He Built the World's #1 Open-Source Coding Agent

Open-source “coding agent” platform: what Open Code is

Open Code is positioned as an open-source alternative to Cloud/hosted coding agent platforms. It’s designed to work with any LLM/model and exposes a model picker, making it easier to adopt new open-source models quickly.

Two-part product architecture

Open Code is described as having:

• A UI/application that users interact with.
• A behind-the-scenes agent loop (“server”) that runs the coding agent by calling the selected LLM.

---

Scale and usage metrics (product adoption signals)

Key adoption and throughput metrics discussed include:

• Weekly active users: 4.6 million (discussion suggests monthly active users may have been even higher)

• Monthly active users: ~13 million by end of June, about a 20x increase since the beginning of the year

• Inference/token throughput: ~7 trillion tokens per day recently (compared to OpenRouter at ~6T/day)

Revenue signals (inference + subscription)

• Revenue via inference (pay-per-token, extrapolated):
  • ~31–33M annualized (from June data)
  • ~38–40M annualized (from last-week data)
• Subscription product adoption:
  • ~160,000 monthly subscribers (subscription launched early March)
  • ~18M annualized revenue attributed to subscription

---

Major growth inflection points

Anthropic/Claude Code “clampdown” episode (inferred)

• Users could use Claude Code subscriptions through Open Code via system prompt handling.
• Anthropic reportedly rejected requests when the system prompt mentioned “open code.”
• Result: press and users treated the products as comparable, creating curiosity and trial.

Open-source model “gap shrinking” timeline

• Early on, Open Code mainly helped users access cloud coding subscriptions.
• Adoption surged as open-weight models improved (a “crop” around Aug/Sep), becoming good enough for real work.

Model equivalence event

Around February of this year, open models reportedly became competitive enough that:

• Gemini 5.2/GLM usage rose sharply
• During a specific 4-week span, users preferred Gemini/GLM over some Anthropic models
• This shift is cited as part of why Open Code launched a subscription product

---

Model usage analytics Open Code publishes

Open Code publishes usage data at opencode.ai/data. The data is based on usage on their subscription tier (paying to use multiple open-source models).

What they publish

• Token volume per day per model
• Unique users per model
• Market-share graphs (token share across model labs)

Key insight claims from the data

• DeepSeek Flash is heavily used by token volume.
• Despite hype, the data suggested GLM wasn’t universally “taking over” DeepSeek:
  • GLM can dip/spike depending on timing
  • DeepSeek usage rebounds after GLM’s initial release
• From a unique-user perspective, top user counts included:
  • Flash (~38k), DeepSeek Pro (~31k), and GLM 5.2 (close to 30k)

Why Flash may be popular (theorized)

• Lower cost allows users to “token budget max”:
  • switch to cheaper models near limits to finish remaining work

---

Geo-based adoption and affordability positioning

Open Code emphasizes adoption in developing markets, with:

• China as #1 (~17%)

Other referenced regions include:

• Indonesia (~4%), Brazil (~5%), Vietnam, etc.

Affordability rationale

A stated driver is that frontier/Cloud Code subscriptions are too expensive (example: $200/month), so:

• open models + Open Code become a more accessible alternative.

---

Enterprise/security “pull” and product-market-fit signal

Open Code claims enterprise traction measured via security questionnaire / security agreement requests.

Common enterprise concerns/questions include:

• Official internal adoption for developers already using it
• Extending access to non-technical users
• Controlling token spend / limiting access to frontier models
• Sometimes requesting visibility/auditing into what teams do (and whether Open Code plans to support this)

Example enterprise workflow: Ramp Slack bot

• Ramp built a Slack bot that runs Open Code behind the scenes
• The subtitles clarify it uses Open Code’s separable agent loop (“server”) independently from the UI.

---

Unit economics: token-based CAC vs ad-based CAC

Growth is framed as token-based experience rather than ad-driven growth:

• Users get the “magic moment” from trying a coding agent via a free tier and/or pay-per-token.

Profitability mechanics discussed

• Pay-per-token can improve margins due to volume discounts from aggregation.
• Early free tier is cost-bearing, but later “whales” (high token spenders/large orgs) drive margin.
• It’s also compared to how major labs subsidize usage to drive adoption.

---

Strategic positioning: marketplace vs lock-in

Open Code positions itself like a marketplace:

• Users can choose models with different cost/quality tradeoffs.
• They argue model competition benefits consumers.
• They expect open-source model labs to specialize into niche roles as models commoditize into utilities.

---

GPU/inference operations insights (high-level)

Open Code reportedly rents GPUs and uses:

• GPU providers for raw inference
• Model labs directly

An operational advantage mentioned:

• Because usage is global, GPU utilization peaks/troughs are less extreme across time zones
• This supports a more stable 24-hour cycle, improving unit economics

---

Product/design choices and branding

Naming/open strategy

• “Open Code” (and earlier “OpenNext”) was a deliberate branding choice to be an open alternative to dominant players—making it harder to be displaced.
• Early support for 70+ models/providers required building a dataset:
  • models.dev (an open database of models/providers)

Developer UX focus

• Open Code emphasizes a modern terminal UI feel for core developers
• It’s described as appealing to “terminal UI connoisseurs”

Origin example (developer-tool culture)

• The team has experience building terminal UIs, including a humorous storefront concept like “buy coffee over SSH”, reflecting strong taste in developer tooling.

---

Main speakers / sources

• Jay V — Founder and CEO of Open Code (primary interviewee)
• Gary Tan — mentioned as the host/regular host (traveling, not speaking in the provided subtitles)
• Host / podcast interviewer (“Lakum” episode) — conducts the interview with Jay (name not explicitly stated in subtitles)

References mentioned in the discussion:

• Dylan Patel (profitability/chasm context)
• Arjun (analytics/data referenced)
• Tibo (lead engineer referenced via a tweet about Codex subscribers)
• Ramp (company example using Open Code)
• Anthropic / Claude Code
• Google / Gemini
• OpenRouter
• YC (Y Combinator) (background/context)