What Gemini 4 Argon is, and when it came out
Koray Kavukcuoglu of Google DeepMind announced Gemini 4 Argon on September 30, 2026, as a "frontier model for real-world coding, enterprise knowledge work, and cyber defense". On October 13 Google opened it to everyone on the paid API under the line "Built to solve your hardest problems", saying it "excels at deep work in complex, real-world tasks" across coding, finance, legal, cybersecurity and science. Google's models page describes it as an "efficient frontier reasoning model with faster responses and lower token spend per turn than Carbon".
Argon was the first of four Gemini 4 models. The others are Carbon (the strongest), Hydrogen (fast and cheapest) and Neon (the middle model and the newest). For the whole family, see Gemini 4 models.
- September 30, 2026: announced, first for "a set of trusted cyber defenders" in Google's Fairwind Program, then paid API customers and Google AI Ultra subscribers. Artificial Analysis uses this as Argon's release date.
- October 13, 2026: broad release in the Gemini API, AI Studio, Android Studio, Gemini for business, Antigravity, the Gemini app and AI Mode in Search.
- November 10, 2026: Carbon went public. In that post Google said Argon "remains the ideal choice for the majority of software engineering and knowledge work", with Carbon for harder, open-ended tasks.
- November 18, 2026: Neon launched. Google Cloud's Neon developer guide calls Neon "the target upgrade" from Argon. Argon is still listed and priced in the API, and Google Cloud's Argon page shows no deprecation notice.
Gemini 4 Argon specs
| Gemini 4 Argon | |
|---|---|
| Announced | September 30, 2026 (Fairwind cyber defenders first) |
| Public API | October 13, 2026 |
| Model ID | gemini-4-argon |
| Context window | 1,048,576 tokens, one limit shared by input and output (Google Cloud) |
| Max output | Google's September 30 post says Argon's output limit went "to an industry-leading 1M tokens, up from the previous 64K". Google Cloud's Argon page: long decodes can run up to 1M minus the input, and each foreground segment is up to 65,536 tokens streamed, 32,768 non-streamed. |
| Thinking levels | LOW, MEDIUM (default), HIGH |
| Google Cloud locations | PayGo in global, us and eu; Provisioned Throughput also in europe-west2 and asia-southeast1 |
| Cyber variant | gemini-4-argon-cyber-permissive, for Fairwind allowlisted users on Google Cloud |
| API price per 1M tokens | $2 input, $0.10 cached, $10 output through November 30, 2026; $4, $0.20 and $20 from December 1. No free tier. |
Gemini 4 Argon pricing vs Neon, Carbon, Hydrogen and Opus 5.5
Gemini API Standard prices per million tokens, from Google's pricing page on November 21, 2026. Google's posts word the end of the promo differently ("after November 30", "from November 30"); the pricing page says the promo price runs "through November 30, 2026" and the new price starts "December 1, 2026", so we use that. Claude Opus 5.5 prices are Anthropic's.
Batch and Flex cost half the Standard price and Priority costs 1.8x. An invite-only Contributor tier is $1 / $5. Argon isn't on the free tier. Vertex AI per-token prices weren't read, so this page gives Gemini API prices only.
| Model | Input | Cached input | Output | Free tier (API) |
|---|---|---|---|---|
| Gemini 4 Argon | $2 until Nov 30, then $4 | $0.10, then $0.20 | $10 until Nov 30, then $20 | No |
| Gemini 4 Carbon | $2 until Nov 30, then $4 | $0.10, then $0.20 | $10 until Nov 30, then $20 | No |
| Gemini 4 Neon | $3 | $0.15 | $15 | No |
| Gemini 4 Hydrogen | $1.50 | $0.075 | $7.50 | Yes |
| Claude Opus 5.5 | $4 | $0.20 | $20 | - |
How to use the Gemini 4 Argon API
Get a paid API key
Argon isn't on the Gemini API free tier, so you need a paid key. You can then pick Gemini 4 Argon in Google AI Studio.
