Home
About
Posts
Snippets
Insights
blog logoJooojub
Home
About
Posts
Snippets
Insights

Daily Briefing

May 24, 2026
AI Insightsbriefing
2026-05-23
31 articles

Anthropic’s Claude Mythos found 10,000 critical vulnerabilities in one month. The patches can’t keep up.

2026-05-23

Summary

Anthropic's AI model Claude Mythos discovered more than 10,000 critical vulnerabilities in one month, but the speed of fixes was not keeping up.

Key Points

  • Claude Mythos is an AI model specialized in detecting vulnerabilities in source code, making great progress in the field of security.
  • Of the more than 10,000 vulnerability candidates discovered, 1,094 were identified as high-risk or critical, but only 97 were patched.
  • The rate at which vulnerabilities are discovered is far outpacing the rate at which they are fixed, posing a serious challenge to the cybersecurity ecosystem.
Notable Quotes & Details

Notable Data / Quotes

  • 10,000+ critical flaws
  • 97 patched
  • 1,094 confirmed high- or critical-severity flaws
  • WolfSSL critical flaw (CVE-2026-5194, CVSS 9.1)
  • 1.5 million USD fraudulent wire transfer prevented

Intended Audience

Cybersecurity experts, software developers, IT decision makers

Berlin’s Peec AI more than doubled revenue to $10M ARR in six months. Its product helps brands show up in ChatGPT.

2026-05-23

Summary

Berlin-based startup Peec AI more than doubled annual recurring revenue (ARR) to over $10 million in six months with its generative engine optimization (GEO) platform.

Key Points

  • Peec AI provides Generative Engine Optimization (GEO) solutions that help companies monitor and improve brand visibility in search results for AI chatbots such as ChatGPT and Gemini.
  • In just 6 months after Series A investment, ARR grew rapidly from $4 million to $10 million, proving rapid growth.
  • Unlike traditional SEO, we manage brand exposure within AI chatbot answers to maximize marketing performance tailored to changes in consumer search behavior.
Notable Quotes & Details

Notable Data / Quotes

  • $10M ARR
  • $21M Series A
  • valuation above $100 million
  • $4M ARR

Intended Audience

Corporate marketers, startup officials, technology investors and strategists

SEO teams are tracking keywords. But are they tracking what ChatGPT says about their brand?

2026-05-23

Summary

Moz Pro has launched a new AI SEO toolkit that allows brands to track how much exposure they get to responses from AI assistants (ChatGPT, Gemini, etc.).

Key Points

  • Existing SEO tools only measure Google search rankings, missing out on brand visibility for AI assistant responses.
  • Moz Pro's new AI SEO toolkit tracks brand mentions and locations in responses from ChatGPT and Gemini through its AI Visibility dashboard.
  • The toolkit also includes tools to help with AI-based prompt research and content planning to support brands’ AI visibility strategies.
Notable Quotes & Details

Notable Data / Quotes

  • Moz Pro
  • AI Visibility dashboard
  • ChatGPT
  • Gemini
  • 20 years

Intended Audience

SEO expert, digital marketer, brand manager

Notes: Content incomplete

Elon Musk has given up on solar power (on Earth)

2026-05-23

Summary

Elon Musk addresses Tesla's strategic change to focus on space-based solar power as a power source for future xAI data centers instead of terrestrial solar power, which was Tesla's original goal.

Key Points

  • xAI is increasing its dependence on fossil fuels by introducing large-scale natural gas turbines without environmental regulations to operate data centers.
  • SpaceX's IPO documents confirm that Musk views space-based solar power, rather than terrestrial solar power, as the future power source for AI data centers.
  • SpaceX claims that space-based solar power can produce energy 24/7 and be more than five times more efficient than ground-based solar power, but high costs and technical difficulties remain challenges.
Notable Quotes & Details

Notable Data / Quotes

  • buy $2.8 billion more
  • SpaceX spent $131 million on 1,279 Cybertrucks
  • xAI has spent $697 million in the last two years on Tesla Megapacks
  • more than five-times the energy

Intended Audience

AI and energy industry workers, Tesla and SpaceX investors, and readers interested in IT technology news.

