The most interesting thing about IBM’s new coding agent isn’t the code it writes. It’s the bill.

Hear me out.

IBM launched Bob in April 2026 as its answer to Claude Code and OpenAI’s Codex. Here’s the twist: Claude is actually one of the models running inside Bob. IBM didn’t try to out-build Anthropic’s model. It built everything around the models instead — the routing, the cost tracking, the audit trail, the mainframe know-how. And that’s exactly where Bob pulls ahead.

Let’s be fair upfront. Claude Code is still the sharper pure coder, and most developers will tell you so. But there are real jobs Bob does that Claude Code doesn’t do, or doesn’t do as well. Here are ten.

1. It picks the right model for each job

Claude Code runs Anthropic models only. You pick one, and every task — brilliant or boring — gets that model at that model’s price.

Bob works differently. It breaks a task into pieces and routes each piece to whichever model fits: Claude for the hard reasoning, Mistral or IBM’s small Granite models for the routine stuff. You wouldn’t assign your most expensive engineer to write boilerplate test cases. Bob applies the same logic to AI.

IBM claims this routing cuts AI compute spend by roughly 40%. Even if the real number lands at half that, it changes your monthly bill. Claude Code has nothing like it.

2. It can tell you what every dollar bought

Ask a team running Claude Code what last month’s AI spend actually produced, and you’ll often get a shrug. Costs are spiky — a multi-agent session can burn several times the tokens of a normal one — and plenty of teams end up bolting on third-party trackers just to see where the money went.

Bob ships that answer in the box. Its Bobalytics dashboard tracks consumption by project, shows which model handled what, and lets you set spending policies centrally. Work is priced in fixed credits called Bobcoins — about 50 cents each — so finance can actually forecast.

“What did this feature cost us?” is a question your CFO understands. Bob can answer it. Claude Code mostly can’t.

3. It speaks mainframe

If your business runs on IBM Z — and most big banks, insurers, and airlines still do — this one matters.

Bob’s Premium Package for Z builds on IBM’s watsonx Code Assistant for Z. It has an Architect mode that maps out how a mainframe application is structured, what depends on what, and what breaks if you touch something. Then it generates and refactors code with full awareness of that environment.

Claude Code can read COBOL and analyze it well. But it has no mainframe toolchain, no z/OS integration, and no decades of IBM institutional knowledge behind it. On the mainframe, Bob is playing at home.

4. It covers IBM i and its forgotten languages

Same story, different platform. Millions of businesses still run on IBM i systems, written in languages like RPG, CL, and DDS that most AI tools barely recognize. There simply isn’t much RPG on the internet for models to learn from.

IBM built a dedicated Bob package for exactly this: RPG, COBOL, CL, SQL, DDS, and the workflows around them. IBM i’s own CTO confirmed Bob replaces the previous watsonx assistant for the platform.

If your codebase includes these languages, Claude Code isn’t really an option. Bob is purpose-built for it.

5. Legacy modernization is a product, not a project

Claude Code can absolutely modernize old code — if you know how to drive it. You design the approach, break up the work, and steer every step yourself.

Bob packages the whole thing: repo-wide refactors, dependency upgrades, framework migrations, automatic test regeneration, and keeping your CI pipeline aligned — as guided workflows with premium packages for Java, IBM i, and Z.

The early numbers are hard to ignore. IBM says one customer, Blue Pearl, compressed a typical 30-day Java upgrade into about 3 days, saving over 160 engineering hours. Another, APIS IT, documented a 20-year-old government system 10x faster and finished migrations in hours instead of weeks. Those are IBM’s own case studies, so apply a discount — but even the discounted version is a lot of money saved.

6. It keeps receipts

Here’s a problem most companies haven’t faced yet but will: an auditor asks who changed a piece of production code, and the honest answer is “an AI did it, and we’re not sure exactly what it looked at.”

Bob’s command line, BobShell, documents every agent action in real time — what was read, what was decided, what was changed, start to finish. When compliance comes asking, there’s a paper trail.

Claude Code has logs and enterprise plans include audit features, but traceability isn’t its core design goal. For Bob, it is. In a regulated industry, that difference decides which tool legal will approve.

7. Security checks happen while the code is written

Most teams scan for vulnerabilities after code is written — in review, or worse, in production. Bob moves that earlier. It scans as code is being authored, flags sensitive data, enforces your company’s policies in real time, and even runs its own AI red-teaming inside the workflow.

Claude Code has a security review command and lets you wire up your own checks. But that’s DIY. Bob ships the security layer pre-assembled, which matters when the developer using the AI is a junior engineer who doesn’t know what to check for.

Catching a flaw at typing time costs minutes. Catching it in production costs a weekend, a war room, and sometimes a headline.

8. Different roles get different guardrails

Bob doesn’t hand every user the same keys. It has role-based modes — Architect, Developer, Security Engineer — and each mode limits what the AI agents can actually do. An architect-mode agent can plan but not push code. That’s how large organizations already think about access, and Bob mirrors it.

Claude Code’s permissions work per-tool and per-command, which suits individual developers fine. But it wasn’t designed around organizational roles. When you’re rolling AI out to 500 engineers with different seniority and clearance levels, role-based control is the model IT departments trust.

9. It’s a complete environment, not an add-on

Claude Code lives in your terminal and plugs into your editor through extensions. That’s great if you’re a developer with an existing setup you love.

Bob is the whole environment — a full IDE built on VS Code, with Ask, Plan, and Agent modes built into the interface, custom modes you can define, plus the BobShell command line for automation. One install, one login, everything included.

For an enterprise standardizing hundreds of developers on one approved tool, “one thing to deploy and govern” beats “a CLI plus extensions plus configuration” every time. Fewer moving parts, fewer support tickets.

10. It arrives already wired into the IBM stack

If your company already runs IBM — watsonx, OpenShift, Z infrastructure, HashiCorp tooling — Bob shows up pre-connected. It can query watsonx.data mid-task, hooks into watsonx Orchestrate, and gives existing watsonx Code Assistant customers a supported migration path. And behind it sits IBM’s consulting and enterprise support machine.

Claude Code can connect to almost anything through MCP, but you build those connections yourself. With Bob, IBM has done the wiring — and there’s one vendor to call when something breaks. Procurement teams love that sentence more than any benchmark.

The honest bottom line

None of this makes Bob the better coder. Reviewers consistently rank Claude Code’s raw coding ability ahead. If you’re one developer who wants the smartest agent in your terminal, Claude Code is still the pick.

But Bob isn’t competing on smartest. It’s competing on cheapest to run, easiest to govern, and safest to put in front of an auditor — while quietly using Claude’s intelligence under the hood.

Here’s the twist, though. Picture the developer who gets the most out of Claude Code — smart enough to build their own guardrails, wire in their own security checks, and watch their own token spend. That developer has basically assembled a coding agent of their own around the model. There’s nothing left to sell them.

But no company buys software for that one developer. An enterprise shopping for an SDLC AI agent is buying for all five hundred engineers — the juniors, the contractors, the team stuck with a twenty-year-old codebase. For them, the roles, the security, the governance, and the cost control aren’t packaging around the product. They are the product. Strip them out, and there’s no purchase left to make. Keep them in, and IBM Bob makes complete sense.

And here’s my bet on how this market shakes out: the winners won’t be the companies with the smartest AI. They’ll be the ones who can prove what every AI dollar actually bought.

Right now, Bob answers that question better than anyone.