Cowry Tech News
AI Cloud

Google opens Gemini 4 Argon to paying customers. The bill and the benchmarks need a closer look.

Gemini 4 Argon is now on the paid Gemini API, in Gemini for business and in Google's top consumer plans. Google says it beats OpenAI and Anthropic on most of its own benchmark table. Early independent tests mostly back the "frontier" label, but they also flag a token-hungry model whose launch price doubles on December 1.

By Cole Harlan · 15 Oct 2026 · 14 min

Google Gemini 4 Argon launch header artwork from the Google Keyword blog
Image: Google

Google has opened Gemini 4 Argon to the public, 13 days after limiting it to vetted security teams. It is the company’s first new flagship since the Gemini 3 series, and Google is pitching it as its answer to OpenAI and Anthropic at the top end.

The general release came on October 13. Google DeepMind SVP and chief AI architect Koray Kavukcuoglu said in a launch post that Argon is “globally available today for developers and enterprises via the API in AI Studio and Gemini platform.” Google AI Ultra subscribers get it in Google Antigravity, the company’s agent-first coding tool. Ultra and Pro subscribers get it in the Gemini app and in AI Mode in Search.

Below: what changed in the model and the API, what the price works out to, how the rival comparisons were built, and what builders and IT buyers still do not know.

What Gemini 4 Argon is

Google’s model card calls Argon “a large scale sparse mixture-of-experts (MoE) transformer-based model” with native support for text, vision and audio inputs. It outputs text only. The card says Argon “is not a modification or a fine-tune of a prior model.” It was trained on Google’s TPUs, using JAX and ML Pathways. Its knowledge cutoff is March 2026.

Google pitches it at long, hard work. The developer docs describe it as a “flagship reasoning model, built for complex software engineering, deep mathematical and scientific problem solving, autonomous multi-step agents, and long decode up to 1M tokens.” The September 30 announcement post listed three target areas: software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense.

Argon is the first model in the Gemini 4 family, according to Google Cloud’s developer guide. The launch post ends with a promise to bring “the rest of the Gemini 4 family” later. Google has not said what those models are or when they ship.

What is new versus earlier Gemini models

The biggest technical change is output length. Google says Argon’s output limit rises to 1 million tokens, “up from the previous 64K tokens.” VentureBeat reported the same jump in its September 30 coverage.

The detail is in the docs. Argon has one shared window of 1,048,576 tokens. Input, thinking and output all draw from it. A single request still stops at 65,536 output tokens by default, or 32,768 for non-streaming direct HTTP calls on Google Cloud. To go further, developers use what Google calls long decode. The response is generated in hops of up to 64K tokens, and each hop picks up where the last one ended. The Google Gen AI SDK follows a continuation_token across those hops automatically. On the Gemini API, background execution does the same on Google’s servers.

Google Cloud’s guide flags a known issue: long decoding on GenerateContent works only on the v1beta1 API, not v1. Developers who pinned v1 will need to change it.

Other changes matter for anyone migrating code:

  • Thinking is set by level only: low, medium (the default) or high. Google Cloud’s guide says setting thinking_budget, or a MINIMAL level, returns an HTTP 400 error.
  • Caching is implicit only. Explicit context caching is not supported. On Google Cloud, implicit caching applies to shared prompt prefixes of 6,144 tokens or more.
  • Agentic video understanding is now generally available, though off by default.
  • Argon does not generate images or audio and is not supported on the Live API, per the model page.
  • Computer use is listed as a preview feature.

Google Cloud’s guide also names a second model ID, gemini-4-argon-cyber-permissive, which defaults to high thinking. Google says it is offering Argon “without cyber guardrails” to approved defenders through its Fairwind Program. The docs do not say who can call that ID.

On quality, Google claims Argon “answers questions more directly with less hedging” and “skips clichés.” The Verge relayed the same claim. That is a vendor description. We found no independent test of it.

Who gets it, and where

Access depends on the plan and the product.

  • Developers: Paid tier of the Gemini API, in Google AI Studio and Android Studio. Google’s pricing page says Argon “is not available on the Free tier or on the Google AI Pro plan.” Ultra subscribers can also use it in AI Studio and Antigravity.
  • Enterprises: Gemini for business, the Gemini Enterprise Agent Platform API on Google Cloud, and Antigravity. Google Cloud’s release notes say the model is generally available in the Gemini Enterprise web and mobile apps and in Workflow Builder. It runs in the global, us and eu multi-regions, with data residency and ML processing support in the US and EU.
  • Consumers: Pro and Ultra subscribers in the Gemini app and AI Mode in Search.

