Gemini 3.8 Flash : mieux que Fable et GPT Sol, c'est quoi ce bordel ?

Gemini 3.8 Flash : mieux que Fable et GPT Sol, c'est quoi ce bordel ?

Gemini 3.8 Flash: better than Fable and GPT Sol, what the hell is that?

Title history
  • Gemini 3.8 Flash : mieux que Fable et GPT Sol, c'est quoi ce bordel ? — observed on 20261101-0040 UTC (YouTube)

by Melvynx · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 Melvynx 👥 67K 📅 September 3, 2026 ⏱ 18 min 👁 18K 📄 expert opinion 🧭 2026-11-01
Available in: English (current) Français

Keywords

Gemini 3.8 Flashbenchmarktoken usageCursorcost per task

Summary

This video reviews Google’s newly released Gemini 3.8 Flash model, focusing on its performance, cost, and real-world coding capabilities. The creator begins by highlighting that Gemini 3.8 Flash achieved top scores on the DeepSWE benchmark, outperforming models like Claude Opus (Fable) and GPT-5.6 while being significantly cheaper. He then analyzes cost per task on Artificial Analysis, noting that the model is about three times cheaper than competitors despite using more tokens. He conducts a series of tests, including a car crash animation test, a Gmail clone, and a landing page, comparing results with other models and raising suspicions of cheating based on response patterns. The video also covers real Cursor usage, showing the model’s performance in fixing code, and discusses token efficiency and speed, highlighting that Gemini consumes more tokens but remains cost-effective. The creator reviews tweets and benchmarks like CursorBench, noting improvements over Gemini 3.7. He concludes by recommending Gemini 3.8 Flash for its value, especially in Cursor, and mentions that he uses it alongside other models. Overall, the video provides a practical, opinionated assessment of the new model.

180 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video offers valuable insights into Gemini 3.8 Flash’s practical performance, particularly its cost-effectiveness and speed. The creator provides concrete examples, such as cost comparisons and token usage data, which are useful for developers considering adoption. However, the argumentation is largely based on anecdotal tests and personal preferences, with limited statistical rigor. The cheating detection, for instance, relies on heuristics and is not thoroughly validated. Nevertheless, the creator clearly distinguishes between objective benchmarks and subjective impressions, giving the viewer a balanced view. The argumentation is persuasive in showing that Gemini 3.8 Flash is a strong value proposition, but it lacks systematic comparison across multiple runs and controlled conditions.

Scientific Rigor, Source Quality, Title Accuracy

The creator references benchmarks like DeepSWE and CursorBench, and uses Artificial Analysis for cost data, but does not provide direct links to these sources beyond the description’s generic links. The sources cited are primarily the creator’s own links to courses and social media, which are not related to the technical claims. This limits the verifiability of the data. The title accurately matches the content: the video indeed tests Gemini 3.8 Flash against other models and discusses its surprising performance. The lack of external sources and reliance on personal testing reduces the overall scientific rigor, though the creator is transparent about his methodology and potential biases.

228 words

Title / Content Match

The title accurately reflects the content: the video tests and evaluates Gemini 3.8 Flash, comparing it to other models like Fable (likely Claude Opus) and GPT Sol (GPT-5.x) in terms of performance and cost.

Quality & Reliability

7/10

The creator presents his own hands-on tests, real benchmark comparisons, and cost analysis, but the evaluation is partially subjective and relies on anecdotal tests rather than systematic methodology.

Chapters

Cited Sources

Contribution & Novelties

The main added value of this video is the practical, hands-on comparison of Gemini 3.8 Flash with other leading models, focusing on cost, token usage, and real coding tasks, rather than just benchmark scores. It provides a user perspective on whether the model is worth integrating into development workflows. The creator also highlights the potential ‘cheating’ issue in benchmarks, which is an important consideration for evaluating AI models.

Pour aller plus loin :

  • Artificial Analysis — Independent platform for comparing AI model costs and performance; relevant for verifying the cost and speed claims.
  • Gemini — Official Google DeepMind page for the Gemini model family, useful for context on the model’s capabilities and versions.
  • Cursor — AI-powered code editor used in the video’s tests; relevant for understanding the integration context.
  • Token efficiency is a key concept; see Transformer models for background on token processing.

143 words

Radar Profile

The radar profile shows balanced scores across quantity and quality of information, with technical level and reliability slightly lower. This indicates the video is informative but relies on subjective testing and lacks rigorous sourcing, making it a moderately reliable commentary.

Reliability 6/10