Gemini 4 Argon: Benchmarks, Availability, and How It Compares

Update (October 24, 2026): Cline's Argon default lasted one day. Cline's October 17 releases (SDK 0.0.95, CLI 3.0.73, desktop 0.0.48) refreshed the model catalog and the Cline provider default reverted to Claude Sonnet 5.5 — the recommended list leads with Sonnet 5.5 again. The October 16 flip is described below as it happened; if you are pinning defaults in Cline today, re-check the picker first.

Between October 13 and 16, 2026, Gemini 4 Argon stopped being a launch announcement and became a default. On October 13, Google published an Antigravity post putting the model in front of AI Pro and Ultra subscribers, and GitHub's changelog added Argon to Copilot across eight surfaces on the same day. Three days later, Cline flipped its first-party provider default to Argon. If you opened any of those tools this week, the new model on top of the list was not a coincidence — it was a coordinated rollout.

The model itself is not new-new: Google announced Gemini 4 Argon on September 30, 2026 as "our next era of frontier intelligence." What changed this week is reach. Argon is Google's flagship reasoning model aimed squarely at hard tasks — complex software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense — and the company is explicit that easier-going "everyday driver" models come later. This article is the field guide we wish existed on day one: where Argon actually lives as of October 17, what Google's internal numbers really say, what the official model card benchmarks show (including where it loses), and honest positioning against Claude Opus 5.5 and GPT-6.1 Sol.

Quick answer

• What it is: Google's flagship reasoning model (gemini-4-argon), positioned for the hardest problems; everyday-driver models are promised later.
• Context window: 1M-token unified context (combined input and output), per Google's model docs and model card. Knowledge cutoff: March 2026.
• Where it is live: Antigravity (Google AI Pro/Ultra, eligible Gemini Enterprise, Gemini platform API), Google AI Studio, GitHub Copilot (Pro+, Max, Business, Enterprise; 8 surfaces), the Gemini platform API, and — since October 16 — the default model on Cline's first-party provider.
• Benchmarks: strongest official scores in Vibe Code Bench (91.9%), LABBench 2 (88.8%), and knowledge-work benchmarks like Vals Index (68.9%); trails Claude Opus 5.5 on Terminal-bench 4.0 and trails GPT-6 Astra on FrontierSWE v2, per Google's own model card.
• vs Opus 5.5: benchmark splits are task-dependent; Argon leads knowledge work and long-context recall, Opus 5.5 leads terminal-heavy automation on the card's own numbers.
• vs GPT-6.1 Sol: not a like-for-like pairing — Argon is Google's flagship, Sol is OpenAI's near-flagship default. Most people will choose by ecosystem, not by score.
• Pricing: Google's model documentation does not publish per-token rates on the pages we checked; third-party coverage reports $2/$10 per million tokens. Treat any price as unconfirmed until Google lists it.

What Gemini 4 Argon is

Argon is the Gemini 4 generation's opening statement, and the positioning language is unusually consistent across Google's surfaces: this is a model for your hardest problems. The Antigravity post describes frontier performance "across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense." The Gemini API docs list it as the flagship reasoning model. The DeepMind model card groups its evaluations into agentic coding, ML engineering, knowledge work, science, computer use, and cybersecurity — not vibe-check categories.

The specs Google has published, verified on October 17, 2026:

Two things Google has not published where we could check: per-token pricing on the model pages, and an "everyday driver" Gemini 4 sibling. The absence of the second is deliberate — the Antigravity post closes by pointing at "future models that can be used as everyday drivers for an even broader range of builders." Read that as: Argon is the ceiling, not the daily default, and Google will sell you the daily default later.

Where you can actually use it

Access widened dramatically in the October 13–16 window. As of October 17:

Google Antigravity. Rolling out in the Antigravity desktop app and CLI for Google AI Pro and Ultra subscribers, on eligible Gemini Enterprise subscriptions, and via the API in Gemini platform. For context on the surface itself — what else runs there, how agents are billed — our Antigravity models overview covers the lineup, and the Antigravity CLI tutorial walks the terminal workflow where Argon does its agentic work.

GitHub Copilot. Added to the model picker on October 13 across eight surfaces: Visual Studio Code, Visual Studio, Copilot CLI, the coding agent, the Copilot app, JetBrains IDEs, Xcode, and Eclipse. Available to Copilot Pro+, Max, Business, and Enterprise. Two admin-relevant details from the changelog: billing is at provider list pricing under usage-based billing, and organization policy is enable-by-default — "new models are automatically enabled unless an administrator has turned off the global default or explicitly disables this model." Rollout is gradual, so the picker may lag the announcement.

