K
Koda Intelligence
boltDaily Signal
S&P 5007,165.08↑ 0.80% NASDAQ24,836.60↑ 1.63% BTC$77,726↑ 0.34% ETH$2,318↑ 0.08% DEEPSEEK V482%· APR SHIP IBM Q1 REV$15.92B↑ 9% CURSOR VAL$60B· TARGET S&P 5007,165.08↑ 0.80% NASDAQ24,836.60↑ 1.63% BTC$77,726↑ 0.34% ETH$2,318↑ 0.08% DEEPSEEK V482%· APR SHIP IBM Q1 REV$15.92B↑ 9% CURSOR VAL$60B· TARGET
THE SIGNAL · 25 APR 2026 · 5 MIN READ

Chinese Labs Now Beat, Match, and Bypass American AI on Three Benchmarks in One Week

DeepSeek V4 runs on Huawei chips, ZEDAI GLM 5.1 tops GPT-5.4 on coding, and Qwen 3.6 scores big on reasoning. China's AI ecosystem just proved it can compete without American hardware.

AI BENCHMARKS · GEOPOLITICS · MARKETS · DIPLOMACY
Lead Story Model Release OpenAI

GPT-5.5 Pro Now Live in API

OpenAI released GPT-5.5 and GPT-5.5 Pro to the API on April 24, one day after the model's announcement. The company says GPT-5.5 matches GPT-5.4 per-token latency while delivering substantially stronger performance in agentic coding, computer use, and scientific research. OpenAI describes it as...

Continue Reading arrow_forward
smart_toy AI

OpenAI launched GPT-5.5 Pro in its API, but the headline belongs to challengers: DeepSeek V4 runs on Huawei silicon, ZEDAI GLM 5.1 outscores GPT-5.4 on coding, and Qwen 3.6-27B hits 0.9 on GPQA.

public World

Iran presented formal peace proposals at multilateral talks in Islamabad, while the Israel-Lebanon ceasefire was extended another three weeks, offering rare dual signals of de-escalation.

trending_up Markets

Sentiment sits firmly in Fear territory as Big Tech's $50 million 2025 lobbying tab and geopolitical uncertainty weigh on investor confidence despite a wave of new model releases.

bolt Wild Card

France charged 88 suspects in a crypto kidnapping ring, exposing a violent underworld where physical coercion has become the preferred exploit against digital wealth holders.

Market Snapshot

query_stats

S&P 500

7,165.08

+0.80%

Nasdaq

24,836.60

+1.63%

Bitcoin

$$77,726.16

+0.34%

ETH

$2,318.12

+0.08%

Crude Oil

$105.33

-0.60%

Fear & Greed

31

Fear

Today's Focus

01

China's Triple Benchmark Surge

DeepSeek V4 running on Huawei chips proves Chinese labs can sidestep U.S. export controls entirely. ZEDAI GLM 5.1 topping GPT-5.4 on coding and Qwen 3.6-27B scoring 0.9 on GPQA mean American frontier models now face credible competition on performance, not just price. The strategic implications for chip sanctions are severe: restriction is no longer containment.

02

Big Tech's Lobbying Blitz

A collective $50 million spent on lobbying in 2025 signals that major AI companies view regulation as an existential variable, not a background nuisance. This spending surge coincides with active congressional debates over compute export rules, liability frameworks, and open-source licensing. The money is buying time, and possibly shaping the rules that will determine which business models survive.

03

Diplomacy's Fragile Window

Iran's peace proposals in Islamabad and the extended Israel-Lebanon ceasefire represent the most concentrated diplomatic activity in the region in months. Neither development guarantees lasting stability; coordinated armed attacks across Mali and a deadly shooting in Louisiana remind us that violence remains decentralized and persistent. Markets have not priced in optimism, and for now, that caution looks warranted.

Listen & Watch

Daily Broadcasts
Daily Deep Dive
Podcast ~22 min

Daily Deep Dive

Video Briefing
play_arrow
Video

Video Briefing

Daily Infographic
Infographic

Intelligence Map

The Wire

AI Intelligence
Policy

Big Tech Spends $50M Lobbying in 2025

An Issue One analysis of federal filings found that seven major tech and AI companies, including Alphabet, Meta, and ByteDance, spent a combined $50 million on federal lobbying during the first nine months of 2025. That works out to nearly...

Issue One / Fortune

Benchmark

Qwen 3.6-27B Hits 0.9 on GPQA

Alibaba's Qwen team released Qwen 3.6-27B on April 21, an open-source model that scores 0.9 on the GPQA benchmark for graduate-level reasoning. At 27 billion parameters, it delivers near-frontier performance in a size class that can run on...

LLM Stats

Open Source

Google Gemma 4 Goes Fully Open Source

Google released Gemma 4, including a 31B parameter variant that outperforms Meta's Llama 4 Maverick on the AIME math benchmark. The model ships under Apache 2.0, making it fully open source rather than merely open-weight. Mashable confirmed the...

Mashable

Model Release

Claude Opus 4.7 Upgrades Complex Coding

Anthropic shipped Claude Opus 4.7 on April 16, describing it as a notable improvement over Opus 4.6 on advanced software engineering, particularly on the hardest tasks that previously required close human supervision. The model also features...

Anthropic

Trend

April Averages One Model Every 72 Hours

April 2026 has seen over 30 significant model releases or updates in roughly 25 days, sustaining a pace of approximately one new model every 72 hours. SearchCans reports that the surge is creating acute decision fatigue for developers, with...

