Gemini 4 Argon vs Gemini 3.8 Flash: What Changes at the Top

On October 13, 2026, Google made Gemini 4 Argon generally available to developers, according to its launch post. That moves Gemini 4 Argon from a defender-only preview to a model any paid Gemini API customer can call, and it puts a direct question to teams that standardized on Gemini 3.8 Flash at high thinking: what do you gain by moving up, and what does it cost? This post compares the two on the numbers Google and independent trackers have published.

2.67xArgon vs Flash price per token$4 / $20 vs $1.50 / $7.50 per 1M, permanent rates

53 vs 41Artificial Analysis Intelligence IndexBoth at high thinking, v4.3.2

57% vs 20%Terminal-Bench 4.0Argon vs Flash, Artificial Analysis

15.1% vs 55.2%AA-Omniscience hallucination rateLower is better

Google

Gemini

Gemini 4 Argon and Gemini 3.8 Flash are both Google models, one generation apart.

Gemini 4 Argon general availability: what changed since launch

Google first announced the model on September 30 with access limited to cyber defenders, which we covered in our launch-day analysis of Argon. General availability is the October 13 step. Google's launch post says Argon is available through the Gemini API paid tier in AI Studio and Android Studio, in the Gemini app and AI Mode in Search for Google AI Ultra and Pro subscribers, and in Antigravity for Ultra subscribers. The Fairwind program still offers cyber defenders access without cyber guardrails on a rolling basis. Per Google's pricing page, the model is not available on the API Free tier.

Google's Gradient Canopy building in Mountain View, California, with a white sculpture in front of its wavy roofline — Google's Gradient Canopy building in Mountain View, California. A file photo of the campus, not an image of Gemini 4 Argon or its launch. Photo: GualdimG, Wikimedia Commons, CC BY-SA 4.0.

Gemini 3.8 Flash, released in September, is covered in our Flash benchmark breakdown. Both models take text, image, video, audio and PDF input, produce text only, and share a 1,048,576-token window, according to Google's model pages for Argon and Flash.

Gemini 4 Argon vs Gemini 3.8 Flash on price

At the permanent rates on Google's pricing page, Argon costs $4.00 per million input tokens and $20.00 per million output tokens, with cached input at $0.20. Flash costs $1.50 and $7.50, with cached input at $0.15. Both input and output work out to 2.67 times Flash's rate. Both models are on lower introductory pricing for now: Argon at $2 and $10 through November 30, 2026, and Flash at $0.75 and $3.75 through December 31, 2026. Neither model lists a long-context surcharge, and grounding with Google Search or Maps is billed identically: 5,000 free requests a month shared across Gemini 3 and newer models, then $14 per 1,000.

Per-token price is not per-task price. Artificial Analysis ran both at high thinking on its Intelligence Index v4.3.2, and its head-to-head page lists $1.99 per task for Argon against $1.24 for Flash, a ratio of about 1.6. The gap is smaller than the rate card because Argon used fewer tokens: 62,000 output tokens per task against 71,000 for Flash. Those costs reflect the introductory rates. Both models' rates double on the permanent schedule, so the same token counts would keep the ratio near 1.6.

$1.99 vs $1.24AA cost per index taskArgon vs Flash at high thinking, introductory rates

62k vs 71kOutput tokens per taskArgon vs Flash

$2,407 vs $1,623Cost to run the full indexArtificial Analysis

Benchmarks: where the gap is wide and where it closes

On Artificial Analysis's current index, Argon at high scores 53 against 41 for Flash at high, both on version 4.3.2. Do not compare this 41 with the 59 Flash scored in September: the index was re-based, so only same-version pairs are meaningful.

Gemini 4 Argon vs Gemini 3.8 Flash, same benchmark both sides

Scores in percent. The gap is widest on long-horizon terminal and software work and narrows to a few points on science and long-context reading, where Flash is level or marginally ahead on AA-LCR.