IT teams should note two settings. In Gemini Enterprise, the Argon toggle is off by default. Google’s release notes say it “defaults to on beginning October 27, 2026.” Admins who want to control the rollout have about a week and a half. In AI developer tools, Argon is enabled by default for licensed users unless the organization runs a custom model list.

Pro subscribers can chat with Argon in the Gemini app. They cannot use it through the Pro plan in AI Studio or the API. The Verge noted that it is not clear when, or if, Argon will be available to Plus subscribers or free users. Google has not said.

What it costs

Google’s pricing page sets two rates. Through November 30, 2026, standard API calls cost $2 per million input tokens and $10 per million output tokens. From December 1, that becomes $4 and $20. Output pricing includes thinking tokens. Cached input costs $0.10 per million now and $0.20 later, a 95% discount.

The page also lists other tiers. Batch and Flex run at half the standard rate: $1 and $5 now, then $2 and $10. Priority costs $3.60 and $18 now, then $7.20 and $36. Search and Maps grounding each include 5,000 free requests a month across Gemini 3 and newer models, then $14 per 1,000.

Google Cloud’s launch post, by product VP Michael Gerstenhaber, gives the same rates for enterprise API use.

When Argon was announced on September 30, Google gave the $2/$10 price but no end date. VentureBeat and Artificial Analysis both noted that gap. The October 13 post closed it.

How that compares with OpenAI and Anthropic

Argon’s list price is low for a frontier model. Anthropic’s pricing page lists Claude Opus 5.5 at $4 input and $20 output per million tokens, and Claude Fable 5.1 at $10 and $50. OpenAI’s model page lists GPT-6 Astra at $10 input and $50 output per million tokens.

At those rates, Argon’s launch price is a fifth of Astra and Fable 5.1, and half of Opus 5.5. The Verge made the same “around a fifth” comparison. From December 1, Argon matches Opus 5.5 per token and stays below Astra.

The Verge warned that lower token prices do not “necessarily translate to lower costs in practice, as models can use different numbers of tokens for the same tasks.” Artificial Analysis put numbers on that. It found Argon averaged 62,000 output tokens per task on its index, against 27,000 for GPT-6 Astra at max effort. At launch prices, Argon cost $1.99 per index task, 60% of Astra’s $3.26. At standard prices, Artificial Analysis estimates $3.98, about 1.2 times Astra.

On those figures, the price edge holds through November. For long reasoning jobs it could shrink or reverse after that.

The benchmarks: Google’s claims, then the independent checks

Google says Argon outperforms “other frontier models on benchmarks across domains.” The figures below are Google’s. Some come from third-party leaderboards; Google ran others itself.

Google’s own posts give these headline scores:

Benchmark Argon score (Google-reported) What it measures
DeepSWE v1.1 77.9% Long-horizon software engineering
AutomationBench (Zapier) 51.3% End-to-end business tasks
LVBench 91.7% Long video understanding
CWE-bench v1 68% (tied first) Fixing security flaws
ERQA 87% Embodied, spatial reasoning
Gray Swan IPI 0.7% attack success Prompt-injection resistance

Google also says Argon ranks first on the Vals Index, a third-party index weighted across finance, coding, legal and tax work.

VentureBeat read the full comparison table Google shared before launch. By its count, Argon led outright on 12 of 18 benchmarks and tied on one. It reported Argon at 19.6% on Harvey’s Legal Agent Benchmark, against 5.4% for GPT-6 Astra and 3.8% for Claude Opus 5.5. It also reported where Argon trails. GPT-6 Astra led on FrontierSWE v2, 65.5% to 55.0%, and on Terminal-Bench Science 0.1, 68.1% to 57.6%. Claude Opus 5.5 led on Terminal-bench 4.0, 66.4% to 57.4%, and on PostTrainBench, 49.3% to 45.3%. VentureBeat’s verdict: “model choice is still workload-dependent.”

What independent evaluators found

We checked two independent leaderboards. Both put Argon at or near the top.

Artificial Analysis scores Argon at high reasoning at 53 on its Intelligence Index. That ties GPT-6 Astra at max effort and sits one point ahead of GPT-6.1 Sol. It is 23 points above Gemini 3.1 Pro Preview. Artificial Analysis said the gains were “driven by lower hallucinations and stronger agentic capabilities.” On its AA-Omniscience test, Argon’s hallucination rate was 15%, against 51% for GPT-6 Astra. Its accuracy there was lower, at 50% versus Astra’s 63%. Argon declines to answer more often, and it also gets fewer answers right.

Artificial Analysis also ranked Argon first on its AutomationBench-AA at 78%. On Terminal Bench 4 it scored 57%, behind Claude Sonnet 5.5, Claude Opus 5.5 and GPT-6 Astra. Google’s own table showed the same gap.

