Hi, I'm Kenneth. Welcome to this week's edition of Tech Venture Navigator.
April has been relentless. Anthropic built an AI model so capable at finding security flaws that they decided not to release it publicly. Instead, they gave it to Apple, AWS, Google, Microsoft, Nvidia, and a handful of others to patch their own systems before anyone else gets models this powerful. OpenAI shipped GPT-5.5 six weeks after GPT-5.4, positioning it as their first true agent runtime. DeepSeek dropped V4 with 1.6 trillion parameters, open-source, trained on Chinese chips, matching Claude on coding benchmarks at one-seventh the price. And YC published 15 categories of companies they want to fund, the most ambitious RFS in their history.
This edition covers the stories that define where we are and where capital is flowing. Let's get into it.
Brought to you by Kondo.
I. Claude Mythos and Project Glasswing: The Model Too Dangerous to Release
Claude Mythos Preview is Anthropic's newest frontier model. It's a general-purpose model with strong coding and reasoning capabilities. But what made headlines is what it did when pointed at cybersecurity.
In internal testing, Mythos identified thousands of zero-day vulnerabilities across every major operating system and every major web browser. It found a 17-year-old remote code execution vulnerability in FreeBSD that gives anyone root access from the internet. It found a 27-year-old bug in OpenBSD. It chained together four separate vulnerabilities to write a browser exploit that escaped both renderer and OS sandboxes. All fully autonomous. No human involved after the initial prompt.
Anthropic's response: don't release it publicly. Instead, they launched Project Glasswing, giving restricted access to 12 launch partners (AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, Nvidia, Palo Alto Networks) plus 40 additional critical infrastructure organisations. Anthropic committed $100M in usage credits and $4M in donations to open-source security foundations.
The UK AI Security Institute tested Mythos independently and confirmed it was the first AI model to complete their full network takeover simulation. They noted limitations in real-world defended environments but called the capability "a significant step change."
Nicholas Carlini from Anthropic's red team said he found more bugs in a few weeks with Mythos than in the rest of his career combined.
Sources: Anthropic | Anthropic Red Team Blog |
✈️ NAVIGATOR’S EDGE: This is Anthropic pivoting from the Pentagon controversy to cybersecurity leadership in a single move. Two months ago, they were banned from federal agencies. Now they're the company that Apple, Google, Microsoft, and the US government's own infrastructure partners are relying on to find vulnerabilities before attackers do.
For founders in cybersecurity: The entire defensive security market just got a new baseline. If Mythos-class models can find zero-days autonomously, every security product that relies on signature-based detection is structurally obsolete. Build for a world where AI finds vulnerabilities faster than humans can patch them.
For investors: Anthropic is building a moat through trust and responsible deployment, not just capability. The Project Glasswing partner list reads like a who's who of global infrastructure. That's enterprise distribution at scale, earned through restraint rather than speed-to-market.
II. GPT-5.5: OpenAI Stops Selling a Chat Model and Starts Selling an Agent
Six weeks after GPT-5.4. That's the cadence now.
GPT-5.5 (codename "Spud") launched April 23. OpenAI's president Greg Brockman called it "the first coding model with serious conceptual clarity." The pitch has fundamentally shifted from "it answers questions better" to "it finishes tasks without you managing every step."
The key capability improvements: you can give GPT-5.5 a messy, multi-part task and it will plan, use tools, check its work, navigate ambiguity, and keep going until done. It excels at writing and debugging code, operating software, researching online, and creating documents and spreadsheets.
Benchmarks: Terminal-Bench 2.0: 82.7%. FrontierMath (Tiers 1-3): 51.7%. Significant gains on agentic coding and computer use over GPT-5.4.
Pricing: Standard at $5/$30 per million input/output tokens. Pro at $30/$180. More expensive than GPT-5.4 but more token-efficient. Nvidia gave 10,000+ staff early access through Codex across engineering, legal, finance, and ops. Not just engineers.
Tom's Guide tested GPT-5.5 against Claude Opus 4.7. GPT-5.5 lost in all seven categories tested. Praised for speed. Criticised for hallucination over admission of uncertainty.
Sources: OpenAI | CNBC | TechCrunch | Wikipedia
✈️ NAVIGATOR’S EDGE: The positioning shift matters more than the benchmarks. OpenAI is no longer selling a chat completion API. They're selling an agent runtime. When a company leads with "it finishes tasks" instead of "it scores higher on benchmarks," they're targeting a different buyer. Enterprise procurement teams buy outcomes, not scores.
For founders: If you were chaining three or four GPT-5.4 calls to complete a multi-step workflow, you can likely collapse that to one or two with GPT-5.5. The reliability improvement on multi-step completion is worth the premium. But test against Claude Opus 4.7 before committing.
III. DeepSeek V4: 1.6 Trillion Parameters, Open Source, Trained on Chinese Chips
One year after R1 crashed Nvidia's stock, DeepSeek is back with something bigger.
