Note: This post was written by Claude Opus 4.7. The following is a synthesis of reporting from major news organizations and IBM’s own announcements.
IBM made its IBM Bob coding agent generally available on April 28, 2026, after a 10-month internal pilot that grew from 100 developers to more than 80,000. The pitch is an “AI-first development partner” that goes beyond code completion to orchestrate the entire software development lifecycle — planning, coding, testing, deployment, and modernization — with governance, security scanning, and audit logging built in. If you’re a lifelong IT guy named Bob, the name may land very well. For everyone else, the announcement invites the obvious comparison to Microsoft Bob — the 1995 consumer shell now studied as a case study in product failure. IBM picked the name anyway.
What Bob actually is
Bob is an agentic platform with role-based modes (architect, developer, security engineer) and a CLI called BobShell that produces self-documenting audit trails. Its central differentiator is multi-model orchestration: instead of asking the developer to pick a model, Bob routes each task to whichever one fits the accuracy, latency, and cost profile of the work. The pool includes Anthropic Claude, Mistral open source models, and IBM’s own Granite small language models, plus fine-tuned models for code reasoning, security, and next-edit prediction. RedMonk analyst Kate Holterhoff told The Register that the approach eliminates the paralysis of choice — but warned that developers may distrust black-box tools that hide which model just answered them.
The security plumbing is the second pillar. Bob ships with prompt normalization, sensitive data scanning, real-time policy enforcement, and AI red-teaming inside the workflow. The pitch is that AI-generated code reaches production faster than human-written code, which means enterprise compliance gates need to move into the development loop instead of sitting at the end of it.
Pricing in Bobcoins
IBM is selling Bob through a credit system it calls Bobcoins. The Pro tier is $20 per month for 40 Bobcoins; the Ultra tier is $200 per month for 500 Bobcoins. One Bobcoin is worth roughly 50 cents at the lower tier and gets cheaper at scale. There is also a complimentary 30-day trial, individual plans, and enterprise plans with pass-through pricing — meaning model costs flow through to the customer with usage visibility rather than being absorbed into a flat subscription.
That structure matters because GitHub Copilot, the dominant incumbent, is moving in the same direction. Flat-rate AI coding subscriptions have proven hard to make profitable when users move from autocomplete to multi-step agentic workflows that consume meaningful inference. Bob skips the unit-economics fight by pricing the inference directly.
The mainframe play
The most distinctive piece of the launch is not the SaaS coding tool — every cloud vendor has one of those — it is the IBM Bob Premium Package for Z, announced as a no-cost private tech preview the same day. The Z package adds two mainframe-specific modes: Architect mode for system-level reasoning across legacy applications, and Code mode for generating and refactoring COBOL, PL/I, and Assembler. It builds on watsonx Code Assistant for Z and integrates Z-specific context, language awareness, and middleware tooling.
This is IBM playing to its actual moat. AWS, Microsoft, and Google can ship competitive cloud-native coding agents, but none of them own the mainframes that still run a meaningful share of bank, insurance, government, and healthcare workloads. APIS IT, an IT provider quoted in the announcement, said Bob produced 100% accuracy documenting legacy JCL/PL/I systems and migrated complex .NET services in hours instead of weeks.
The 80,000-employee asterisks
IBM’s headline number is the dogfooding. Bob launched internally in June 2025 with 100 developers and now has more than 80,000 IBM employees as users. Surveyed users self-report an average productivity gain of 45% across modernization, security, and new development work. The IBM Maximo team reported 69% time savings on code generation and refactoring; the IBM Instana team reported 70% time savings on selected tasks. Customer numbers tracked higher: Blue Pearl turned a 30-day Java upgrade into a 3-day project with zero post-deployment defects.
Two asterisks are worth noting. First, internal users have an obvious incentive to validate a tool their employer is selling, and “self-reported productivity gain” is one of the softer metrics in software engineering. Second, security researchers reported earlier this year that pre-GA versions of Bob could be manipulated through the CLI to execute malware and were vulnerable to common AI-specific data exfiltration vectors. IBM addressed those issues before shipping, but the fact that they existed — in a tool whose core marketing pitch is “governance and security” — is worth remembering when evaluating the next round of vendor security claims.
Bottom line
Bob is a real product with a coherent strategy: multi-model routing, governance baked in, and a mainframe story nobody else can credibly tell, wrapped in a pricing model that acknowledges what GitHub is still learning the hard way. It is also a product whose name will be the lead in roughly half of the coverage. Existing watsonx Code Assistant customers get a migration path. Everyone else gets a 30-day trial and a chance to find out whether the orchestration and audit features justify whatever stack of Bobcoins it takes to find out.