability and release date
- 30 September 2026: announced. Rolled out first to trusted cyber defenders through the Fairwind Program.
- 13 October 2026: generally available on the paid Gemini API as
gemini-4-argon. No free tier. - 14 October 2026: listed on RouterPlex as
gemini-4-argon. - 27 October 2026: paused on RouterPlex.
- 28 October 2026: Google announced Gemini 4 Carbon, again for cyber defenders first.
- 7 November 2026: the RouterPlex pause was lifted, but
gemini-4-argonstill has no working upstream channel. - 30 November 2026: last day of the $2 / $10 introductory price.
- 1 December 2026: Google's list becomes $4 / $20.
Google says it is taking part in the U.S. government's voluntary pre-release access process. It is also adding safeguards in four areas before a broad release: misuse (cyber and CBRN), prompt injection, misalignment monitoring of the model's chain of thought and actions, and hardened sandboxes.
What Google says Argon already does inside Google
These are Google's own examples. They are not independent tests.
- C/C++ to Rust migrations, from libraries like re2 and libgav1 up to the 800K+ line Fuchsia Zircon kernel. The Rust port of the libgav1 video decoder runs 2.7× faster with identical output after Argon replaced 32K lines of SIMD code.
- Memory savings across Google data centers: over 300 TiB freed so far, with 500 TiB to 1 PiB expected in total.
- Quantum algorithm work: beat a published baseline by 40% in minutes.
Is Gemini 4 Argon on RouterPlex?
Not in a usable way. RouterPlex listed gemini-4-argon on 14 October 2026 and paused Gemini routes on 27 October 2026. The pause was lifted on 7 November 2026, but our upstream still has no working channel for Argon, so requests fail or time out. We will update this post when Argon answers.
Abliterated models for cyber security work
Abliteration is a weight edit that removes a model's refusal behaviour. The model then answers security questions that stock models, Argon included, often decline. RouterPlex hosts five abliterated routes. USD per 1M tokens, from the RouterPlex catalog, checked 10 November 2026.
See the abliterated models API hub for live prices and usage rules.
| Model id | Input | Cache read | Output | Max output | Page |
|---|---|---|---|---|---|
glm-5.3-abliterated | $2.00 | $0.30 | $4.00 | 131,072 | page |
glm-5.3-flash-abliterated | $0.25 | $0.06 | $1.00 | 131,072 | page |
deepseek-v4.1-flash-abliterated | $0.50 | $0.07 | $1.50 | 131,072 | page |
glm-4.7-flash-abliterated | $0.12 | $0.09 | $0.60 | 24,000 | page |
gemma-4-26b-abliterated | $0.25 | input rate | $0.75 | 8,192 | page |
All five support tool calling. The first three have a 1M-token context and image input. GLM-4.7 Flash is text only with a 200K context, and Gemma 4 has a 256K context with image input. These are RouterPlex-set rates. RouterPlex stores the requests sent to abliterated models (not the responses) for 1 to 365 days for security and abuse prevention, as the privacy policy and section 05 of the terms state. Requests to other models are not stored.
Security teams use them for work they are authorized to do. They red-team their own LLM apps with real attack prompts. They draft pentest plans and reports for systems in scope. They read how a disclosed CVE works, then write detection rules and patches. Incident responders triage suspicious scripts, logs and malware samples. Others practice on CTF challenges.
The terms still apply: sexual content of any kind is banned on every model, and these routes must not be used against systems you do not own or are not authorized to test.
Be clear about the gap. Abliteration removes refusals. It does not add skill. On the ExploitGym chart Google published with Gemini 4 Carbon, stock GLM-5.3 solves 15.0% against Argon's 42.0%. On SEC-bench Pro, stock DeepSeek-V4.1-Flash scores 62.8% against Argon's 83.4%. What you get is a capable model that answers the question, at a price you can see. Read what abliteration is for the background.