Google’s new anything-to-anything AI model is wild

2026-05-23

Summary

Analyzes the video generation function and performance of Google's newly introduced 'Omni' generative AI model and addresses improvements and limitations compared to the previous model, 'Veo'

Key Points

  • Google unveils 'Omni', a new generative AI model family that can transform various inputs such as text, photos, and video.
  • The first model, 'Omni Flash', was introduced into 'Flow', an AI video creation and editing platform, and has improved character consistency and ability to utilize real-world knowledge compared to the previous model, 'Veo'.
  • Despite improvements, there are still limitations in the image creation process where object consistency is broken or visual errors (glitches) occur.
Notable Quotes & Details

Notable Data / Quotes

  • Omni
  • Omni Flash
  • Flow
  • I see

Intended Audience

Tech enthusiasts, content creators, and industry insiders interested in AI technology

Nous Research Releases Contrastive Neuron Attribution (CNA): Sparse MLP Circuit Steering Without SAE Training or Weight Modification

2026-05-23

Summary

Nous Research has published a Contrastive Neuron Attribution (CNA) technique that can control a model's rejection mechanism for harmful requests without modifying the model's weights.

Key Points

  • CNA techniques analyze differences in activation of MLP neurons to identify specific neurons that distinguish between harmful and beneficial prompts.
  • It was revealed that the model's rejection behavior was not generated by fine tuning, but rather a structural characteristic that already existed in the base model.
  • Our tests show that adjusting just 0.1% of neurons in total MLP activation can reduce the rejection rate for harmful requests by more than 50% while maintaining the model's output quality.
Notable Quotes & Details

Notable Data / Quotes

  • 0.1% adjustment in MLP activation
  • Reduction in rejection rate by more than 50%
  • Maintain output quality above 0.97
  • Llama and Qwen architecture (applies 1B to 72B parameters)

Intended Audience

AI researcher, model interpretability researcher, LLM safety and alignment expert

Towards Speed-of-Light Text Generation with Nemotron-Labs Diffusion Language Models

2026-05-23

Summary

NVIDIA's Nemotron-Labs diffusion language model is a new model that improves speed and accuracy by breaking away from the existing sequential text generation method by generating and refining tokens in parallel.

Key Points

  • To overcome the limitations of sequential token generation in existing autoregressive methods, we introduced a diffusion language model (DLM) method that generates multiple tokens in parallel and iteratively refines them.
  • It maximizes GPU efficiency to increase inference performance, and generates tokens that can be modified on their own, which is advantageous for text correction and specific context filling tasks.
  • It provides three generation modes: autoregressive, diffusion, and self-guessing, allowing you to flexibly adjust speed and accuracy to suit your work situation.
Notable Quotes & Details

Notable Data / Quotes

  • 3B
  • 8B
  • 14B
  • NVIDIA Nemotron Open Model License
  • NVIDIA Source Code License
  • NVIDIA Megatron Bridge framework

Intended Audience

AI model developers and researchers, latency-sensitive application developers

DeepSeek makes V4 Pro price discount permanent

2026-05-23

Summary

DeepSeek has officially made the existing discounted pricing for the V4 Pro API permanent and updated its model specifications and pricing structure.

