Gemini 4 Argon: Low Hallucination Rate, 4th on the Vals Index
By Matt Cullerton, founder of Mavric and Apso
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: hallucination rates, accuracy, Terminal Bench 4, AutomationBench-AA and pricing.
- Vals Index, Vals AI, results dated November 4, 2026: Argon's 4th-place rank and per-benchmark scores.
- Vibe Code Bench, Vals AI, dated November 4, 2026: Argon's second-place rank at 91.91%.
Published November 4, 2026. Based on a TikTok by @vibewithkevin. Full video notes (markdown)