Google's New AI Was Benchmarked on a Setting You Don't Get

By TechPulse USA Newsroom — Mountain View, California | 15 October 2026

Two days after Google put its most capable AI model in paying users' hands, the story of Gemini 4 Argon has split in two. One half is the rollout Sundar Pichai announced on X: a frontier model, finally public, doing serious work inside Google's own engineering org. The other half is the fine print nobody put in the launch tweet — and it is the more interesting half.

We pulled up Google's October 13 blog post and Artificial Analysis's live leaderboard side by side and read them line by line before writing a word of this.

Rollout day two: what actually happened on October 13

The announcement came straight from Pichai's X account: "Gemini 4 Argon is here." The details underneath it matter more than the sentence. Argon went live globally for developers and enterprises in the paid tier of the Gemini API and the Gemini platform, for Ultra subscribers inside Google Antigravity, and for Ultra and Pro subscribers in the Gemini app and AI Mode in Search. That ended a two-week restricted run: a small circle of cybersecurity professionals in Google's Fairwind Program got first access, deliberately unshielded, to throw real vulnerability research at it.

The pricing is introductory and the clock is loud: $2 per million input tokens and $10 per million output, with cached inputs at a 95% discount, through November 30. After that, rates double to $4 and $20. Google is also giving 10,000 scientists six months of free access to Antigravity's highest usage tier, and early enterprise testers include Deloitte and LexisNexis.

Reuters reported the launch arrives after a rocky stretch — a scrapped Gemini 3.5 Pro that Pichai had once promised for June, and a DeepMind overhaul after co-founder Demis Hassabis stepped aside. DeepMind chief AI architect Koray Kavukcuoglu says the model "excels at deep work in complex, real-world tasks." But "deep work" turns out to have a settings menu.

The benchmark asterisk: your MEDIUM isn't their HIGH

This is the story that changes how every Argon number should be read. Argon ships with three thinking levels — LOW, MEDIUM and HIGH. The standard model defaults to MEDIUM. The cyber-permissive Fairwind variant defaults to HIGH. And the Intelligence Index scores that Artificial Analysis published — the scores Google's defenders have been quoting at rivals — reflect the HIGH configuration.

Tech Times, which caught this first, put the precise numbers on it: on Artificial Analysis's Intelligence Index, Argon at high reasoning scores 52.56. That is within half a point of GPT-6 Astra at 52.67 — a tie, more or less. But it sits five points below Claude Sonnet 5.5 (56.00) and five below Claude Opus 5.5 (57.62). Everyday Pro and Ultra subscribers in the Gemini app at the default MEDIUM setting are on a different — and likely lower — capability profile than the charts suggest.

Google's own methodology page says most Argon results use its highest thinking setting; it did not hide it. But a launch-tweet benchmark and a methodology footnote are different documents, and the internet only read one of them. Sam Witteveen, chief executive of Red Dragon AI, told VentureBeat: "This is a really, really, really good model. But is it the best model out there? I actually don't think it is," he said.

Two leaderboards, two different stories

Argon is genuinely strange on independent tests: first place on one respected leaderboard, eighth on another, and both results can be true at once. Vals AI's Vals Index — GDP-weighted across finance, coding, legal and tax work — has Argon at No. 1 with 68.90%, ahead of Claude Sonnet 5.5 (67.04%) and Opus 5.5 (66.97%). It also tops Vals' Finance Agent v2 (65.40%) and AutomationBench-AA's agentic business tasks (77.5%).

The cooler leaderboard is Artificial Analysis's Intelligence Index: Argon at 53, level with GPT-6 Astra and Claude Fable 5.1, eighth of 216 models. Terminal-Bench 4: 57% — a 53-point jump over Gemini 3.1 Pro Preview, but behind Sonnet 5.5, Opus 5.5 and Astra. DeepSWE v1.1, though: Argon led outright at 77.9% against Opus 5.5's 74.2% and Astra's 74.1%.

The number that should travel furthest is not the ranking. It is the hallucination stat. On Artificial Analysis's AA-Omniscience test — which punishes wrong answers but not refusals — Argon posted a 15% hallucination rate, the lowest of any model scoring above 45 on the index.

Astra sat at 51%; GPT-6.1 Sol at 54%. In practice, Argon would rather say it doesn't know than guess. The trade-off: its answer accuracy was 50%, thirteen points below Astra. The most honest frontier model, and on raw factuality one of the least accurate.

The Chrome refactor: six million lines, weeks not months

Photo: Lines of code on a monitor. Google says Argon helped its Chrome team complete an architectural refactor across hundreds of Chromium files in weeks. (Markus Spiske, Wikimedia Commons, CC0)

All company-reported, but the internal engineering story is the one that would sell Argon to a CTO. Google says Argon helped the Chrome team separate core webpage handling from a roughly six-million-line legacy folder of browser features inside Chromium's 40-million-line codebase. The work spanned hundreds of files and took weeks, not the months the team expected. The payoff: live webpages moving between tabs, side panels and floating windows without reloading — and an entire category of browser crashes eliminated.

The supporting numbers are just as loud. More than 100,000 employees now use Argon monthly through an internal version of Antigravity, per VentureBeat. Agentic code submissions rose more than 35%; rollbacks fell by nearly a third. Argon reviewed 134 aging bugs and flagged 97 as stale or duplicated in under three hours. Its optimizations, now live in the open-source Agent Development Kit, cut in-memory session prep time by 40% and sped up cold imports by 21%. It made the open-source xTB quantum chemistry package up to 2.3x faster at structure prediction and 1.8x at conformer search, and delivered a 2.85x inference speedup for Gemma 4 E2B on an Nvidia A100 via vLLM.

