---
title: "Gemini 4 Argon: Low Hallucination Rate, 4th on the Vals Index"
date: "2026-11-04"
video_id: "7704442822035787039"
tiktok_url: "https://www.tiktok.com/@vibewithkevin/video/7704442822035787039"
creator: "@vibewithkevin"
category: "AI"
tags: ["gemini 4 argon", "claude code", "hallucination rate", "vals index", "claude.md"]
excerpt: "Gemini 4 Argon has a 15% hallucination rate but only 50% accuracy, and ranks 4th on the Vals Index. What that means for using it as a planner for Claude Code."
---
Gemini 4 Argon is Google's model that Artificial Analysis reported with a 15% hallucination rate on its AA-Omniscience test, the lowest among top models. That figure is real, but it sits beside a 50% accuracy score, and the Vals Index page dated November 4, 2026 ranks Argon 4th rather than first. A TikTok from @vibewithkevin uses these numbers to argue for a CLAUDE.md line that has Claude Code ask Argon for a plan before it writes any code.

## What the TikTok recommends

The video builds on an X post by "Vox" (@Voxyz_ai), dated Oct 31, which reads: "For architecture questions, or asking how to simplify your code, the strongest model right now is Gemini 4 Argon. But don't let it write code. It hallucinates a lot." The post continues that building still goes to Opus 5.5.

The speaker turns that into a setup. A `## Planning` section in `~/.claude/CLAUDE.md` tells Claude Code: "On architecture or simplify questions, ask Argon for a plan first:", followed by a `gemini-api generate --model gemini-4-argon` command and a "PAID API KEY" tag. Claude then puts Argon's plan next to its own, says where they differ, and says which one deletes more code. Claude builds, and Argon only plans.

The video shows the snippet as text on screen. It never shows the command running, a plan coming back, or a comparison. The command is cut off at "read-", so the read-only flag is implied but not shown, and I could not confirm that a CLI called `gemini-api` exists. The Vox post also says Codex Astra should review the finished work, a step the video drops.

## Does Gemini 4 Argon really have the lowest hallucination rate?

Yes, under Artificial Analysis's definition. Its article dated September 30, 2026 reports a 15% hallucination rate on AA-Omniscience for Argon. It calls this the lowest among models scoring 45 or more on its Intelligence Index, where Argon scores 53. The same article gives 51% for GPT-6 Astra (max) and 54% for GPT-6.1 Sol (max).

The number the video leaves out is accuracy. Argon answered 50% of AA-Omniscience questions correctly. GPT-6 Astra answered 63%, and Argon also trails Gemini 3.1 Pro Preview by 5 points. On the overall AA-Omniscience score, Argon gets 42, Astra 43 and GPT-6.1 Sol 42.

A low hallucination rate on this test means the model says "I don't know" more often instead of guessing. It does not mean Argon gets more answers right. The speaker touches on this when he says Argon is "most likely to admit it doesn't know instead of guessing", but the on-screen card shows the 15% figure by itself.

## Is Argon number one on the Vals Index?

Not as of November 4, 2026. The video says "On the VALS index, which grades finance, legal, tax, and coding work, Argon debuted at number one." The Vals Index page, with results dated November 4, 2026 under version 2.2, lists Gemini 4 Argon in 4th place at 69.24% accuracy. It costs $13.37 per test and takes 44 min 58 s per test.

The index covers four sectors weighted by approximate share of U.S. GDP: finance, coding, legal and tax. Argon's scores on that page:

- Finance Agent v2.1: 67.7%
- Excel Modeling Benchmark v1.1: 74.8%
- Terminal-Bench 4: 57.6%
- Vibe Code Bench v1.1: 91.9%
- Code Migration v1.1: 74.4%
- Legal Research: 54.8%
- HLAB-Vals-1.1: 27.1%
- Tax Agent: 44.6%

On the separate Vibe Code Bench page, also dated November 4, Argon ranks second at 91.91%. That is a coding result, not the overall index. I could not confirm a first-place debut from a page I read.