Cline. Since October 16, Argon is the default model on the first-party Cline provider — more on this below, because it is the change most likely to have touched you without warning.

Google AI Studio and the Gemini platform API. The API docs confirm Argon is live to try in AI Studio, and gemini-4-argon is callable on the Gemini platform for pay-as-you-go API users.

The October 13 Antigravity post: what Google actually claimed

The most quotable document of the week is Google's own Antigravity post, which is essentially an internal-usage report dressed as a launch note. Every number below is from that post, dated October 13, 2026 — they are Google's claims about Google's tools, so weight them accordingly.

Scale: "over 100K users across Alphabet rely on Argon inside our internal version of Antigravity each month." Over half of Cloud sales, almost half of finance, and over a third of business and operations Googlers use the tool. Whatever else this is, it is not a demo.

Engineering throughput: over a four-week period, Google saw a 35% increase in agentic code alongside a 30% decrease in post-submit rollbacks. The pairing is the interesting part — more agent-written code that gets reverted less often is the actual claim, not just volume.

The YouTube comments frontend: the flagship case study. Argon's audit found that redundant emoji rendering alone accounted for nearly 45% of server overhead, leading to 12 targeted optimizations. The end state: the frontend's CPU footprint cut in half, server comment-load latency down 30 ms, and client end-to-end latency down 20 ms.

Bug triage: given 134 open bug reports across multiple services, Argon closed 97 stale or duplicate bugs in under 3 hours — a 72.4% queue reduction — with documented commit proof for each, and sorted the remaining 37 by owner with none left unassigned.

Language migration: a custom Antigravity skill used Argon to rewrite legacy Fortran into Rust, achieving a 1.8x performance boost plus memory safety.

Security audit — the one to remember: auditing an access-control change batch, Argon found four critical security bugs. Two earlier frontier models had audited the identical code and both concluded there were zero security holes. Note the correct count: two prior models, four bugs found. This is also a quiet shot in the model wars — "our model finds what the competition's flagships miss" is the entire argument for a hard-tasks flagship.

Beyond engineering: localization of hundreds of docs pages into 17 languages in hours, a sales-vetting workflow that cut research time by roughly 90% (~10 hours/week saved), 80 FP&A workflows where 15.5 hours of human direction produced about 335 human-equivalent hours over seven days, and a 3,200-line Python/SQL/dashboards data-center pipeline taken to production in days. Enterprise-knowledge-work anecdotes, exactly where Argon is positioned.

Gemini 4 Argon benchmarks: what the model card shows

Google's DeepMind model card is the only official numbers source, and — to Google's credit — it includes benchmarks where Argon loses. All scores below are from the card, published October 2026:

The head-to-head rows on the same card tell you where the edges are:

• Argon leads knowledge work decisively. Vals Index: Argon 68.9% vs Claude Opus 5.5 at 67.0% and GPT-6 Astra at 63.1%. Harvey's Legal Agent Benchmark: Argon 19.6% vs Astra 5.4% and Opus 5.5 at 3.8% — a wide gap in exactly the legal/finance territory the launch post emphasizes.
• Argon trails on pure SWE and terminal work. FrontierSWE v2: Astra 65.5% and Opus 5.5 at 62.3% versus Argon's 55.0%. Terminal-bench 4.0: Opus 5.5 leads at 66.4%, with Astra at 58.2% and Argon at 57.4%. PostTrainBench also goes to Opus 5.5 (49.3% vs 45.3%).
• Long-context recall holds up. GraphWalks stays at 99.7% BFS F1 at 128k and only sags to 84.2% across the 256k–1M range — consistent with the 1M unified window being real, not nominal.

Third-party aggregation is still settling, but the early picture is consistent with the card rather than with the hype. MindStudio's comparison roundup (October 2026) confirms the Vibe Code Bench lead, repeats that Opus 5.5 "currently holds an edge" for terminal-heavy automation and polished 3D/visual generation, and notes that Artificial Analysis's composite index puts Argon mid-pack — roughly even with Claude Fable 5.1. A hands-on review at Analytics Vidhya also caught the model making an unsupported claim in one launch-summary test, a useful reminder that flagship does not mean hallucination-proof. Treat all third-party numbers as indicative, not authoritative.

