Gemini 4 Argon on Felo: 1M-Token Output for Search and API
Gemini 4 Argon raises output to 1M tokens and leads Google's knowledge-work benchmarks. What it changes for Felo Search, the Felo API, and long research.
Google announced Gemini 4 Argon on September 30, 2026. The headline is not a benchmark point. It is the output ceiling: 1 million tokens in a single pass, up from 64,000.
That number changes what a research session can end with. Instead of asking a model to draft a report section by section, you can hand it the full source set and ask for the finished document. Felo Search is built around that workflow (search, read, synthesize, cite), so a model that can write at length fits the product better than one that stops after a few pages.
Access is still rolling out. Gemini 4 Argon is not generally available from Google yet, and Felo Search access is being enabled alongside that rollout. The Felo API is in closed beta, with Argon access coming soon. This guide covers what the model is, what Google's numbers show, where it falls short, and how to prepare on Felo.
What Gemini 4 Argon is
Argon is Google's first frontier model since Gemini 3, and it ships under a new naming scheme rather than as an update to the Gemini 4.0 line. That is why the name carries no version number. Google is treating it as a separate release, not a point upgrade.
Its defining change is the output limit. Earlier Gemini flagships capped output at 64K tokens. Argon raises that to 1M, which Google says lets a model solve a hard problem in one trajectory instead of many round trips. The target workload is long, multi-step knowledge work: legal analysis, financial research, document review, and long-form drafting.
This is a different shape from the coding-first frontier models Argon is being compared to. Google's published leads cluster on reading, reasoning, and writing at length, not on agentic software engineering.
A 1M output ceiling is not the same as a 1M context window
A context window is everything the model can read in one request: your prompt, the documents you attach, and the conversation so far. Max output is how much the model can write back in a single response. Many models accept very long input and still stop generating after 64K tokens.
That mismatch is why long reports usually get assembled in pieces. The model can read the whole filing bundle but cannot write the whole analysis in one go, so you ask for section one, then section two, then a conclusion, then a pass to smooth the seams.
Argon removes that constraint at the output end. The practical result is a different request pattern: one instruction for the complete deliverable, with the model holding the argument together from start to finish.
Specs at a glance
| Specification | Gemini 4 Argon |
|---|---|
| Announcement | September 30, 2026 |
| Max output | 1,000,000 tokens |
| Previous Gemini output ceiling | 64,000 tokens |
| Introductory input price | $2.00 per 1M tokens |
| Introductory output price | $10.00 per 1M tokens |
| Cached input | 95% off |
| Later list price | $4.00 input / $20.00 output per 1M tokens |
| Rollout | Staged; not generally available |
Two pricing details matter. The introductory rate is temporary, and Google has not said when it ends, so budget on the later number. Cached input stays 95% off in both periods, which matters for long sessions that reuse the same document set on every turn.
The benchmarks Google published
Argon leads on knowledge-work evaluations by wide margins and ties or trails on some software-engineering tasks.
| Benchmark | What it measures | Argon | Nearest comparison |
|---|---|---|---|
| DeepSWE v1.1 | Long-horizon software engineering | 77.9% | Claude Opus 5.5 at 74.2%, GPT-6 Astra at 74.1% |
| CWE-bench v1 | Finding and patching vulnerabilities | 68% | Ties for first |
| AutomationBench | End-to-end business tasks | 51.3% | Ranked first |
| LVBench | Long video understanding | 91.7% | Highest score Google reports for Argon |
| Harvey Legal Agent | Long-horizon legal work | 19.6% | Claude Fable 5.1 at 6.7%, GPT-6 Astra at 5.4% |
| Vals Finance Agent v2 | Financial research and analysis | 65.4% | Claude Fable 5.1 at 58.9%, Claude Opus 5.5 at 58.6% |
The Harvey result is the widest gap in the published set: 19.6% against 6.7% for the next model, and roughly three times the nearest competitor. The Vals result is narrower but shows the same ordering.
On GraphWalks, which tests whether a model can actually use a context between 256K and 1M tokens, Google reports a lead of more than ten points over other flagships. That is the evaluation closest to the work Felo users do: not a short question, but a large set of sources that has to stay usable through a long answer.
One independent data point sits outside Google's table. Artificial Analysis measured a 15% hallucination rate for Argon, the lowest among the 36 frontier models it plots. In practice, that means the model is more willing to say it does not know instead of producing an unsupported answer. That matters when the output is a long document where errors are hard to spot.
Treat all of this as directional. The benchmark scores are vendor-reported, Google has not disclosed the thinking or effort setting used for the headline numbers, and no independent lab has replicated them. Bloomberg has reported that some Google employees are skeptical the benchmark lead survives contact with real work. That is normal for a launch-week model, and it is why the first thing to test is your own workload, not the leaderboard.
Where Argon is not the right pick
Argon's advantage is uneven, and Google's own numbers show where it falls away.
It trails GPT-6 Astra on FrontierSWE v2, a from-scratch project build, and on Terminal-Bench 4.0, which covers running a Linux shell. On Arena.ai's crowd-voted leaderboards it places first in Text Arena but only eighth in Code Arena: WebDev and Agent Arena. If your work is agentic coding or terminal automation, a coding-first model is the better tool. Argon is for the reading and writing end of the job.
That distinction matters inside Felo. A long research report and a coding agent should not default to the same model just because one model is newer. Pick the model for the task, not the leaderboard.