## Does Argon trail Claude at editing and running code?

On Terminal Bench 4, yes, by a few points. The Artificial Analysis article lists Argon at 57%, Claude Opus 5.5 at 60%, Claude Sonnet 5.5 at 64% and GPT-6 Astra at 59%. Terminal-style tasks are the closest published measure of "run it and fix what breaks." The gap to Opus 5.5 is 3 points and to Sonnet 5.5 it is 7.

Argon is not weak at code across the board. The same article puts it first on AutomationBench-AA at 78%, ahead of Claude Sonnet 5.5 at 71%, and it scores 91.9% on Vibe Code Bench. The claim that Argon "hallucinates a lot" when writing code comes from the Vox post. None of the pages I read measure code-writing hallucination directly, so the planner and builder split is a workflow preference rather than a measured result.

## What Gemini 4 Argon costs

The "PAID API KEY" tag matches the pricing in the Artificial Analysis article. Standard prices are $4 per 1M input tokens and $20 per 1M output tokens. Discounted prices of $2 and $10 apply for at least one month, and the article says Google had not confirmed an end date. On September 30 the model was "being rolled out to selected users," so access was limited at launch. I did not check the current Gemini API docs for the model id, so `gemini-4-argon` is the video's string and not a confirmed identifier.

## Bottom line

The 15% hallucination figure is accurate, but it comes from Argon answering less often, and its accuracy on the same test is 50% against 63% for GPT-6 Astra. The Vals Index ranked Argon 4th on the date the video was posted. Using Argon as a read-only planner and Claude as the builder is one person's setup, and the video shows no runs, timings or quality comparison to back it.

## Sources

- [Gemini 4 Argon: Google is back as one of the top three labs in intelligence achieved, Artificial Analysis, September 30, 2026](https://artificialanalysis.ai/articles/gemini-4-argon-google-top-three-labs): hallucination rates, accuracy, Terminal Bench 4, AutomationBench-AA and pricing.
- [Vals Index, Vals AI, results dated November 4, 2026](https://www.vals.ai/benchmarks/vals_index): Argon's 4th-place rank and per-benchmark scores.
- [Vibe Code Bench, Vals AI, dated November 4, 2026](https://www.vals.ai/benchmarks/vibe-code): Argon's second-place rank at 91.91%.