Gemini 4 Argon vs Claude Opus 5.5

This is the fairest flagship-vs-flagship pairing available right now, because Google's own model card publishes direct comparisons. Anthropic's Opus 5.5 has been the current Opus flagship since its September 22 launch, anchoring the 5.5 generation (Sonnet 5.5 on September 28, Haiku 5.5 on October 7).

The card's split, side by side:

How to read this honestly:

• Argon wins knowledge work; Opus wins the terminal. If your day is document-heavy analysis, legal/finance workflows, or long-context research over million-token corpora, Argon's card numbers are stronger. If your day is hands-on-keyboard agentic coding in a terminal — multi-step execution, tool loops, repo-scale SWE tasks — Opus 5.5's numbers lead, and third-party reviews agree with the card on this.
• One caveat applies to the whole table: these are scores Google chose to publish, on Google's selected harness. Independent replication is early. The directional gap on Harvey's (5x) is probably real; treat two-point gaps as noise until third parties replicate.
• Ecosystem decides as much as scores. Opus 5.5 lives in Claude Code, the Claude apps, and Copilot. Argon lives in Antigravity, Copilot, Cline, and the Gemini platform. Your harness lock-in is worth more than three benchmark points.
• Positioning is not symmetric. Both are flagships today, but Google has already told you Argon's "everyday driver" successor is coming — meaning Argon is positioned as the specialist you escalate to, while Opus 5.5 currently is Anthropic's daily driver at the top tier. Pricing is one place Argon may undercut: third-party reporting puts its API list price at $2/$10 per million tokens (reported to rise in December), which would make it cheaper than Opus-class rates — but Google has not confirmed pricing on the pages we checked, so treat that as provisional.

Gemini 4 Argon vs GPT-6.1 Sol

The second-highest-volume comparison this week is also the most apples-to-oranges. Per OpenAI's current model docs: GPT-6 Astra is the flagship ("our most capable model for complex work across code, apps, and research"), GPT-6.1 Sol is the recommended default for complex coding and agentic workflows at "near-Astra performance ... at a lower cost," and GPT-6 Luna is the efficiency tier. So "Argon vs Sol" pits Google's ceiling against OpenAI's workhorse — and the right mental model is Argon-vs-Astra, or Sol-vs-whatever-Google-ships-as-the-everyday-driver.

What can be said with sources attached:

• No official head-to-head exists. Google's model card compares Argon against GPT-6 Astra, not Sol. Anyone quoting precise Argon-vs-Sol benchmark deltas is aggregating third-party runs — interesting, not authoritative.
• Pricing is where they genuinely meet. GPT-6.1 Sol is verified at $2/$10 per million tokens — and the third-party reporting on Argon's pricing lands on the same $2/$10 figure (unconfirmed by Google). If both hold, the price excuse for choosing between them disappears, same as it did between Sol and Sonnet 5.5.
• Context windows are close but not identical. Sol runs a 1,050,000-token window; Argon's card documents 1M unified input+output. Near-wash in practice.
• Selection logic: for high-volume daily agentic coding where you want good-enough-at-1/5-flagship-cost, Sol's entire design brief is that job. For the occasional hard task where you want the model to find the four critical bugs the other models missed, you escalate to a flagship — Argon or GPT-6 Astra, both documented ceilings. Note also that OpenAI is retiring GPT-5.5 from Codex sign-in as of October 14, 2026, so laggards on old defaults are being pushed down this same decision tree.

Cline's default model is now Gemini 4 Argon

If you use Cline, this is the section that explains what changed under you. On October 16, 2026, Cline shipped four synchronized releases — desktop v0.0.47, CLI v3.0.72, VS Code 4.1.25, and SDK 0.0.94 — and the headline change is a default-model flip. From the release notes, verbatim:

With no model chosen, the Cline provider now uses Gemini 4 Argon, the new first model in Cline's recommended list.

The specifics worth knowing:

• Scope: only the first-party Cline provider flipped. OpenRouter still defaults to Claude Sonnet 5.5 — the release notes are explicit about it. If you route through OpenRouter, nothing changed for you.
• The recommended list changed too: Gemini 4 Argon is the new first entry, and Grok 4.7 was dropped from the list. The SDK release notes confirm the bundled default (first Cline recommended model) moved from Claude Sonnet 5.5 to Argon.
• You are not entirely safe if you use another provider: the same releases switched catalog defaults to Argon for ten other providers — including Google Gemini, GitHub Copilot, and Google Vertex AI among them — with the release summary warning that "if you use one of those providers without pinning a model, expect a different default." The practical takeaway: pin your model explicitly if the default matters to your workflow or your bill.
• The catalog also grew from 6,862 to 6,952 models in the same pass, and the free-model list swapped Step 5 Preview for Nemotron 3.5 Super VL.