Key Points

  • The 75% discount on DeepSeek-V4-Pro API is officially permanent.
  • The input cache hit price of all models has been reduced to 1/10 of the existing launch price.
  • The existing deepseek-chat and deepseek-reasoner model names were abolished and the specifications were reorganized to V4 Flash and Pro models.
Notable Quotes & Details

Notable Data / Quotes

  • DeepSeek-V4-Pro API price reduction of 75% becomes permanent
  • Cache hit input token price reduced by 1/10 compared to launch price
  • Pricing effective from 12:15 UTC on April 26, 2026
  • 75% discount promotion ends May 31, 2026 15:59 UTC

Intended Audience

Developers and companies using AI model API

Antigravity 2.0 ranks #1 in OpenSCAD Architectural 3D LLM Benchmark

2026-05-23

Summary

This is a benchmark result that evaluated the ability of AI coding tools to implement buildings as parametric CAD codes using OpenSCAD.

Key Points

  • Google Antigravity 2.0 / Gemini 3.5 Flash High topped the benchmark with 4.5/5 points.
  • The assessment item was the ability to implement Pantheon as code based solely on reference images and prompts.
  • The performance of the model was determined by its ability to handle spatial geometry and the ability to verify the final mesh rather than simple tool accessibility.
Notable Quotes & Details

Notable Data / Quotes

  • Google Antigravity 2.0 / Gemini 3.5 Flash High: Quality 4.5/5 (1st place)
  • Cursor: Time 5/5, Quality 1.4/5 (fastest speed, lowest quality)
  • Codex 5.5 High: Quality 3.0/5 (good detail, but PNG and STL mismatch)

Intended Audience

AI researcher, parametric CAD developer, AI coding tool user

U.S. researchers face new restrictions on publishing with international collaborators

2026-05-23

Summary

Some grant managers at the U.S. NIH and NASA are imposing unofficial restrictions on publishing papers with co-authors from foreign institutions, causing confusion in the research field.

Key Points

  • Even though the research was conducted in the United States, prior approval or deletion of the existing paper is requested simply because there are co-authors from overseas institutions.
  • Without new official guidelines, administrators notified individuals individually, which had the side effect of encouraging researchers to preemptively exclude foreign co-authors in the future.
  • Academics are concerned that this measure violates standard publishing ethics and will seriously undermine international collaborative research practices.
Notable Quotes & Details

Notable Data / Quotes

  • In 2017, 30% of NIH-funded papers were collaborative studies involving both U.S. and non-U.S. authors.
  • MDI Biological Laboratory excluded 16 out of 22 papers scheduled to be reported due to issues with foreign co-authors.

Intended Audience

Scientific researchers, academic institution managers, and research policy makers

Show GN: Automating operations with CS Clothes that emerge little by little

2026-05-23

Summary

In order to streamline operational CS work, we developed a tool to automate DB and code question-answering in natural language using Claude CLI.

Key Points

  • Implemented an internal tool that automates CS queries through a web server running on a MacBook.
  • Claude CLI converts natural language questions into SQL queries or answers by referring to GitHub code and DB schema.
  • For security purposes, restrictions such as SELECT-only execution, LIMIT 1000 setting, and 30 second timeout are applied.
Notable Quotes & Details

Notable Data / Quotes

  • SELECT ONLY
  • LIMIT 1000
  • 30 second timeout

Intended Audience

IT operators and developers

Show GN: /idea2planning, which immediately turns ideas that come to mind into planning documents.

2026-05-23

Summary

This is an introduction to '/idea2planning', a skill for Claude Code that automatically converts ideas that come to mind into structured planning documents (1-pager, PRD, wireframe).

Key Points

  • It offers a three-stage automatic conversion pipeline that turns your ideas into 1-pager, PRD, and ASCII wireframes.
  • You can create a customized plan for your organization by referring to actual example documents provided by users in the RAG method.
  • It is designed with a structure that separates templates, principles, and example files to facilitate team-level improvement and expansion.
Notable Quotes & Details

Notable Data / Quotes

  • GitHub: https://github.com/pentaxzs/skill-idea2planning
  • Installation path: ~/.claude/skills/idea2planning

Intended Audience

Product managers (PMs) and product designers (PDs) who frequently need to write planning documents

Is personalized AI memory actually a problem worth solving or am I just coping[D]

2026-05-23

Summary

This raises the need to build a personalized database that goes beyond the simple memory function of current AI and evolves by learning the user's way of thinking and cognitive patterns.