## Video description and verification notes

**The file is a 113.7 s vertical clip (audio 113.714 s, video 113.667 s) at 1080x1920, 30 fps, in an MP4 (mov,mp4) container. The video is HEVC Main (hvc1) at 482,248 bit/s. The audio is HE-AACv2 stereo at 44.1 kHz and 48,066 bit/s. The file is 7,630,415 bytes at an overall 536,814 bit/s, and the format id is bytevc1_1080p_536809-1. The post timestamp is epoch 1793830384, which is 2026-11-04 22:13:04 UTC. The handle is @vibewithkevin and the nickname is "Vibe with Kevin". The audio credit is "original sound" by Vibe with Kevin. The description is 190 characters, with no URL. Engagement counts were captured 2026-11-05 13:03 UTC: 953 views, 32 likes, 3 comments, 14 reposts and 34 saves. I read ten frames, numbers 1, 10, 20, 25, 32, 37, 43, 50 and 56 in the burned-in layout plus the spoken timeline, and the frame times below are inferred from the 2 s extraction interval (frame N is at roughly 2(N-1) s). That inference matches the transcript, because frame 10 shows the tweet while the speaker is naming it near 18 s. I did not read the other frames. The transcript is 66 segments and roughly 330 words. The frames show a bearded man in a home setting speaking to camera. Frame 1 carries the caption card "Gemini plans, Claude builds, with one CLAUDE.md line". Frame 10 (about 18 s) shows a screenshot of an X post by "Vox" (@Voxyz_ai, Oct 31) reading "For architecture questions, or asking how to simplify your code, the strongest model right now is Gemini 4 Argon. But don't let it write code. It hallucinates a lot." The post continues "Building still goes to Opus 5.5. Once it's done, have Codex Astra give it a picky" and is cut off by "Show more", under a yellow label "STRONGEST FOR ARCHITECTURE". Frames 20 and 25 (about 38 s and 48 s) show a dark editor card titled "~/.claude/CLAUDE.md" with the heading "## Planning", the text "On architecture or simplify questions, ask Argon for a plan first:", and a typed command "$ gemini-api generate --model gemini-4-argon read-" that is still being typed in frame 25, beside a yellow "PAID API KEY" tag. Frame 32 (about 62 s) shows the caption "NOW, THE WARNING". Frame 37 (about 72 s) shows a card "AA-OMNISCIENCE · GEMINI 4 ARGON" with "15%" and the text "hallucination rate at launch. GPT-6 Astra: 51%". Frame 43 (about 84 s) shows a quoted passage tagged "The Batch · We're thinking" that says Argon "may not be the most powerful coder" but Vals AI's index shows strengths in finance, taxes and law, and that it is the frontier model most likely to admit it does not know a factual answer. Frame 50 (about 98 s) shows the labels "CLAUDE → BUILDS", "ARGON → PLANS" and "EDIT · RUN · FIX WHAT BREAKS". Frame 56 (about 110 s) is the speaker alone. No terminal output, no Gemini response, no plan comparison and no run of the setup appears in any frame I read. The spoken argument is that adding one CLAUDE.md line makes Claude Code ask Gemini 4 Argon for a read-only plan before building, because Argon is strong on architecture and honest about what it does not know, but weaker than Claude at editing and running code. The last transcript segment (id 65, "Thank you.") runs 113.12 s to 113.18 s against 113.714 s of audio, so it falls inside the audio and the transcript is not fabricated. Whisper wrote "cloud" for Claude, "Vox" is the on-screen name, and "payday API key" is "PAID API KEY" in frame 25.**

## The hallucination figure is real, but "lowest" comes with 50% accuracy

At 71.9 s the video says "Argon launched with the lowest hallucination rate of any top model, 15%", and at 73.6 s "OpenAI GPT-6 Astra, for example, was at 51." The Artificial Analysis article dated September 30, 2026 supports both numbers. It reports a 15% hallucination rate on AA-Omniscience for Argon, which it describes as the lowest among models scoring 45 or more on its Intelligence Index, and 51% for GPT-6 Astra (max). It gives 54% for GPT-6.1 Sol (max). So the figures and the "lowest of any top model" wording hold under the article's own definition of "top", which is a score of 45 or more.

The video omits what sits beside that number. The same article lists Argon's accuracy at 50%, which is 5 points below Gemini 3.1 Pro Preview and 13 points below GPT-6 Astra at 63%. Its overall AA-Omniscience score is 42, against 43 for Astra and 42 for GPT-6.1 Sol. A low hallucination rate on this test means the model refuses or says it does not know more often, and it does not mean it answers more questions correctly. The speaker half acknowledges this at 83.74 s to 86.78 s ("most likely to admit it doesn't know instead of guessing"), and frame 37 shows the 15% figure alone. A viewer who reads the card as "Argon is right more often" would be wrong: on this test Astra answered 63% correctly and Argon 50%.

The video also describes the test as one that "takes off points for wrong answers, but not for I don't know answers" (65.7 s to 69.3 s). That matches how the test is used to separate abstention from error, and I did not fetch a methodology page to check the exact scoring rule, so I treat the description as consistent with the article but not independently confirmed.