Sources: Artificial Analysis head-to-head page (Intelligence Index v4.3.2), BenchLM, Google launch materials. Google-reported rows are vendor figures, not independent tests.

The widest gaps are on long, agentic work. On Terminal-Bench 4.0 Argon scores 57% against 20% for Flash, and on FrontierSWE v2 55.0% against 19.6%, per BenchLM's listings of the Proximal leaderboard. Google-reported figures point the same way but narrower: OSWorld-2.0 69.2% against 59.0%, Vals Finance Agent v2 65.4% against 61.4%, and LABBench2 88.8% against 86.2%. On long-context retrieval the two are level, with Flash a point ahead on AA-LCR (81% to 80%).

Argon works longer and harder to solve highly complex scientific, coding, and agentic problems.
Google, Gemini 4 Argon launch post

Reliability shows the biggest swing. BenchLM, citing Artificial Analysis, lists a hallucination rate of 15.1% for Argon (rank 2 of 185) against 55.2% for Flash (rank 68). It does not state a benchmark version, and lower is better. Google adds its own results: 87% on ERQA, 57% on RareBench, 52.2% on LabBench, first place on the Vals Index, and a 0.7% attack success rate on the Gray Swan indirect prompt injection benchmark across 15 adaptive attempts. These are vendor figures.

Speed, thinking levels and where Flash still fits

Argon is built to think longer, and Google's docs warn that "hard prompts can take several minutes" and recommend background execution or client timeouts of at least 15 minutes. Artificial Analysis lists 24.48 seconds to the first answer token for Flash at high thinking, and had published no speed figures for Argon when we checked.

Both models offer low, medium and high thinking, with medium the default. Artificial Analysis's Flash page shows index scores of 33, 40 and 41 across the three levels, so the step from medium to high adds one point on Flash. More thinking can also hurt lookup-shaped work: on Roboflow's vision evals, visual identification fell from 97.9% at low effort to 96.9% at high, at 2.3 times the cost per sample. Google advises running evaluations at all three levels and choosing the lowest that meets your quality bar. For quick lookups and short answers, Flash at low or medium remains the sensible tool; Argon is for work that needs the extra depth.

Migrating to Gemini 4 Argon: the developer gotchas

Google's latest-model guide lists several changes that matter when swapping the model ID:

No explicit caching. The cachedContents API is not supported. Implicit caching is on by default for prompts of at least 4,096 tokens, with a 95% discount on cache hits. Flash supports explicit caching.
Sampling parameters are ignored. Temperature, top_p and top_k are removed from requests, and if sent they are ignored.
Raise output limits. The default maximum output is 65,536 tokens. If your app caps output below 32K, Google recommends at least 32,768, because Argon can spend more of the limit on thinking.
Long decode needs background execution. Responses above 65,536 tokens, up to 1M, run in the background, and you must join the text from all model output steps.
Lengthen timeouts to at least 15 minutes for hard prompts.

Like Flash, Argon has no Live API and no image or audio generation.

How Metir uses the two models

Gemini 4 Argon is now available in Metir's model picker as the high-tier Gemini model, replacing Gemini 3.8 Flash at high thinking. Gemini 3.8 Flash remains for the fast and thinking tiers. The split follows the evidence above: depth where the extra reasoning pays back, speed where it does not.

Takeaways

Argon is about 2.7 times Flash per token but about 1.6 times per task on Artificial Analysis's index, since it uses fewer tokens.
The gap is largest on terminal, software and agentic benchmarks, and close on long-context retrieval and several science tests.
Budget for latency and rewrite caching if you rely on explicit caches.
Both models carry introductory prices that end within three months, so model cost at permanent rates.

Sources:

Image credits

Header image: Google's Gradient Canopy building in Mountain View, California, by GualdimG, Wikimedia Commons, CC BY-SA 4.0 (source). In-body image: the same building seen from Shoreline Blvd, by GualdimG, Wikimedia Commons, CC BY-SA 4.0 (source). Both are file photos of Google's campus, not images of Gemini 4 Argon or its launch.