V4-Pro: 1.6 trillion total parameters (49B active per token). V4-Flash: 284B total (13B active). Both support 1M token context by default. Both MIT licensed. Both open-source on Hugging Face.
The efficiency numbers are staggering: At 1M token context, V4-Pro uses only 27% of the inference compute and 10% of the KV cache memory compared to DeepSeek V3.2. This makes million-token context practically deployable for the first time at this scale.
Benchmarks: SWE-bench Verified: 80.6% (Claude Opus 4.6: 80.8%). Codeforces rating: 3,206 (highest ever for any model at release). Terminal-Bench: 67.9% (Claude: 65.4%). Near-frontier performance at a fraction of the price.
The geopolitical signal: V4 was trained entirely on domestic Chinese hardware, specifically Huawei Ascend chips and Cambricon accelerators. Not Nvidia GPUs. Huawei confirmed its latest AI cluster can support V4. This is proof that frontier-class models can now be trained outside the Nvidia ecosystem.
Pricing: V4-Flash at $0.14 per million input tokens. That's a 7x gap to Claude on coding benchmarks that are within 0.2 percentage points of each other.
✈️ NAVIGATOR’S EDGE: DeepSeek V4 changes the cost calculus for every AI startup running inference. If your product sits on top of an LLM, the price difference between Claude and DeepSeek on equivalent coding tasks is now 7x. That's not a rounding error. That's a different business model.
For investors: The "trained on Chinese chips" signal is the one to watch. If Huawei Ascend can support frontier training, the US chip export controls are less effective than assumed. The geopolitical implications for Nvidia's moat are significant.
IV. YC Summer 2026 RFS: 15 Categories, Three Themes, One Message
YC just published its most ambitious Request for Startups ever. 15 categories. Deadline: May 4. This isn't a vague "we like AI" blog post. It's a partner-attributed wishlist of exactly what they want to fund.
Three themes running through the entire list:
1. AI is the operating system, not a feature. Company Brain, AI Operating System for Companies, Software for Agents, SaaS Challengers, Dynamic Software Interfaces, AI-Native Service Companies. All variations on one belief: the software stack is being rebuilt AI-native from the ground up.
2. Silicon is the new frontier. Inference chips for agent workflows, electronics in space, semiconductor supply chain. Hardware is back. The AI stack runs on chips, chips run on a fragile supply chain managed with spreadsheets. YC sees opportunity across all three layers.
3. The physical world is open for business. Low-pesticide agriculture, counter-swarm defence, industrial capabilities in space, hardware supply chain. The most "atoms" YC has ever been.
The categories founders will skip but shouldn't:
Software for Agents. Everyone is building agents. Almost nobody is building what agents need: machine-readable interfaces, APIs instead of buttons, documentation agents can parse. Aaron Epstein's framing: "Make Something Agents Want." This is the AWS of the agent economy.
AI-Native Service Companies. Gustaf Alstromer upgraded this from the Spring RFS. The thesis: services spend globally is multiples of software spend. If AI can deliver a service at software margins, the market is enormous. Insurance brokerage, accounting, tax, audit, compliance, healthcare admin. The pitch is a services contract, not a SaaS subscription.
Counter-Swarm Defence. A swarm of cheap drones recently took out an AWS data centre. A Patriot missile costs $3M. An FPV drone costs $500. Tyler Bosmeny's thesis: the winning companies will look more like Cloudflare than Raytheon.
Chris Tottman's deep dissection of the RFS is worth reading in full. His key insight: the shift from Spring ("aim AI at unglamorous industries") to Summer ("we are at the beginning of a new hardware era") reflects something structural. YC is making layer-zero bets now, not just layer-two applications.
Sources: Y Combinator | VC Corner | Founders Corner | VC Cafe
V. The Model Race and Funding Radar
Seven days of seismic moves across models and capital:
OpenAI's Pentagon Deal
Hours after Anthropic's ban, OpenAI announced a deal to deploy models on the Pentagon's classified networks. Altman's internal memo stated OpenAI shares Anthropic's red lines on surveillance and autonomous weapons. But critics note the contract language references "all lawful purposes," leaving significant room for interpretation. (Axios)
Altman also called Anthropic's supply chain designation "a very bad decision" and "an extremely scary precedent." 11 of his own employees signed an open letter opposing it.
