← Back to feed

The AI coding market fractured in 2026 into at least four distinct tiers. Tier 1: autocomplete (Copilot, Tab9). Tier 2: interactive coding agents (Cursor, Devin). Tier 3: something most of these companies are not trying to build. On May 5, Blitzy raised $200 million to own Tier 3.

What the Numbers Say

Blitzy scored 66.5% on SWE-Bench Pro at the time of announcement. SWE-Bench Pro is the long-horizon variant of SWE-Bench, testing multi-step real-world issue resolution across full repositories — not the constrained 500-instance Verified subset. Most frontier models cluster between 58% and 62% on Pro. Blitzy’s architecture, which layers thousands of agents over a knowledge graph of the codebase rather than routing a single model at a task, outpaced individual frontier model baselines at launch.

The $200 million Series A was led by Northzone, with strategic capital from Liberty Mutual Strategic Ventures, Erie Insurance Strategic Ventures, and BAL Ventures — three insurance companies investing in the solution to their own problem. PSG, Battery Ventures, Jump Capital, Morgan Creek Digital, and Defiant also participated. Total funding now exceeds $204 million.

The Architecture

Blitzy does not work like Cursor or GitHub Copilot. The platform begins by reverse-engineering the target codebase — handling repositories up to and exceeding 100 million lines — into a dynamic knowledge graph that maps every system, service, and dependency interaction. That graph is persistent; it updates as the codebase changes.

From there, thousands of specialized agents run in parallel for days or weeks of uninterrupted inference. A single execution cycle calls Claude, GPT, or Gemini more than 100,000 times. The output includes not just code but automated testing and validation.

Pricing reflects the scale: initial evaluations cost up to $250,000. Full projects run $500,000 to $10 million annually. This is not a developer productivity tool. It is priced as infrastructure for decisions a CIO makes once per decade.

The Market

The architectural bet is that enterprise legacy modernization is a categorically different problem from the interactive coding assistance that Cursor and Copilot solve. The proof of concept is in the buyer list: State Street and QAD have both reported 5× engineering velocity gains. Dozens of Global 2000 enterprises across 10 industries are in production deployment.

The specific target is the 220 billion lines of COBOL that still run 95% of US ATM transactions, along with comparable legacy stacks at major insurers, banks, and government contractors. These systems have been deferred for decades because the cost and risk of modernization was too high relative to available engineering capacity. Blitzy’s argument is that the capacity constraint is now solvable with enough agents running long enough, with comprehensive enough codebase understanding that a human engineer would take months to acquire.

Blitzy was founded in 2023 by Brian Elliott and Sid Pardeshi, a former Nvidia master inventor with more than 27 patents. The company is headquartered in Cambridge, Massachusetts, and has more than doubled headcount in the last six months. New capital goes to research expansion and regulated-sector go-to-market: government, financial services, insurance.

What to Watch

The SWE-Bench Pro figure will be challenged. As of announcement, Blitzy’s 66.5% put it ahead of standalone frontier model baselines, but those baselines move quickly. The durable advantage, if it holds, is not in the benchmark score — it is in the knowledge graph architecture that makes 100-million-line codebases tractable. No other platform has demonstrated equivalent capability in that size class in production.

The strategic investor structure — three insurance companies, a private equity sponsor (PSG), and multiple growth funds — suggests the company has already mapped a specific customer pipeline. Insurance modernization alone is a credible $10B+ market opportunity.