Call an abliterated model
Set ROUTERPLEX_API_KEY in your environment. Never paste a key into code.
curl https://api.routerplex.com/v1/chat/completions \-H "Authorization: Bearer $ROUTERPLEX_API_KEY" \-H "Content-Type: application/json" \-d '{"model": "glm-5.3-flash-abliterated","messages": [{"role": "system", "content": "You assist an authorized penetration test. Scope: staging.example.com only."},{"role": "user", "content": "Explain how CVE-2021-44228 works and write a Sigma rule that detects exploitation attempts in web server logs."}]}'
The same routes run in Claude Code and the Anthropic SDK:
export ANTHROPIC_BASE_URL=https://api.routerplex.comexport ANTHROPIC_AUTH_TOKEN=$ROUTERPLEX_API_KEYclaude --model glm-5.3-abliterated
See the Claude Code setup guide for the full configuration.
Where it goes wrong
- A timeout or 5xx on
gemini-4-argonmeans our upstream still has no working channel for it. Switch the model id. - A
402on an abliterated route means it needs paid top-up credit. Bonus credit, match credit and the $1 trial do not cover these routes. - A
402anywhere else means the balance hit $0 or the key hit its spend cap. The balance never goes negative. deepseek-v4.1-flash-abliteratedcan take 18 to 60 seconds. Raise your client timeout.
Start security work on RouterPlex
Create a RouterPlex account, add a $5 top-up with a $0 fee, set a hard spend cap on the key, and call glm-5.3-flash-abliterated. The same key reaches the rest of the model catalog. The terms apply to every request, and you are responsible for staying in scope.
Common questions
Frequently asked questions
What is Gemini 4 Argon?
Gemini 4 Argon is Google DeepMind's new frontier model, announced on 30 September 2026 by Koray Kavukcuoglu. Google built it for long, multi-step work: software engineering, finance and legal research, and cybersecurity defense. Its output limit is 1M tokens per response, up from 64K.
How much does the Gemini 4 Argon API cost?
Google's introductory price is $2 per 1M input tokens and $10 per 1M output tokens through 30 November 2026. From 1 December 2026 it becomes $4 input and $20 output per 1M. Thinking tokens are billed as output. Google prices cached input at 95% off.
Can I use Gemini 4 Argon today?
Yes, on Google's paid Gemini API, where gemini-4-argon has been generally available since 13 October 2026. There is no free tier on Google's side. On RouterPlex the model is listed but does not answer as of 10 November 2026: the Gemini pause was lifted on 7 November, and our upstream still has no working channel for Argon.
Is Gemini 4 Argon on RouterPlex?
Not in a usable way. RouterPlex listed gemini-4-argon on 14 October 2026. The Gemini pause that began on 27 October was lifted on 7 November 2026, but Argon still does not answer on our upstream as of 10 November 2026. For authorized security work, RouterPlex hosts five abliterated models, from $0.12 input and $0.60 output per 1M tokens on glm-4.7-flash-abliterated. They bill from paid top-up credit only.
How does Gemini 4 Argon compare with GPT-6 Astra and Claude Opus 5.5?
In Google's launch table, Argon scores highest or ties on 14 of 19 rows, including DeepSWE v1.1 (77.9% vs 74.1% Astra and 74.2% Opus 5.5) and AutomationBench (51.3% vs 41.4% and 42.5%). It trails on FrontierSWE v2, Terminal-bench 4.0, PostTrainBench, Terminal-Bench Science 0.1 and OSWorld-2.0. Google compiled the table. RouterPlex has not re-run it.
What is the Gemini 4 Argon context window?
Google's model page lists a unified 1,048,576-token window shared by input, thinking and output. A normal synchronous request can return up to 65,536 output tokens. Google's 1M-token output needs its long-decode mode, which runs as a background job on the Interactions API.
How do I set the Gemini 4 Argon thinking level?
Argon supports three thinking levels: low, medium and high. On Google's API, medium is the default. Higher levels spend more output tokens, and thinking tokens bill at the output rate, so use low for extraction and short answers.
What can I use instead of Gemini 4 Argon for cyber security work?
On RouterPlex, the abliterated routes answer security questions that stock models often refuse, such as red-teaming your own apps, CVE analysis, detection rules and pentest reports for systems you are allowed to test. glm-5.3-flash-abliterated costs $0.25 input and $1.00 output per 1M tokens. They score below Argon on Google's charts, and they bill from paid top-up credit only.
Run the smallest paid test.
Add $5, cap the key, and verify the result with your own workload. No subscription, and credit never expires — a first top-up of $25+ is matched with $25 extra.