Why it matters beyond Cline: Cline's recommended list is a distribution slot that shapes what hundreds of thousands of developers try first. Google putting Argon in that slot — and OpenRouter keeping Sonnet 5.5 as the counter-signal — is the clearest snapshot yet of how the model vendors are fighting for default position rather than just capability position.

Can you use Gemini 4 Argon for free?

Straight answer: there is no confirmed free tier for Argon itself as of October 17, and we will not invent one. What exists, by cost:

• Google AI Studio — closest to free. Google's own docs banner the model as live to try in AI Studio, which requires no product subscription. Whatever usage limits apply there are the loosest legitimate way to put hands on Argon today.
• Antigravity — subscription route. Requires Google AI Pro or Ultra (or an eligible Gemini Enterprise seat).
• Copilot — paid tiers only. Pro+, Max, Business, or Enterprise, and billed at provider list pricing under usage-based billing once you're in.
• Gemini platform API — pay per token. The predictable route for builders; pricing specifics await an official listing.

If your actual goal is "frontier coding model, minimal cost," the honest framing is that Argon is a new option, not a cheap one — the genuinely low-cost routes remain the ones we usually cover, like DeepSeek's Anthropic-compatible endpoint. What Argon changes is the free-trial calculus on Google's side: AI Studio means you can evaluate the flagship before deciding whether AI Pro/Ultra is worth a subscription.

FAQ

What is Gemini 4 Argon?

Google's flagship reasoning model, announced September 30, 2026 and rolled out across Antigravity, GitHub Copilot, Cline, Google AI Studio, and the Gemini platform API during October 13–16, 2026. It is positioned for the hardest tasks — complex software engineering, enterprise knowledge work, cybersecurity defense — with "everyday driver" models in the Gemini 4 family promised for later.

What is the Gemini 4 Argon context window?

1M tokens as a unified context window (combined input and output, with long decode up to 1M), per both the Gemini API docs and the DeepMind model card. The knowledge cutoff is March 2026. Long-context recall on the card's GraphWalks benchmark stays high across the 256k–1M range.

Gemini 4 Argon vs Claude Opus 5.5 — which is better?

Task-dependent, and Google's own model card says so. Argon leads knowledge-work benchmarks (Vals Index 68.9% vs 67.0%, Harvey's Legal 19.6% vs 3.8%) and long-context recall; Opus 5.5 leads terminal-heavy agentic coding (Terminal-bench 4.0 66.4% vs 57.4%) and FrontierSWE v2 (62.3% vs 55.0%). Choose by workload and by which harness your team already lives in.

Gemini 4 Argon vs GPT-6.1 Sol — which should I use?

They occupy different rungs: Argon is Google's flagship, while GPT-6.1 Sol is OpenAI's near-flagship default sitting below GPT-6 Astra. For high-volume everyday agentic coding, Sol is the cost-rational default at $2/$10 per million tokens. For the hardest single tasks, escalate to a flagship — Argon or Astra. No official benchmark head-to-head between Argon and Sol exists; be suspicious of precise numbers claiming otherwise.

Why did my Cline default model change to Gemini 4 Argon?

Because Cline's October 16 releases (VS Code 4.1.25, desktop 0.0.47, CLI 3.0.72, SDK 0.0.94) made Argon the first model in the recommended list and the default for the first-party Cline provider. OpenRouter's default remains Claude Sonnet 5.5, and you can pin any model explicitly — which is the recommended fix if you want your old default back.

Is Gemini 4 Argon available on free plans?

Not on any plan Google or its partners call "free," as verified October 17. The closest no-subscription route is trying it in Google AI Studio. Antigravity requires Google AI Pro/Ultra or eligible Gemini Enterprise; Copilot requires Pro+, Max, Business, or Enterprise. API access is pay-per-token via the Gemini platform.

Related reading

• Antigravity models overview — what runs in Google's agentic IDE and how the lineup fits together
• Antigravity CLI tutorial — the terminal workflow where Argon does its agentic work
• Cline desktop tutorial — configure providers and pin models after the default flip
• GPT-6.1 Sol pricing and context window — OpenAI's default-tier model, verified down to cache rates
• GPT-6 Astra guide — OpenAI's flagship, the fair comparison point for Argon
• Claude Haiku 5.5 — the third member of Anthropic's 5.5 generation