What this changes in Felo Search
Felo Search pairs search with document understanding. A model with a 1M-token output ceiling changes what a single session can produce.
Research can end in a document. You can gather sources, then have the model write the synthesis in one pass instead of assembling it section by section across many prompts. The result is a draft with fewer seams and a clearer argument.
Whole document sets become tractable. Long-context retrieval is the capability Argon is strongest at. Contract sets, filings, and long reports stop being a retrieval problem you solve by chunking and start being a reading problem the model can hold.
Drafting where tone matters. Argon's reported profile, strong text and moderate coding, suits teams that need prose that reads as written, not assembled.
Argon access inside Felo Search is still in progress. Until it opens, every other model in the Felo lineup is available now. When Argon does open, the workflow is the same one Felo already supports:
- Bring the whole question. Give Felo the document set or the long thread, not a summary of it. Argon's advantage starts where the context stops fitting in a prompt.
- Ask for the finished artifact. With a 1M-token output ceiling, you can request the full report or complete draft in one instruction rather than in sections.
- Check it against the sources. Felo keeps citations attached to the answer, which is what makes a long model-generated document reviewable.
What this changes for the Felo API
The Felo API is in closed beta, and Gemini 4 Argon access is coming soon. There is no public date yet, and the API model catalog does not list Argon today.
When it lands, the integration path should look familiar. Felo API Platform exposes models through OpenAI-compatible Chat Completions and Responses endpoints, plus an Anthropic-compatible Messages endpoint. The pitch is one key and one base URL across model families, so switching to Argon should be a model-ID change rather than a rewrite.
The practical question for API users is not how to call Argon. It is what to call it with. A 1M-output model rewards requests that carry the full context and ask for a complete deliverable. It punishes the habit of sending a short prompt and stitching together sections. If your current pipeline does the stitching because the model could not hold the whole job, Argon is the reason to revisit that design.
Long output is not free. At the introductory rate, 1M output tokens cost $10; at the later rate, $20. A single long report is cheap by professional-services standards, but a batch job that generates long documents for thousands of records is not. Budget on the later price and measure output tokens per task, not just input.
The rollout, stage by stage
Google is releasing Argon in stages, and the first stage is not the public one.
| Stage | Status | What it means |
|---|---|---|
| Trusted cyber defenders | Available now | Selected partners in Google's Fairwind Program receive Argon first, without cyber guardrails |
| Paid API customers and Google AI Ultra | Announced next | Google says wider access begins here; no date has been given |
| Developers, enterprises, consumers | Not yet scheduled | Follows the paid tiers; Google has not committed to a timeline |
A subscription alone does not grant access today. That applies to Google AI Ultra and to Felo plans as well: Felo Search access is being enabled alongside Google's rollout, not ahead of it.
How to prepare
Three things are worth doing before access opens.
Collect the document sets you would never paste into a chat box. The model's advantage is proportional to the context you can give it. Have a contract set, a filing bundle, or a year of research notes ready.
Write the request as a deliverable, not a question. "Draft the full report, with sections for X, Y, and Z, citing each source" is a better test than "what does this mean?" The output ceiling only pays off if you ask for the whole thing.
Compare against your current model on the same task. Run one long task through Argon and through whatever you use today, then compare review time, not just generation time. A faster draft that needs a full rewrite is not faster.
FAQ
Is Gemini 4 Argon available on Felo now?
Not yet. Argon access inside Felo Search is in progress, and the Felo API is in closed beta with Argon access coming soon. The rest of the Felo model lineup is available now.
What is Gemini 4 Argon best at?
Reading and writing at length. Google's published leads are on knowledge work such as Harvey's Legal Agent Benchmark and Vals Finance Agent v2, long-context retrieval on GraphWalks, and long video understanding on LVBench at 91.7%. Its coding scores are competitive on DeepSWE v1.1 but weaker on from-scratch project builds.
How much does Gemini 4 Argon cost?
Google's introductory API price is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. That rises to $4 and $20 per million once the introductory period ends. Google has not said when the increase takes effect.
How is Gemini 4 Argon different from Gemini 4.0 Pro?
Argon is a separate release with a new naming scheme, not an update to the Gemini 4.0 line. Its defining change is the output ceiling, raised from 64K to 1M tokens, and it is the first Gemini flagship since Gemini 3 rather than an incremental refresh.
Is Gemini 4 Argon better than GPT-6 Astra and Claude Opus 5.5?
It depends on the workload. Argon leads both on DeepSWE v1.1 and on knowledge-work evaluations. It trails GPT-6 Astra on FrontierSWE v2 and on Terminal-Bench 4.0. There is no single winner across all tasks.
Does Gemini 4 Argon hallucinate less?
Artificial Analysis measured a 15% hallucination rate, the lowest it has recorded among frontier models. That means the model is more likely to decline rather than invent an answer, which matters most on long research output.
The bottom line
Gemini 4 Argon is not a better coding model. It is a model built to finish long knowledge work in one pass, and that is the gap Felo Search and the Felo API are set up to use: search across sources, hold the context, and produce a cited document.
Access is still staged. Google is starting with Fairwind partners, Felo Search access is in progress, and the Felo API is in closed beta. If your work is legal, financial, or research-heavy, the useful preparation is not another benchmark comparison. It is having a real document set and a real deliverable ready for the day the model shows up.
→ Try Felo Search with the models available today: https://felo.ai/search
→ Explore the Felo API Platform: https://openapi.felo.ai/models