On LMArena’s text leaderboard, as of the October 13 update, gemini-4-argon-high ranks first with a score of 1525, marked “Preliminary,” on 4,903 votes. Meta’s Muse Wonder Max is second at 1513, and Claude Opus 5.5 high is third at 1508. The margins sit inside overlapping rank spreads, and the vote count is still small.

Neither source tests every Google claim. We found no independent check of the DeepSWE, LVBench, ERQA or Gray Swan scores. Those stay single-source, attributed to Google.

What it means for builders

For developers, the main change is long output. An agent can now run much longer before it hands work back to a person. That has a cost in code. Teams need to handle continuation tokens, pick the v1beta1 endpoint, set max_output_tokens high enough, and budget for thinking tokens that are billed as output.

The migration checklist in Google’s docs is short. Change the model ID. Review max_output_tokens. Drop thinking_budget. Replace explicit caches. Google says it also ships a gemini-api-dev skill that a coding agent can use to migrate a codebase.

Third-party support arrived on day one, Google’s developer guide says, across tools including Vercel, LiteLLM, OpenRouter and Pydantic. Google also says Argon is available in Managed Agents through the Antigravity agent.

Google’s examples of Argon at work are its own. It says Argon in Antigravity raised agentic code submissions by more than 35% and cut code rollbacks by nearly a third inside Google. It says Argon helped speed up Gemma 4 inference 2.85 times on an A100 via vLLM, and that Argon agents produced a Rust video decoder 2.7 times faster than an earlier Rust port. These are vendor claims. Some point to public code, such as pull requests and the google/adk-python repository, but we have not reviewed that code.

What it means for enterprises

Google is selling Argon to the enterprise as a model for long, checkable work. The Google Cloud post says it can “help automate multi-step ledger reconciliations, spreadsheet formulas, and variance analyses.” It quotes customers. Harvey’s head of applied research, Niko Grupen, said the company “saw substantial improvements over Gemini 3.1 Pro across our legal agent evaluations.” ServiceNow’s Nenshad Bardoliwalla said Argon “delivered roughly 20% improved task success than Gemini 3.8 Flash” on its NowAI-Bench in early access. Both are partner statements published by Google.

The safety story is part of the pitch. The model card says Google “cannot rule out” that Argon reached its Frontier Safety Framework’s critical capability levels for biological risk and for cybersecurity, and cannot rule out a lower alert threshold for stealth and situational awareness. In each case Google says it built a safety case or risk assessment and applied safeguards that bring risk below those levels. The launch post says the model’s shortcut attempts on coding benchmarks fell to 2%, and that it runs a monitor that reads the agent’s reasoning to block “unintended workarounds.”

Some early tests are mixed. Customer-support software firm IrisAgent published a benchmark run on its own support suites. It found Argon never invented an answer on a ticket it should have declined. It also found Argon’s slowest 1% of responses were about four times slower than Claude Sonnet 5.5 and about nine times slower than GPT-6 Sol. IrisAgent said it is “not routing production traffic to it yet.” That is one vendor’s test on one workload. Buyers will want their own tests.

How we got here

Argon follows a long gap. The Verge said it is “Google’s first major frontier release since Gemini 3 in late 2025,” and that rivals’ models had left Google trailing. VentureBeat, citing reports from Reuters and Axios, described investor worries over a delayed Gemini 3.5 Pro and a leadership shake-up in August that put Kavukcuoglu in charge of DeepMind.

The rollout was staged. On September 30, Google released Argon only to “trusted cyber defenders” through Fairwind. It said it was “actively engaged in the U.S. government’s voluntary process for pre-release model access.” The general release came 13 days later. The Verge noted it landed a day after Meta announced its own frontier model, Muse Wonder Full.

Open questions

  • After December 1: Prices double. Artificial Analysis’s math suggests Argon’s cost per task could then exceed GPT-6 Astra’s.
  • The permissive cyber model: Google says Fairwind is expanding to “hundreds of additional” partners. The docs do not set out who can call gemini-4-argon-cyber-permissive, or how use is audited.
  • Rate limits: We did not find published Argon rate limits for the paid tier in the pages we read.
  • Plus and free users: No date has been given.
  • Latency: Google’s model card warns of “occasional slowness or timeout issues.” IrisAgent’s results point the same way. No broad independent latency study has been published.
  • The rest of Gemini 4: Google says more models are coming. It has not said which, or when.
  • Unverified numbers: DeepSWE, LVBench, ERQA and Gray Swan scores are still Google’s alone.

The next checkpoint is December 1, when the introductory price ends and Artificial Analysis’s cost-per-task estimate rises to $3.98.

Sources