Week of Feb 24 - March 2
Company | Amount | Stage | Sector | Key Investors |
|---|---|---|---|---|
OpenAI | $110B | Growth | Frontier AI | Amazon ($50B), Nvidia ($30B), SoftBank ($30B) |
Wayve | $1.2B | Series D | Autonomous Driving | Uber, Microsoft, Nvidia, Mercedes, Nissan |
MatX | $500M | Series B | AI Training Chips | - |
SambaNova | $350M | Series E | AI Inference Chips | Vista Equity, Intel Capital |
Axelera AI | $250M | Series B | Edge AI Chips | BlackRock, Samsung Catalyst |
Basis | $100M | Series B | AI Accounting | Accel |
Profound | $96M | Series C | AI Marketing | - |
Rowspace | $50M | - | AI Financial Intelligence | - |
Harper | $47M | - | AI Insurance | - |
Sources: TechStartups | Analytics Insight | AI Funding Tracker
✈️ NAVIGATOR’S EDGE: The pattern this week: AI chip alternatives are getting real funding (MatX, SambaNova, Axelera pulled in $1.1B combined in one week). Autonomous driving is back (Wayve's $1.2B from actual car manufacturers, not just VCs). And vertical AI for regulated industries (Basis for accounting, Harper for insurance, Rowspace for finance) continues to attract serious capital.
February 2026 closes as the most consequential month in AI investment history: an estimated $195B+ in tracked AI-related capital, driven by OpenAI's $110B, Anthropic's $30B, Waymo's $16B, and dozens of $100M+ rounds.
YC Summer 2026 deadline (May 4) and which categories attract the most applications
Project Glasswing disclosures as partners begin publishing vulnerability patches
DeepSeek V4 adoption and whether the 7x price gap triggers a pricing response from Anthropic or OpenAI
GPT-5.5 API rollout and the "different safeguards" OpenAI flagged
Prediction market volume trajectory as Polymarket and Kalshi integrate with more retail platforms.
Miles Clements is a Partner @ Accel where he helps to lead their growth fund. At Accel, Miles has led or invested in Atlassian, Cursor, Linear, and more.
Podcast: Apple |
VC Research Corner
GTM Effectiveness Collapse (Chris Tottman)
B2B GTM effectiveness fell from 78% in 2018 to 47% in 2025. More than half of every sales and marketing dollar is structural waste. 40-60% of deals end in no decision. The core insight: your ACV is your GTM compass. If your motion costs more than your price point can sustain, no amount of hiring fixes it. Five GTM motions exist on a spectrum from PLG to named accounts. Most founders are running the wrong one. (Founders Corner)
The Big Number: 7x
The price gap between DeepSeek V4 and Claude Opus on coding benchmarks that are within 0.2 percentage points of each other. V4-Pro at $3.48 per million output tokens. Claude at $25. Same performance, radically different economics. That gap is going to force pricing decisions across the entire foundation model market within the next quarter.
✈️ NAVIGATOR'S EDGE: If your pitch deck still positions your product as "AI-enhanced SaaS," delete that slide. The market just wiped $830 billion from companies with that exact positioning. Reframe around autonomous outcomes, proprietary data loops, and outcome-based pricing. The question VCs are asking in every meeting this week: "Could a Cowork plugin replicate what you do?" You need a bulletproof answer.
Claude Mythos Preview - Not publicly available, but Project Glasswing partners are using it for defensive security. The capability benchmark for where AI cybersecurity is heading.
GPT-5.5 - OpenAI's agent runtime. Try it on messy, multi-step tasks where you'd normally chain multiple calls.
DeepSeek V4 - Open-source, MIT licensed, 1M context, fraction of frontier pricing. Test V4-Flash for high-volume production workloads.
Lovable - Full-stack app generation via chat. The gold standard for non-technical founders shipping MVPs.
Bolt.new - In-browser development using WebContainers. Instant full-stack environment generation.
Averi - GEO (Generative Engine Optimisation). Optimises your brand to be cited by Perplexity and ChatGPT.
Lemlist — The outreach tool we recommend for breaking through the noise with personalised fundraising outreach.
We sit at the centre of 560k+ founders, operators, and LPs. Our goal is to eliminate the "random walk" of fundraising by using our data to make the one warm connection that actually matters.
For Founders: If you are raising a Seed or Series A and your roadmap is built for the agentic era, we want to see it. We don't just invest; we plug you into the distribution loops of our entire community. 👉 Startup Intake Form: Apply here
For Investors: Stop digging through generic deal flow. Tell us exactly what "high-signal" looks like for your mandate, and we'll route the outliers directly to your inbox. 👉 Investor Intake Form: Apply here
Fundraising in 2026 is a game of conviction and speed. Let's stop wasting time on "maybes" and start building rounds that move the needle.
Pass this on to a founder who’ll use it.
Thanks for reading this edition. If you got value from it, share it with a founder, teammate, or friend who’d use it too it keeps this community sharp and growing. Not for you? You can always update your preferences anytime.
Want more practical ideas on how to raise smarter and build better?
Join the crew below and follow me on LinkedIn to never miss an update.
If you’d like tailored help you can book time with me on hubble here.
Stay going,
— Kenneth Kelly
Founder @Tech Venture Navigator
What do you think about Tech Venture Navigator?
💡 Love it | 🔥 Useful | 👍 Solid | 🤷♂️ Okay | 🚫 Skip it or Leave feedback here.
Got ideas, topics, or frameworks you’d love to see covered? Just hit reply or drop me a message, we read every note.