## The "number one on the Vals index" claim is out of date: 4th place on 4 November 2026

At 79.58 s to 82.12 s the video says "On the VALS index, which grades finance, legal, tax, and coding work, Argon debuted at number one." Frame 43 shows a Batch passage saying only that the index "shows its strengths in fields like finance, taxes, and law", and the frame does not say Argon ranks first. The live Vals Index page, whose results are dated November 4, 2026 under version 2.2, lists Gemini 4 Argon in 4th place with 69.24% accuracy, a cost of $13.37 per test and 44 min 58 s per test. The page says the index covers four sectors weighted by approximate share of U.S. GDP: finance, coding, legal and tax. Argon's sector benchmarks on that page are Finance Agent v2.1 at 67.7%, Excel Modeling Benchmark v1.1 at 74.8%, Terminal-Bench 4 at 57.6%, Vibe Code Bench v1.1 at 91.9%, Code Migration v1.1 at 74.4%, Legal Research at 54.8%, HLAB-Vals-1.1 at 27.1% and Tax Agent at 44.6%.

A search result summary from Vals pages indicated that Argon led the earlier v2.1 leaderboard on October 18 at 68.90%. I did not open that page, so I cannot confirm the "debuted at number one" claim from a source I read. "Debuted" is a past-tense word, so the video is not strictly wrong if Argon led at launch, but a viewer watching on November 4 or later would see it ranked 4th. The fetched Vals page for Vibe Code Bench, also dated November 4, 2026, ranks Argon second at 91.91%, which is the coding result, not the overall index. Verdict: partly supported and out of date. The first-place claim is unverified from a page I read, and the current position is 4th.

The video also says "other people are saying" Argon is most likely to admit it does not know. The Batch passage in frame 43 is the source on screen, and it makes that statement. I could not fetch The Batch itself, so I only record that the frame shows this text and that it agrees in direction with the 15% figure and 50% accuracy above.

## "Argon still trails Claude's models for editing files, running them, and fixing what breaks": consistent with the Terminal Bench 4 numbers

At 94 s to 98.6 s the video says Argon "still trails Claude's models for editing files, running them, and fixing what breaks." The Artificial Analysis article gives Terminal Bench 4 scores of 57% for Argon, 64% for Claude Sonnet 5.5, 60% for Claude Opus 5.5 and 59% for GPT-6 Astra. The Vals page gives Argon 57.6% on the same benchmark inside the coding sector. Terminal-style tasks are the closest published proxy for "run it and fix what breaks", so the claim holds for Terminal Bench 4. The gap to Opus 5.5 is 3 points, and to Sonnet 5.5 it is 7 points, so "trails" is accurate while "by a lot" would not be.

The same article lists Argon at number 1 on AutomationBench-AA at 78%, ahead of Claude Sonnet 5.5 at 71%, and the Vals page gives Argon 91.9% on Vibe Code Bench. Argon is therefore not weak at code by every measure, and the video's split is a workflow preference. The speaker says at 25.2 s that Argon "still hallucinates a lot" when writing code. The tweet in frame 10 makes the same statement, and I found no benchmark in the pages I read that measures code-writing hallucination specifically, so that point is the tweet author's judgment.

## The CLAUDE.md line: what the frames show and what they do not

At 42.58 s to 47.56 s the video says "On architecture or simplifying questions, Claude asks Argon for a plan first through Google's Gemini API tool with a payday API key." Frame 25 shows the snippet, with the tag "PAID API KEY". The tag is consistent with the pricing in the Artificial Analysis article: standard prices are $4 per 1M input tokens and $20 per 1M output tokens, and discounted prices of $2 and $10 apply for at least one month, which the article says Google had not confirmed an end date for. The article also says Argon was at that time "being rolled out to selected users", so on September 30 access was limited. I did not check the current Gemini API docs for the model id, so the string `gemini-4-argon` in the frame is the video's text and not a verified identifier.

At 50.14 s the video states "Argon can't edit any files or do any work." The command in frame 25 is cut off at "read-", presumably a flag such as "read-only", but the frame ends before the text completes. I did not see the flag finished in a frame I read. The frames never show a `gemini-api` tool being installed or run, and I could not verify that a CLI with that name exists. The read-only property would plausibly follow from calling a text-generation endpoint without tool access, since the model then returns text and nothing else, but that is an inference and not something the video demonstrates.