Then the cornhole: Google says Argon watched a person play the game, built a simulation, and helped a humanoid robot reproduce the throws — scoring 87% on the Embodied Reasoning benchmark.

The $1.99-a-task price that doubles at the end of November

Artificial Analysis's cost math is the part developers should screenshot. At introductory pricing, Argon costs about $1.99 per Intelligence Index task — roughly 60% of GPT-6 Astra's cost for equivalent intelligence. After November 30, that flips: $3.98 a task, about 120% of Astra's current rate. The promo is real, but also a deadline; anyone building on Argon now is locking in architecture on a price that expires in six weeks.

One quirk in the developer guide before you port anything: on Argon, temperature, top_p, top_k and candidate_count are accepted — and silently ignored. Reasoning depth is a named level (low, medium, high; medium default). Test all three directly.

This is also the week Google's free tier got meaningfully thinner — when Google shipped three new Gemini models last month, free accounts were left on Flash-Lite only, and AI Plus subscribers are now being told they will lose manual Pro-model selection from November 13 unless they upgrade.

Who gets what: Pro, Ultra, Plus and the Fairwind door

Photo: The Chromium logo. The six-million-line refactor happened inside Chromium's roughly 40-million-line codebase. (The Chromium Authors, Wikimedia Commons, public domain)

The access matrix is tiered enough to deserve its own section. In the Gemini app and AI Mode in Search, both Pro ($19.99/mo) and Ultra ($99.99/mo) subscribers get Argon. But the multi-threaded agentic coding in Antigravity — where the model edits multiple files, runs tests and iterates on its own — is Ultra-only, as is AI Studio access. Pro gets Argon-powered code generation in the app's Canvas mode, a lighter single-session environment. And for the most demanding programming tasks, Google says Gemini 3.1 Pro stays available anyway, which tells you something about Argon's rough edges.

When Pichai posted his announcement on X, the platform's owner replied in the thread with a single word: "Congrats!" — Elon Musk, magnanimous for a day. Pichai used the launch to show off multi-step tool use: Argon generated an exploded-view animation, lighting, camera moves and audio from a raw Pixel CAD file — running, he said, 1,000 iterative rounds to check its renders and refine the code. "It has a much stronger grasp of coding and multi-step tool use," he wrote.

Meanwhile the Fairwind Program, which hands vetted cyber defenders an Argon variant without the usual cyber guardrails, is expanding to hundreds of partners. One early-access partner, security firm Wiz, used Argon to find and fix a file-download bypass in healthcare software that exposed a master key and administrator access — protecting, Google says, hundreds of thousands of patient records.

Our take: the scoreboard is honest, the frame is slippery

Here is the judgment, plainly stated. The thinking-level fine print is not a scandal — Google documented it, and Anthropic's Opus 5.5 is also evaluated at maximum effort — but it is slippery framing. A benchmark run at HIGH is being quoted at people who will use MEDIUM — the difference between the advertised product and the one subscribers actually hold.

But dismissing Argon over the asterisk would be the dumber mistake. The 15% hallucination rate is the single most interesting number in this launch cycle: a frontier model that declines to answer is a different tool than one that guesses confidently, and for legal briefs, financial analysis and medical-adjacent work it is arguably the better one. Pair that with six million lines of the world's most-used browser codebase refactored in weeks, and the picture is not a hype cycle. It is Google — after months of delays, a cancelled 3.5 Pro and a leadership shakeup — finally shipping something that does the boring, expensive, multi-step work it kept promising.

Meanwhile the competitive frame keeps tightening: Anthropic turned profitable and Wall Street noticed, and OpenAI's GPT-6 line is now being out-honested, if not out-reasoned, at 60% of the price — for six more weeks, anyway.

Gemini 4 Argon: your questions answered

Is Gemini 4 Argon free?

No. The Gemini app requires Pro ($19.99/month) or Ultra ($99.99/month), the API is billed per token, and Antigravity's agentic coding is Ultra-only. Free accounts stay on Flash-Lite, and AI Plus subscribers lose manual Pro-model selection from November 13 unless they upgrade.

Which thinking mode will I actually get?

In the Gemini app, MEDIUM — the default. The HIGH scores on benchmark charts come from the Fairwind variant and developers who set thinking_level to high in the API. Everyone else gets medium.

How much does the API cost?

Through November 30: $2 per million input tokens, $10 per million output, with cached inputs discounted 95%. From December 1, those rates double to $4 and $20. Artificial Analysis puts the per-task cost at about $1.99 now and $3.98 after the promo ends.

Is Argon actually better than GPT-6 Astra?

It depends what you mean by better. On the Intelligence Index they tie at 53 (52.56 vs 52.67 — Astra, by a nose). Argon hallucinates far less: 15% vs Astra's 51% on AA-Omniscience — but answers fewer factual questions correctly: 50% vs 63%. On professional workflows it leads the Vals Index outright. The safer pair of hands; Astra the sharper recall.

When do free users get Argon?

Google has not said. Pichai's September pledge was to expand availability "as soon as possible," with no date attached. Given the free tier just lost Flash and Pro models, do not hold your breath.

Sources: Reuters; Tech Times; Artificial Analysis; VentureBeat; InfoWorld; Google DeepMind blog; ANI via The Hindu BusinessLine; digit.in; Gizbot; Neowin. Internal rollout and engineering figures are Google-reported, not independently replicated. This is a follow-up to our October 1 report on the Gemini 4 Argon launch.

TechPulse USA is a newsroom tracking how artificial intelligence is actually being bought, sold and deployed — not just announced.