Key Points

  • Currently, the memory function of AI (Claude, ChatGPT, etc.) is at the level of simply storing only the user's factual information, so it has limitations in reflecting the actual way of thinking.
  • Beyond simple session-to-session memory, you need a dynamic database that continuously learns users' thinking patterns, common failure areas, and effective explanation methods.
  • As the system interacts with the user, it can refine its cognitive profile and provide more personalized responses.
Notable Quotes & Details

Intended Audience

A community of developers and researchers interested in AI technology

OpenAI is hiring a $445,000 researcher. Requirements? Be 'tasteful and strategic.'

2026-05-23

Summary

There is news that OpenAI is hiring a ‘sensible and strategic’ researcher, offering an annual salary of $445,000.

Key Points

  • OpenAI is recruiting researchers with a high annual salary of $445,000.
  • As a recruitment requirement, a 'tasteful and strategic' attitude was emphasized along with technical capabilities.
  • This recruitment news is attracting attention in the IT community.
Notable Quotes & Details

Notable Data / Quotes

  • $445,000
  • 'tasteful and strategic'

Intended Audience

IT industry workers and those seeking employment in the AI ​​field

Notes: Content incomplete

Why can't people just run gemini and claude code using their own gpus?

2026-05-23

Summary

Regarding the computing resource limitations of Gemini and Claude Code, the question arises as to why users cannot run AI directly using their personal GPUs.

Key Points

  • Gemini and Claude Code are perceived to have limited or reduced functionality due to lack of resources or cost issues.
  • There is a demand from engineers and general users to directly run AI using idle GPUs on personal PCs.
Notable Quotes & Details

Intended Audience

AI developers, engineers, and tech community users

Notes: The text is very short

I wish there was a “Canva for AI training” already

2026-05-23

Summary

In order to solve the high technical entry barrier in the AI ​​model fine-tuning process, we believe that a user-friendly platform like 'Canva' that can easily learn and deploy models without infrastructure knowledge is needed.

Key Points

  • The current AI learning process is overly difficult and painful for ordinary users due to complex infrastructure management and technical errors.
  • Once the learning platform is abstracted, the focus of AI development will shift from a focus on technology to a focus on data quality and creativity.
  • The emergence of AI learning platforms such as 'Canva' or 'Shopify' that can complete everything from data upload to model deployment with a few clicks is inevitable.
Notable Quotes & Details

Notable Data / Quotes

  • Canva for AI training
  • Shopify moment for AI model training

Intended Audience

Non-experts and developers interested in developing AI models

I built a cognitive architecture where the AI has actual needs that drift between sessions — not prompt engineering, actual state variables

2026-05-23

Summary

This is about a case in which the cognitive architecture 'PHI // DRIFT' was developed with actual needs and state variables fluctuating between sessions.

Key Points

  • Implementing an AI architecture with real desires rather than prompted engineering through seven homeostatic state variables that fluctuate between sessions.
  • Memories are scored based on emotional significance and time passage rather than simple vector similarity.
  • Applying the shadow concept of Jungian psychology, a module that tracks behavioral patterns is included as a core element of the architecture.
Notable Quotes & Details

Notable Data / Quotes

  • 7 homeostatic state variables
  • 9 months
  • CPU-only mini tower

Intended Audience

AI developer, cognitive architecture researcher

Gemma4 26b a4b Apex quant is quite good

2026-05-23

Summary

User reviews about the outstanding performance and efficiency of the new Apex quantization technology for the Gemma 4 26B model.