Call it from the Gemini API
Use the model ID
gemini-4-argon. Thinking defaults to MEDIUM; set HIGH for harder tasks or LOW when speed matters more. The context is 1,048,576 tokens shared by input and output.# Gemini API: model ID gemini-4-argon (paid tier only) curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-4-argon:generateContent" \ -H "x-goog-api-key: $GEMINI_API_KEY" \ -H "Content-Type: application/json" \ -d '{"contents": [{"parts": [{"text": "Review this contract clause and list the risks"}]}]}' # thinking levels: LOW | MEDIUM (default) | HIGHOr use it on Google Cloud (Vertex AI)
The same model ID works on Google Cloud's Gemini Enterprise Agent Platform. PayGo runs in
global,usandeu. For very long answers, long decoding continues up to 1M tokens minus the input, in segments of up to 65,536 tokens streamed.Consider Neon for new work
Google Cloud names Neon as Argon's target upgrade. Neon defaults to HIGH thinking and adds XHIGH. Check the differences in Gemini 4 Neon before you switch.
Where you can use Gemini 4 Argon
- Gemini app and AI Mode in Search: Google's October 13 post makes Argon available to Google AI Pro and AI Ultra subscribers.
- Gemini API and Google AI Studio: paid tier only.
- Google Antigravity: for AI Ultra subscribers.
- Android Studio and Gemini for business.
- Google Cloud:
gemini-4-argonand, for Fairwind allowlisted users,gemini-4-argon-cyber-permissive.
Gemini 4 Argon benchmarks
These are Google's own numbers from its September 30 and October 13 posts. Where Google claims a lead without a score, we say so; those claims weren't checked on the benchmark owners' sites.
| Benchmark | Gemini 4 Argon | Reported by |
|---|---|---|
| DeepSWE v1.1 | 77.9%, new state of the art | Google (Sep 30) |
| Zapier AutomationBench | 51.3%, #1 | Google (Sep 30) |
| LVBench | 91.7%, state of the art | Google (Sep 30) |
| CWE-bench v1 | 68%, tied #1 | Google (Sep 30) |
| Vals Index | #1 (no score given) | Google (Oct 13) |
| Vals Finance Agent v2, Harvey Legal Agent Benchmark | Leading (no scores given) | Google (Sep 30) |
| ERQA | 87% | Google (Oct 13) |
| RareBench (Gamow Labs, private) | 57%, top score | Google (Oct 13) |
| LabBench (Gamow Labs, private) | 52.2%, top score | Google (Oct 13) |
| Gray Swan indirect prompt injection | 0.7% attack success over 15 adaptive attempts | Google (Oct 13) |
What else Google reports
- Safety, per Google's October 13 post: coding shortcut attempts on 2% of tasks, adversarial harmful-request violations cut nearly in half, and under 1% of sub-tasks unlocked in bio red teaming.
- Google later compared Carbon with Argon. Carbon scored 0.0% on Gray Swan prompt injection (Argon 0.7%), 92% on agentic honesty (Argon 70%), 2.17% on Gray Swan automated cyberattacks (Argon 9.86%), and more than 2x Argon on an internal secure-code benchmark.
- Hydrogen's launch post puts Hydrogen at 51 on the Artificial Analysis Intelligence Index at high, against Argon's 50.
Gemini 4 Argon on Artificial Analysis and Arena
Two independent sources rank Argon very differently. Artificial Analysis runs a fixed set of evals; Arena ranks models by human votes in blind chats. Ranks move quickly, so each figure has its date.
- Artificial Analysis, November 21, 2026: Intelligence Index 50 at high and 48 at medium, #33 of 95 models. Output speed 69 tokens per second at high, 44.5 seconds to first token, and $4.97 per Intelligence Index task.
- For comparison on the same AA pages: Neon 61 (xhigh), Carbon 60 (high), Hydrogen 54 (xhigh). Cost per task: Hydrogen $3.16, Argon $4.97, Neon $6.44, Carbon $7.76. Output speed: Hydrogen about 161-166 tokens per second, Neon about 110, Argon 69, Carbon 55.