The video also says Claude puts Argon's plan "next to its own plan, tells you where they differ, and tells you which one deletes more code" (53.3 s to 57.5 s). This is a description of prompt behavior. No frame shows an example of such a comparison, and the "banger ideas" claim at 105.98 s is the speaker's report from a few days of use. Neither can be checked from the artifact. The speaker also tells viewers at 60 s to "Screenshot this and give it to your cloud or codex agent", which means the setup text lives in the frame and not in a link. The description has no URL, so there was no linked page to read.

The tweet in frame 10 ends with advice that Codex Astra review the finished work, and the video drops that step and keeps only the plan-with-Argon, build-with-Claude pair. The video also names the model "Opus 5.5" through the tweet and says "everyone's still sticking with Opus 5.5" at 30.88 s, which is a generalization from one tweet and a personal setup and has no measured basis in the sources I read.

## Timeline and attribution

The tweet in frame 10 is dated Oct 31 on screen, and the video was posted on 2026-11-04 22:13 UTC, so the video repeats a four-day-old opinion. The benchmark article is dated September 30, 2026, so the 15% and 51% figures are about five weeks old at posting and were produced at launch. The Vals results the video leans on moved between October 18 and November 4 according to the pages and snippets above, and the video presents the launch-time ranking without a date. The speaker says he has used the setup "for the past few days", and no frame contains a dated artifact of the work itself. The setup is one person's workflow, and the video gives no timing, cost or quality comparison of runs with and without Argon.

## Key Takeaways

- **Verified:** At 71.9 s the video says Argon had "the lowest hallucination rate of any top model, 15%" and that GPT-6 Astra was at 51. The Artificial Analysis article reports 15% for Argon and 51% for GPT-6 Astra (max) on AA-Omniscience, with "top" meaning an Intelligence Index score of 45 or more.
- **Partial correction:** The video presents 15% as a reliability gain without the accuracy figure. The same article lists Argon's AA-Omniscience accuracy at 50% against 63% for GPT-6 Astra, so part of the low rate comes from answering less often.
- **Partial correction:** At 82.12 s the video says Argon "debuted at number one" on the Vals Index. The Vals Index page dated November 4, 2026 (version 2.2) ranks Argon 4th at 69.24%. The earlier first-place result is unverified from a page I read.
- **Clarification:** The claim that Argon trails Claude on editing and fixing code matches Terminal Bench 4: Argon 57%, Claude Opus 5.5 60%, Claude Sonnet 5.5 64% and GPT-6 Astra 59%. Argon scores 91.9% on the Vals Vibe Code Bench, so the planner/builder split is a workflow choice and not a clear coding failure.
- **Unverified:** The read-only property of the `gemini-api generate --model gemini-4-argon` command is not confirmed, because the command is cut off at "read-" in frame 25 and no run appears in any frame I read. The existence of a `gemini-api` CLI is also unconfirmed, and the quality of Argon's plans ("banger ideas") rests on the speaker's report.
- **Context:** The claims about Argon's honesty come from a launch-week article and a Batch excerpt shown in frame 43. Rankings and pricing changed in the weeks before posting, and the discounted API prices were temporary according to the Artificial Analysis article.

## Resources

- [Gemini 4 Argon: Google is back as one of the top three labs in intelligence achieved, Artificial Analysis, September 30, 2026](https://artificialanalysis.ai/articles/gemini-4-argon-google-top-three-labs) establishes the 15% and 51% hallucination rates, the 50% versus 63% accuracy, the Intelligence Index score of 53, the Terminal Bench 4 scores, the AutomationBench-AA result and the $4/$20 and $2/$10 per 1M token prices I cite.
- [Vals Index, Vals AI, results dated November 4, 2026](https://www.vals.ai/benchmarks/vals_index) establishes Argon's 4th-place rank at 69.24%, the four-sector structure and the per-benchmark scores I cite.
- [Vibe Code Bench, Vals AI, dated November 4, 2026](https://www.vals.ai/benchmarks/vibe-code) establishes Argon's second-place rank at 91.91% on that coding benchmark.