Key Points

  • Using mudler's apex quant, we achieved a speed of 38 tps without quality loss in 90k context.
  • It showed stable performance over longer context processing than the previously used larger capacity quantization model.
  • Tested via llama.cpp Vulkan on RX 9060 XT 16 GB environment.
Notable Quotes & Details

Notable Data / Quotes

  • 38tps at 90.000 context
  • APEX-I-Compact (15gb)
  • RX 9060 XT 16 GB
  • unsloth ud-q5kxl quant (21.2gb)

Intended Audience

LLM Developer, Local AI Model User

G4-MeroMero-26B-A4B-it-uncensored-heretic Is Out Now, a Finetune of gemma-4-26B-A4B-it, With KLD of 0.0152 and 12/100 Refusals!

2026-05-23

Summary

A new uncensored version based on the gemma-4-26B-A4B-it model, the G4-MeroMero-26B-A4B-it-uncensored-heretic model, has been released.

Key Points

  • In response to user requests, a 26B-A4B version of the censorship release model was released, taking speed and memory efficiency into consideration.
  • Distributed in Safetensors and GGUF formats for use in a variety of environments.
  • It has lighter hardware requirements than the existing 31B version.
Notable Quotes & Details

Notable Data / Quotes

  • KLD of 0.0152
  • 12/100 Refusals
  • G4-MeroMero-26B-A4B-it-uncensored-heretic
  • gemma-4-26B-A4B-it

Intended Audience

Local LLM user and artificial intelligence model developer

meituan-longcat/LongCat-Video-Avatar-1.5 · Hugging Face

2026-05-23

Summary

'LongCat-Video-Avatar 1.5', an open source framework for creating high-quality audio-based human avatar videos, has been released.

Key Points

  • Equipped with Whisper-Large, the naturalness of lip movements has been greatly improved.
  • It supports a variety of functions such as voice-text based video creation and video continuation, and provides commercial-level stability.
  • DMD2-based technology enables efficient image generation with 8-step inference.
Notable Quotes & Details

Notable Data / Quotes

  • 8-Step Inference
  • MY License
  • 508 image-audio source pairs

Intended Audience

AI video creation technology developer and digital human creator

Llama.cpp VS LiteRT on a custom Xiaomi 12 Pro 24/7 Server (V2 Redesign)

2026-05-23

Summary

This is a case of modifying the Xiaomi 12 Pro smartphone to build an AI server that runs 24 hours a day and comparing the performance of Llama.cpp and LiteRT.

Key Points

  • Using the Snapdragon 8 Gen 1-based Xiaomi 12 Pro, the cooling system and power supply unit (PSU) were custom designed and converted into a 24/7 AI server.
  • Performance benchmarking with the gemma-4-E4B model shows that Llama.cpp has a low load, while LiteRT is slightly faster in generation but has high CPU utilization and power consumption.
  • If you are having difficulty utilizing GPU acceleration and your purpose is simply to use the LLM server, we advise that it is more efficient to purchase a mini PC.
Notable Quotes & Details

Notable Data / Quotes

  • Llama.cpp: Prompt 30.6 t/s, Generation 5.7 t/s
  • Cooling automation: 40°C on, 35°C off
  • PSU protection: 4.3V crowbar circuit applied

Intended Audience

Tech enthusiasts interested in DIY hardware modifications and building on-device AI

What is the current best Small Language Model that can be run without GPU?

2026-05-23

Summary

This is a discussion among users about the best small language model (SLM) that can be efficiently run on CPU without GPU.

Key Points

  • We are looking for a compact language model with high performance and speed that can run in a CPU environment without GPU acceleration.
  • The key criteria for model selection are accuracy and inference speed.
  • We are asking users to share their experiences about which deployment stack they use in real environments.
Notable Quotes & Details

Intended Audience

Developers, users interested in running AI models locally

Notes: Content incomplete

Best Buy is selling this 2TB Corsair SSD for over 60% off right now - and I vouch for it

2026-05-23

Summary

The news is that Best Buy is selling Corsair's 2TB external SSD at a 62% discount.