- Arena Text leaderboard, November 20, 2026: Gemini 4 Argon (High) is #2 with 1520, behind Carbon (1527) and ahead of Neon (1513). Claude Opus 5.5 (High) is #5 with 1506. Hydrogen isn't ranked on that board.
Gemini 4 Argon vs Neon vs Carbon
Prices are Gemini API Standard rates per million tokens. AA scores are the highest Intelligence Index figure Artificial Analysis shows for each model.
| Argon | Neon | Carbon | |
|---|---|---|---|
| Google's description | Efficient frontier reasoning, lower token spend per turn than Carbon | "Balances depth and efficiency" | "Our strongest model" |
| Public API | Oct 13, 2026 | Nov 18, 2026 | Nov 10, 2026 |
| API price in / out | $2 / $10 to Nov 30, then $4 / $20 | $3 / $15 | $2 / $10 to Nov 30, then $4 / $20 |
| Thinking levels | low, medium (default), high | low, medium, high (default), xhigh | low, medium (default), high |
| Context | 1,048,576 | 1,048,576 | 1,048,576 |
| Output per streamed segment | 65,536 | 65,536 | 65,536 |
| AA Intelligence Index (Nov 21) | 50 (high), #33 | 61 (xhigh), #6 on Nov 19 | 60 (high), #8 |
| AA cost per Index task | $4.97 | $6.44 | $7.76 |
| Arena Text (Nov 20) | #2, 1520 | #3, 1513 | #1, 1527 |
| Gemini app | AI Pro and Ultra | AI Pro and Ultra | AI Ultra |
| Antigravity | AI Ultra | Paid AI Pro and Ultra | AI Ultra |
Argon, Neon or Carbon: which to pick
- Argon vs Neon: Google Cloud calls Neon the target upgrade from Argon, and Google's model-picking guide in AI Studio lists Neon, Carbon and Hydrogen but not Argon. Neon scores 11 points higher on AA's Intelligence Index (61 vs 50), defaults to HIGH thinking and adds XHIGH. Argon ranks higher on Arena (#2 vs #3) and costs less per AA task ($4.97 vs $6.44). Per token, Argon is cheaper through November 30 and 33% more expensive from December 1.
- Argon vs Carbon: same price and same thinking levels. Google calls Carbon its strongest model and says it uses more tokens than Argon to explore and verify. On AA, Carbon scores 10 points higher (60 vs 50) but costs about 56% more per task ($7.76 vs $4.97) and runs slower (55 vs 69 tokens per second). On Arena, Carbon is #1 and Argon #2, 7 points apart. On Google's own safety tests, Carbon resists prompt injection and cyberattacks better.
- Argon vs Hydrogen: Hydrogen costs $1.50 / $7.50, has a free API tier, runs more than twice as fast on AA and scores slightly higher there (54 at xhigh, 51 at high, vs Argon's 50). See Gemini 4 Hydrogen.
- Rule of thumb: if you already run Argon and it works, it is still available and priced. For new work, Google points to Neon for complex tasks with someone waiting, Carbon for multi-hour runs and security audits, and Hydrogen for fast, high-volume work.
Gemini 4 Argon vs Claude Opus 5.5
Only data from the same referee for both models, plus each vendor's list prices. We don't have an Artificial Analysis Intelligence Index figure for Opus 5.5 from a source we checked, and Argon isn't in the AA Coding Agent Index results we have, so neither appears here.
| Gemini 4 Argon | Claude Opus 5.5 | Note | |
|---|---|---|---|
| Arena Text (Nov 20, 2026) | #2, 1520 (High) | #5, 1506 (High) | Same board, human votes |
| API price in / out per 1M | $2 / $10 to Nov 30, then $4 / $20 | $4 / $20 | Same price from December 1 |
| Cached input per 1M | $0.10, then $0.20 | $0.20 | Same from December 1 |
| Context | 1,048,576 tokens, shared by input and output | 1M input, 128K output | - |
What that means
- On Arena's text board, people preferred Argon's answers over Opus 5.5's often enough to put it three places higher, 14 points apart. Arena measures chat preference, not agentic coding.