SearchCans

Enterprise

Microsoft Launches Foundry MAI Human-Centric Models

Microsoft introduced its Foundry MAI models in the company's AI playground, emphasizing a human-centric design philosophy. Details remain limited, but the release aligns with Microsoft's broader push into enterprise AI tooling alongside its...

femaleswitch.app

Hardware

Google TurboQuant Shrinks Models for Edge Deployment

Google released TurboQuant, a memory compression technique designed to make large models viable for smaller, low-latency deployments. The approach targets edge and on-device scenarios where full-precision models are impractical. It reflects a...

SearchCans

Open Source

Open Models Now Match Closed on Key Benchmarks

Multiple open-source and open-weight models released in April, including Qwen 3.6-27B, Kimi K2.6, and Gemma 4 31B, now match or exceed proprietary systems on benchmarks like GPQA (0.9), AIME, and SWE-bench Pro. DeepSeek V3.2 ships under MIT...

AI Frontier Network

The Globe

World Affairs
Build

Claude Managed Agents: Spin Up Production AI Agents Through Anthropic's New Hosted Framework

Anthropic launched Claude Managed Agents in April 2026, letting developers define tasks, tools, and guardrails while Anthropic handles orchestration and hosting. Over 1,000 startups are already using the service to skip the infrastructure work of building agentic systems from scratch. Sign up through the Anthropic dashboard, define your agent's scope and tool access, and deploy without managing your own orchestration layer.

Productivity

Criteo GO: Let Generative AI Optimize Your Ad Campaigns Without a Media Buyer

Criteo GO uses generative AI to handle ad creative generation, audience targeting, and budget allocation for small teams that lack dedicated media buyers. The platform is designed to give startups access to enterprise-grade advertising optimization at a fraction of the cost. Connect your product catalog, set a daily budget, and let the system iterate on creative and placement automatically.

Coding

Manus: Hand Off Entire Multi-Step Projects to an End-to-End AI Agent

Manus operates as a general-purpose AI agent that can browse the web, write and execute code, manage files, and deliver completed work products rather than just suggestions. Describe a project in plain language, such as scraping a dataset, analyzing it, and producing a formatted report, and Manus handles the full pipeline. It is particularly useful for non-engineers who need technical output without writing a single line of code.

Creativity

Canva AI 2.0 Agentic Suite: Delegate Multi-Step Design Workflows to AI Agents

Canva debuted its 'AI 2.0' agentic tools on April 16, adding agents that can handle chained design tasks like resizing assets across formats, applying brand guidelines, and generating copy in one pass. Instead of prompting one step at a time, describe the full deliverable, such as 'create a launch campaign kit with social posts, a banner, and an email header,' and the agent sequences the work. The suite is rolling out inside the existing Canva workspace at no additional cost for Pro subscribers.

Mindset

Project Glasswing: Audit Your Codebase With the Same AI Finding OS-Level Vulnerabilities

Anthropic's Project Glasswing, backed by AWS, Apple, Google, Microsoft, NVIDIA, and others, uses the unreleased Claude Mythos model to find high-severity vulnerabilities in major operating systems and browsers. The defensive security findings will be shared publicly so smaller teams can benefit. Start monitoring Anthropic's Glasswing disclosures now and cross-reference them against your own dependency stack to patch before exploits proliferate.

Productivity

Motion: Let AI Schedule Your Calendar Around Deep Work, Not Just Meetings

Motion uses AI to auto-schedule tasks, meetings, and deadlines across your calendar, dynamically reprioritizing when plans change. The key move is to input every task with an estimated duration and deadline, then let the algorithm protect blocks for focused work instead of manually defending your time. At $19 per month for individuals, it replaces the mental overhead of daily planning with a continuously optimized schedule.

The Arena

Competitive Intel

OpenAI

Cybersecurity push accelerates with trusted access program, safety fellowship, and Axios compromise response

Google DeepMind

Gemini Robotics-ER 1.6 deployed into Boston Dynamics Spot and Orbit AI for industrial autonomy

Anthropic

Federal ban looms as Trump administration targets Anthropic over guardrail refusal; company acknowledges Claude quality regression

Meta AI

Employee surveillance program feeds AI agent training as Muse Spark scores 58% on Humanity's Last Exam

THE DOJO · BUILD TODAY

Deploy managed agents, automate ads, and stress-test your stack against open-source challengers.

  1. Spin up Claude Managed Agents. Anthropic's new hosted framework lets you deploy production AI agents without managing infrastructure. Start with a single retrieval-augmented workflow and measure latency against your current pipeline.
  2. Benchmark your code against GLM 5.1. ZEDAI's MIT-licensed model topped GPT-5.4 on SWE-bench Pro. Pull it locally, run your hardest engineering tasks through it, and compare cost-per-token against your current provider.
  3. Hand off a multi-step project to Manus. Use the new end-to-end AI agent to automate a complete research-to-deliverable workflow. Document where it breaks so you know exactly which steps still need human oversight.
THE BOTTOM LINE

The bottom line: the benchmark moat is gone.

Three Chinese models matched or beat American frontier systems in one week, and one of them did it on domestic chips. For builders, this means open-source alternatives are now genuinely competitive on performance, cost, and licensing. Lock in your tool choices based on real benchmarks, not brand loyalty. The next quarter will reward teams that test aggressively and switch fast.

Want this every morning?

AI analysis, world news, markets, and tools. One briefing, delivered free.

One email per day. No spam. Unsubscribe anytime.