Key Points

  • The Corsair EX400U 2TB external SSD is on sale for $330, a 62% discount at Best Buy.
  • It supports USB-C and Thunderbolt 4 and offers read speeds of up to 4,000 MB/s and write speeds of 3,600 MB/s.
  • It has a compact design that is MagSafe compatible and convenient to use with Apple devices, including the iPhone.
Notable Quotes & Details

Notable Data / Quotes

  • 62% off
  • $330
  • 2TB
  • 4,000 MB/s
  • 3,600 MB/s

Intended Audience

Users who require large storage space and fast transfer speeds, such as gamers, creative professionals, and mobile device users

Best Buy and Amazon has dropped major SSDs prices - I found the best storage deals

2026-05-23

Summary

Best Buy and Amazon are selling SSDs at steep discounts ahead of Memorial Day.

Key Points

  • Amid the recent rise in PC component prices, we offer discounts on SSDs from major brands such as Western Digital, Samsung, and SanDisk.
  • This is a high-capacity SSD upgrade opportunity for PC, laptop, PS5, and Xbox Series X|S users.
  • Major discount products were introduced, including a 62% discount on SanDisk 8TB SSD and a discount of more than $1,100 on WD Black 4TB SSD.
Notable Quotes & Details

Notable Data / Quotes

  • SanDisk 8TB SSD 62% discount
  • Save over $1,100 on WD Black 4TB SSD
  • WD Black SSD read/write speeds up to 7300/6600 MB/s
  • WD Black SSD endurance up to 2,400TB

Intended Audience

Consumers looking to purchase high-capacity storage devices for PCs and gaming consoles at an affordable price

Google Cloud Introduces Cross-Engine Iceberg Support in BigQuery

2026-05-23

Summary

Google Cloud has introduced cross-engine support for Apache Iceberg in BigQuery, enabling interoperability and unified data management across multiple platforms without data duplication.

Key Points

  • BigQuery provides a preview of the serverless Iceberg REST catalog, allowing multiple engines, including Spark, Flink, and Trino, to query the same Iceberg table without duplicating data.
  • Managed support for metadata management, table maintenance, and synchronization tasks reduces operational complexity and improves efficiency.
  • We expanded the cross-cloud lakehouse ecosystem by linking the Iceberg catalog with external platforms such as AWS, Azure, Databricks, and Snowflake and supporting AI workflows.
Notable Quotes & Details

Notable Data / Quotes

  • Apache Iceberg
  • BigQuery
  • Next ’26

Intended Audience

Data Engineer, Data Architect, Cloud Platform Manager

Claude Mythos AI Finds 10,000 High-Severity Flaws in Widely Used Software

2026-05-23

Summary

Anthropic's project 'Glasswing' and AI model 'Claude Mythos Preview' discovered more than 10,000 vulnerabilities in large-scale open source software.

Key Points

  • The Claude Mythos Preview model identified over 10,000 potential vulnerabilities in over 1,000 open source projects.
  • Of the vulnerabilities identified, 1,726 were confirmed to be real threats, of which 1,094 were classified as high or critical risk.
  • Anthropic called for accelerated security measures, noting that the slow pace of patching vulnerabilities compared to the speed of discovery is a major security challenge.
Notable Quotes & Details

Notable Data / Quotes

  • Over 10,000 vulnerabilities discovered
  • 6,202 high or critical vulnerability candidates
  • 1,726 True Positives
  • 1,094 actual high or critical vulnerabilities
  • Critical flaw in WolfSSL CVE-2026-5194 (CVSS score: 9.1)
  • 97 upstream patches and 88 advisories issued
  • How to prevent a $1.5 million wire transfer fraud

Intended Audience

Security expert, software developer, IT system administrator

Pichai: "Google coding is lagging behind other companies... We will turn the tables to a virtuous data cycle structure."

2026-05-23

Summary

Google CEO Sundar Pichai acknowledged that he was lagging behind in the AI ​​coding competition and announced a strategy to overcome this through a virtuous data cycle structure.