- Through November 30, Argon costs half as much per token as Opus 5.5. From December 1 the two list at the same price.
- For coding agents, the agent matters as much as the model. Opus 5.5 runs in Claude Code; Argon runs in Google Antigravity and through the Gemini API. See Claude Code vs Codex for how the two main coding agents compare.
A new model is half of it. The other half is a job.
Models keep getting better and cheaper, and the agents around them still wait for the next prompt. Crewmark turns coding agent sessions into AI employees with a role, a goal in numbers, working hours and a heartbeat, so they take the next issue from a board on their own. Claude Code and Codex employees run natively; other agents, such as OpenCode, DeepSeek Harness and Gemini, connect through ACP (acpx). An employee uses whatever model its agent runs.
Pick a role
From the AI employee roles, such as an AI software engineer.
Put the work on a board
Assign issues, make the regular ones repeat, and let the heartbeat wake the employee to take the next one.
See what it can do
Jobs with setup and rules for each agent are in Claude Code use cases and Codex use cases.
FAQ
What is Gemini 4 Argon?
Google's first Gemini 4 model, a frontier model for real-world coding, enterprise knowledge work and cyber defense. It has a 1M-token context shared by input and output, three thinking levels (LOW, MEDIUM by default, HIGH), and the model ID gemini-4-argon.
When was Gemini 4 Argon released?
Google announced it on September 30, 2026, for cyber defenders in its Fairwind Program first. Public API access started on October 13, 2026, along with the Gemini app for AI Pro and Ultra subscribers. Artificial Analysis lists September 30 as the release date.
How much does Gemini 4 Argon cost?
In the Gemini API, $2 per million input tokens, $0.10 cached and $10 output through November 30, 2026. From December 1 it's $4, $0.20 and $20. Batch and Flex are half price, Priority is 1.8x. There is no free API tier.
What are Gemini 4 Argon's benchmarks?
Google reports 77.9% on DeepSWE v1.1, 51.3% on Zapier AutomationBench, 91.7% on LVBench, 68% on CWE-bench v1 and 87% on ERQA. Independently, Artificial Analysis scores it 50 on its Intelligence Index (#33 of 95 on November 21), and Arena's text board ranks it #2 with 1520 (November 20).
Gemini 4 Argon vs Neon: which should I use?
Google Cloud calls Neon the target upgrade from Argon. Neon scores 61 to Argon's 50 on Artificial Analysis and adds XHIGH thinking, while Argon ranks higher on Arena and costs less per AA task. From December 1, Neon ($3 / $15) is cheaper per token than Argon ($4 / $20). See Gemini 4 Neon.
Gemini 4 Argon vs Carbon: what's the difference?
Both cost the same and have the same thinking levels. Google calls Carbon its strongest model, for harder, open-ended tasks, and said Argon remains the ideal choice for most software engineering and knowledge work. Carbon scores 60 to Argon's 50 on Artificial Analysis but costs more per task and runs slower. See Gemini 4 Carbon.
Is Gemini 4 Argon better than Claude Opus 5.5?
On Arena's text leaderboard (November 20, 2026), Argon is #2 with 1520 and Opus 5.5 is #5 with 1506. Argon costs half as much per token through November 30, and the same from December 1. We don't have independent coding results that include both, so test them on your own tasks.
Is Gemini 4 Argon being retired?
Google hasn't said so. Argon is still listed and priced in the Gemini API, and Google Cloud's Argon page has no deprecation notice. Google Cloud does name Neon as the upgrade target.
Checked against Google's Argon posts, the Gemini API pricing and models pages, Google Cloud's Argon and Neon docs, Artificial Analysis and Arena on November 21, 2026.