Gemini 4 Argon: lowest hallucination rate, #1 automation
New models on Complete
Google's Gemini 4 Argon is out, and it is built for business work that has to run without someone checking every line. Independent testing by Artificial Analysis puts it #1 on AutomationBench, gives it the lowest hallucination rate of any top model, and at launch pricing it costs less per task than every model it competes with on quality, except GPT-6.1 Sol. That launch price runs until November 30, 2026.
Gemini 4 Argon is available for enterprise teams on Complete.
What Gemini 4 Argon is
Gemini 4 Argon is Google DeepMind's newest closed-weights model. It rolled out to selected users on September 30 and launched broadly on October 13, 2026.
Reads text, images, video, audio and PDFs. Answers in text.
Holds about 1 million tokens of context, enough for a large contract set or a full quarter of reports in a single request.
Writes up to 65,536 tokens per response, with a long-output mode that extends to roughly 1 million.
Always reasons before answering. Effort can be set to low, medium or high.
On the Artificial Analysis Intelligence Index v4.3.2, Argon scores 53. That ranks it 8th of 214 models, where the median score is 27. It ties GPT-6 Astra and sits a few points behind Claude Sonnet 5.5 (56) and Claude Opus 5.5 (58).
Argon has the lowest hallucination rate of any top model
Artificial Analysis measures how often a model gives a confident wrong answer instead of admitting it doesn't know, using its AA-Omniscience test. Argon's hallucination rate is 15%. That is the lowest of any model scoring 45 or more on the index. GPT-6 Astra comes in at 51% and GPT-6.1 Sol at 54%.
For a finance close, a compliance check or a customer reply, a wrong answer stated with confidence costs more than an honest "I don't know." Argon gives that answer far more often than its peers.
Argon also answers fewer of those knowledge questions correctly: its AA-Omniscience accuracy is 50%, 13 points below GPT-6 Astra. Argon knows a little less and guesses a lot less, which suits work where an unflagged mistake is the expensive outcome.
Argon finishes automation tasks better than any model tested
AutomationBench, run by Artificial Analysis, scores how well a model carries out multi-step business automations from start to finish. Argon scores 78%, ranked first. Claude Sonnet 5.5, the next model, scores 71%.
The rest of its work record points the same way:
AA-Briefcase, which grades realistic business deliverables: 1494 Elo with a 65% rubric pass rate, the highest Artificial Analysis has recorded.
Terminal-Bench 4.0, which tests hands-on technical tasks in a command line: 57%. Only Claude Sonnet 5.5 (64%), Claude Opus 5.5 (60%) and GPT-6 Astra (59%) score higher.
Describe the process. Argon writes and runs the steps. A named person approves the result before anything ships.
It costs half its standard price until November 30
Google is running launch pricing on Argon through November 30, 2026.
Input, per 1M tokens | Output, per 1M tokens | Cached input | |
|---|---|---|---|
Launch price, through Nov 30, 2026 | $2.00 | $10.00 | 95% off |
Standard price, from Dec 1, 2026 | $4.00 | $20.00 | 95% off |
Token prices only tell part of the story, because models use different amounts of tokens to finish the same job. Artificial Analysis also reports the cost of running a full index task. At launch pricing, an Argon task costs $1.99. The same task costs $3.26 on GPT-6 Astra, which scores the same 53, and $5.98 on Claude Opus 5.5.
Intelligence score against cost per task. Source: Artificial Analysis Intelligence Index v4.3.2. Argon is plotted at launch pricing.
Model | Intelligence Index | Cost per task |
|---|---|---|
Claude Opus 5.5 | 58 | $5.98 |
Claude Sonnet 5.5 | 56 | $5.46 |
Gemini 4 Argon, launch price | 53 | $1.99 |
Gemini 4 Argon, standard price | 53 | $3.98 |
GPT-6 Astra | 53 | $3.26 |
GPT-6.1 Sol | 52 | $0.72 |
Grok 4.7 | 46 | $3.74 |
DeepSeek V4.1 Flash 1017 | 45 | $0.26 |
At $1.99, Argon sits on the value frontier: no model on the chart scores higher for less money. From December 1 its task cost doubles to $3.98, about 1.2 times GPT-6 Astra. Workloads you run and measure in November get the low price, and they also tell you whether Argon's accuracy and automation results are worth the standard rate for your processes.
Where Argon fits, and where it doesn't
Argon takes its time. Artificial Analysis measures a 47.1 second wait before the first token appears, against a 3.85 second median for models in its tier. Once it starts, it writes at about 81 tokens per second on Google's API. It is also verbose: it used 110 million output tokens to run the full index, against a median of 87 million.
That profile suits work that runs in the background: monthly reporting, document review, reconciliations, scheduled automations. A live chat with a customer waiting on the other end is better served by a faster model.
Use Argon for | Use a faster model for |
|---|---|
Scheduled reports and reconciliations | Live chat and instant replies |
Reviewing contracts, filings and long PDFs | High volume, low stakes classification |
Multi-step automations that run unattended | Anything where a 45 second wait breaks the flow |
Run Gemini 4 Argon on Complete
Gemini 4 Argon is available for enterprise teams on Complete. Your ops, finance and compliance owners describe the process, the agent builds and runs it on Argon, and every consequential step waits for a named approver.
Book a demo to run Gemini 4 Argon on your own workflows.
Frequently asked questions
Is Gemini 4 Argon better than GPT-6 Astra?
On the Artificial Analysis Intelligence Index, they score the same: 53. Argon has a 15% hallucination rate against 51% for Astra, while Astra answers 13 points more knowledge questions correctly. Argon costs $1.99 per task at launch pricing against $3.26 for Astra. From December 1 its cost per task doubles to $3.98, about 1.2 times Astra.
Where does Gemini 4 Argon run?
It is a closed model with no downloadable weights. It runs on the Gemini API, Google AI Studio, Vertex AI on Google Cloud, Antigravity, OpenRouter and the Gemini app. It is also available for enterprise teams on Complete.
What inputs does Gemini 4 Argon support?
It reads text, images, video, audio and PDFs, and responds in text.
When does the Gemini 4 Argon launch price end?
Google's introductory pricing of $2 per million input tokens and $10 per million output tokens runs through November 30, 2026. Standard rates of $4 and $20 apply from December 1.
Sources
Benchmark results are from Artificial Analysis Intelligence Index v4.3.2 as of October 22, 2026. Pricing is Google's published Gemini API pricing.