Key Points

  • CEO Pichai officially admitted that Google's coding AI is less competitive than Antropic or Open AI.
  • The core cause of the difference in coding competitiveness is analyzed as the lack of a surface for collecting developer real-time workflow data.
  • We plan to bridge the gap with a 'data flywheel' structure that increases model performance by securing and strengthening real-time data.
  • CEO Pichai announced the prediction that AGI (Artificial General Intelligence) will be possible around 2030.
Notable Quotes & Details

Notable Data / Quotes

  • Token usage mentioned at Google I/O ‘doubles every few weeks’
  • CEO Pichai's expected time to achieve AGI is '2030'

Intended Audience

AI technology industry insiders, investors, IT developers

Why did Chinese telecommunications companies jump into the AI ​​token rate system?

2026-05-23

Summary

Chinese telecommunications companies are introducing token-based billing plans and AI services in earnest to monetize their investments in AI computing infrastructure.

Key Points

  • Major Chinese telecommunications companies such as China Mobile, China Telecom, and China Unicom have launched AI token plans and various AI model support services for businesses and general users.
  • Telecom companies are converting massive AI infrastructure investments into commercial services and building new revenue models through packaged products combined with security.
  • This move is interpreted as the Chinese government's policy intention to utilize cheap energy and real estate in the western inland to develop an AI data center base and make it the core of the global AI economy.
Notable Quotes & Details

Notable Data / Quotes

  • As of March, daily token consumption exceeded 140 trillion (40% increase compared to the beginning of this year)

Intended Audience

AI and communications industry insiders, investors, IT policy experts

[Ahn Gwang-seop AI Truth] Is Google taking control of AI watermarks as well?

2026-05-23

Summary

Google's AI watermarking technology 'SynthID' has become a de facto standard for AI content trust infrastructure along with the open standard C2PA, and we analyze Google's strategy to lead the AI ​​verification ecosystem through this.

Key Points

  • Google DeepMind's AI watermarking technology 'SynthID' is expanding its influence by being applied to 100 billion images and videos and 60,000 years of audio.
  • Major AI companies such as OpenAI, Nvidia, and Kakao are complementing the limitations of C2PA, the existing metadata method, by introducing Synth ID into their products.
  • Google is taking a platform strategy to design and dominate the overall AI content reliability verification ecosystem by standardizing verification technology and securing infrastructure leadership.
Notable Quotes & Details

Notable Data / Quotes

  • SynthID application: 100 billion images and videos, 60,000 years of audio
  • C2PA member companies: more than 6,000 as of January 2026
  • Number of SynthID verifications: over 50 million times

Intended Audience

AI industry practitioners, IT strategists, and general readers interested in technology trends

Notes: Content incomplete

Gemini 3.5 Flash, performance has taken a leap... The cost of turning is 5.5 times

2026-05-23

Summary

Google's new model Gemini 3.5 flash has greatly improved performance, but the cost has increased 5.5 times compared to the same task, raising controversy over its efficiency.

Key Points

  • Gemini 3.5 Flash outperformed leading models such as Claude Opus 4.7 and GPT-5.5 in coding and agent tasks.
  • With the leap in performance comes a significant increase in operating costs, with agent tasks costing 75% more than Gemini 3.1 Pro.
  • It shows some performance limitations in difficult tests that require extensive knowledge, and its price positions it as a middle ground between existing models.
Notable Quotes & Details

Notable Data / Quotes

  • 5.5x increase in cost for the same task
  • Total cost of agent work 75% higher than Gemini 3.1 Pro
  • Input $1.5, output $9 per million tokens

Intended Audience

AI technology developers, companies and officials using artificial intelligence models

PreviousDaily Briefing
NextDaily Briefing
Jooojub
System S/W engineer
Explore Tags
Series
    Recent Post
    © 2026. jooojub. All right reserved.