IBM Bob

IBM hit general availability with Bob on April 28, 2026 — its bid to be the AI coding agent that buyers in regulated industries actually feel safe deploying. The pitch is the inverse of every consumer-friendly vibe coding tool on this site: instead of “describe an app, ship in a day,” Bob is “we already have 80,000 IBM employees using it internally, here’s how it scales without your CISO writing a memo.”

For the non-technical founders this site is built for, Bob is mostly the wrong tool. But it’s worth understanding what IBM is doing here, because it tells you something about where the AI coding category is heading for buyers who can’t use the Cursor-and-Claude-Code stack.

What Bob actually is

Bob is a multi-agent platform that wraps the full software development lifecycle — planning, coding, testing, deployment, and modernization — with enterprise controls layered on top. The headline architectural choice is dynamic model routing: Bob picks which model to send a task to based on accuracy, performance, and cost, drawing on a mix of frontier models (Anthropic Claude, Mistral open source) plus IBM’s own fine-tuned Granite models for code reasoning, security, and next-edit prediction. You don’t pick the model. Bob does.

The CLI is called BobShell, and it produces what IBM calls “self-documenting agentic processes” — every action the agent takes is traceable from start to finish. That’s a procurement-friendly feature, not a developer-quality-of-life one. It exists so a compliance team can answer “what exactly did this AI do in our codebase?” with a real log file.

The governance stack is where the actual product differentiation sits: prompt normalization (so a sloppy prompt doesn’t leak data), sensitive data scanning (so you don’t accidentally hand customer PII to a frontier model), real-time policy enforcement, and AI red-teaming inside the development workflow rather than bolted on after the fact.

Pricing

Bob uses a token-based pricing model on top of monthly subscriptions:

• Pro — $20/month, 40 Bobcoins
• Ultra — $200/month, 500 Bobcoins

Each Bobcoin is roughly 50 cents of upstream model spend, so Bob’s pricing is effectively pass-through. IBM’s framing is that this gives organizations real visibility into AI spend and lets them tie costs to outcomes rather than to “experimentation.” The honest read: it’s a usage-based model with a more enterprise-flavored wrapper, similar in spirit to how Cursor and Replit Agent 3 charge.

Performance claims

IBM is reporting 20–80% productivity gains across SDLC tasks, 90%+ time savings on repetitive work, and a 45% average productivity gain across the 80,000 internal IBM users it shipped Bob to during its run-up to GA. Take the headline numbers with the usual skepticism — vendor-published productivity stats from internal usage almost never survive contact with independent benchmarking — but the scale of the internal rollout is notable. Bob isn’t a beta dressed up for the press release; it’s a product IBM has eaten its own dog food on.

Who Bob is for

If your organization has a procurement department, an internal AI usage policy, audit requirements, and a CISO who has already had at least one uncomfortable conversation about Cursor or Claude Code, Bob is on the shortlist. The governance features aren’t decorative — they’re the actual reason to buy this over a consumer-tier AI coding tool. Pass-through pricing with usage visibility is a legitimately differentiated answer to “we don’t know how much we’re spending on AI.”

If you’re a solo founder, a two-person startup, or a PM trying to validate an idea, this is not your tool. The pricing isn’t punishing, but the entire product is built around enterprise concerns you don’t have. Use Cursor or Claude Code instead.

Bottom line

Bob is IBM’s serious entrant in the AI coding agent category, and unlike a lot of “enterprise AI” launches it’s an actual product, not vaporware. The dynamic model routing is a real architectural bet, the governance stack solves a real buyer problem, and the internal-usage scale is genuinely impressive for a GA launch. None of that makes it the right starting point for the audience this site serves — but if you ever transition into a role where a CIO is asking which AI coding tool to standardize on across 5,000 engineers, Bob deserves a real evaluation.

Was this helpful?

YesCould be better