AI coding tools: who captures the value?
A commercial map of subscriptions, consumption and workflow control in AI-assisted software development.
Public-source desk research, checked 16 October 2026. Company disclosures describe vendor claims and published terms. Commercial conclusions are Offaxis interpretation. Numerical scenarios are explicitly synthetic. We have not independently tested products or audited private company financials.
Investment of attention: follow the budget owner
Our thesis: a durable coding-tool business needs a repeatable claim on an engineering budget. A convincing code demonstration can win a trial; renewal depends on a team completing useful work with acceptable review effort. The commercial unit therefore moves from an individual prompt toward a maintained software change. This is Offaxis analysis, not a measured productivity result.
The market has at least three competing routes to that budget: a model provider selling intelligence, a development platform distributing assistance inside an existing workflow, and an application vendor packaging a particular way of working. These are analytical categories; a company can occupy more than one. Owning a model is one possible advantage. Owning the place where work is assigned, reviewed and accepted is another.
What public evidence establishes
GitHub’s pricing page describes AI credits consumed by multiple Copilot features and organizational controls. It also identifies license and policy management among differences between individual and organizational offerings. This establishes a published packaging strategy, not Copilot’s standalone profit margin. [1]
Anthropic’s business-plan announcement describes premium seats, additional usage, spending limits and usage analytics. These are mechanisms for selling both access and consumption while giving administrators control. The announcement’s customer testimonials are company-published evidence, not independent productivity studies. [2]
Where the money can go
| Business position | Revenue mechanism | Commercial advantage to test | Failure mode |
|---|---|---|---|
| Model provider | Metered inference and business plans | Task quality and inference efficiency | Usage grows faster than contribution |
| Development platform | Seats plus consumption | Distribution into existing engineering work | Bundled access goes unused |
| Workflow application | Subscriptions or task allowances | Fewer handoffs for a defined customer | Users switch when a platform bundles the feature |
| Integration specialist | Setup, maintenance and support | Customer-specific operational knowledge | Every deployment becomes custom consulting |
This map is an editorial framework. It does not assign market shares or claim that every vendor follows one revenue model.
The renewal equation
For a buyer, total monthly cost includes seats, additional consumption, integration, review and rework. For a vendor, contribution includes revenue less inference, execution infrastructure, support and other directly attributable service costs. Customer value and vendor margin are related but distinct: a useful tool can still be expensive to supply.
Illustrative sensitivity, not vendor data: suppose a team spends $1,000 on access and usage in a month, then incurs $2,000 of incremental review and rework. Across 200 accepted changes, the incremental cost is $15 per accepted change. At only 100 accepted changes, it becomes $30. These figures do not demonstrate savings. A valid comparison needs the cost and quality of equivalent work without the tool, including review that would have happened anyway.
Competition and defensibility
Our hypothesis is that generic feature differentiation is vulnerable to bundling. A stronger position would combine reliable task completion, integration with the customer’s repository and an administrative purchase that is easy to renew. Integration can create switching friction, but friction alone does not establish a lasting advantage. Portable configuration and better competing tools can reduce it.
What would change our view?
We would look for cohort renewals, paid expansion after trials, accepted-change cost, recurring defect rates and contribution after serving heavy users. Published user counts without a paid definition would not answer those questions. A buyer can run a bounded trial on comparable tasks and record reviewer time, failure recovery and consumption together. The key falsifier is straightforward: if output increases while accepted, maintainable work does not, the commercial value case weakens.
Sources and evidence
Stay in the loop
Sign up for Offaxis news and updates. Free email signup; paid membership is separate.