Skip to content
← Back to blog
Research Papers·August 18, 2026·6 min read

Claude's API is doing more coding work automatically. Its chat app is doing less.

Anthropic's March 2026 index: coding tasks rose 14% in its API since August, fell 18% in Claude.ai. The split shows where AI work is really going.

Anthropic just told on itself. In the March 2026 edition of its Economic Index, the company that makes Claude reports that coding tasks are quietly leaving its own chat product for its API, and the reason isn't that people stopped using AI to write code. It's that the same task changed shape: what used to be a back-and-forth conversation in Claude.ai is increasingly a workflow running unattended through the API. If you're deciding this year how much of your roadmap to hand to an agent, that shift, not another "AI writes X% of our code" headline, is the number worth reading closely.

What the Economic Index actually tracks

Anthropic's Economic Index samples about a million conversations across Claude.ai and Anthropic's own first-party API, classifies each one against the Department of Labor's O*NET occupational task list, and separately tags whether the interaction looks like automation (the model completes a task with little ongoing input) or augmentation (the human stays in the loop through feedback, iteration or validation). This edition, subtitled "Learning curves," was published March 24, 2026, drawing on a sample pulled February 5 to 12, 2026, compared against the same categories from Anthropic's report in August 2025. Anthropic also publishes the underlying dataset on Hugging Face, so the classification behind these numbers is checkable rather than taken on faith, which is more than you can say for most vendor usage claims.

Coding is migrating from the chat window to the pipeline

Here's Anthropic's own sentence: "Coding tasks continue to migrate from augmentative usage in Claude.ai to more automated workflows in our first-party API traffic." The number behind it: "Since August 2025, the share of tasks in this category has increased by 14% in the API and decreased by 18% in Claude.ai." Read plainly, two different populations are moving in opposite directions. People typing directly into a chat window are doing proportionally less coding through Claude than they were seven months ago. Code arriving through the API, meaning through someone else's IDE plugin, CI pipeline, or agent framework, is doing proportionally more, and doing it with less of a human approving each step along the way.

Concentration is the tell, not just the category share

The more decision-relevant number sits one layer down. Anthropic also tracks how spread out usage is across its top ONET tasks. On the API, "the top 10 ONET tasks now [account] for 33% of traffic, up from 28%" over the same window. On Claude.ai, it runs the other way: "the top 10 tasks made up 19% of all traffic in February, down from 24% in November." A shrinking top-10 share in chat means people are asking Claude to help with a wider, messier variety of things. A growing top-10 share in the API means fewer, narrower, more repeatable jobs are eating a bigger slice of that traffic, which is exactly the shape automation takes: not "AI does everything" but "AI does this one well-bounded thing, over and over, without anyone watching it happen."

What this data doesn't say

It's worth being precise about the limit of this evidence, because the honest reading is narrower than the exciting one. Anthropic is describing the mix of its own product traffic, not a claim about the software industry as a whole, and a shift toward "automated" classification says nothing about whether the automated code was correct, whether it shipped, or whether someone caught a bad merge before it hit production. It measures how a task was handled, not how well. Treat this as a real signal about where usage is concentrating, not as proof that unattended AI coding is now safe to trust by default.

A worked example

Picture a 35-person Series B company building a usage-based billing product, the kind of team this data describes well: senior engineers who used to paste a tricky migration script into Claude.ai and iterate on it line by line, and who have since wired an agent into their CI pipeline to open pull requests for a narrow set of jobs (dependency bumps, test scaffolding, the boring 80% of a schema migration) with no human in that loop until review. That's the automation share growing on the API side of the ledger. What doesn't move to the agent is the 20% of that migration that touches a table nobody fully understands anymore, the part where a senior engineer's judgment about what could go wrong in production is the entire value of having them on the team. The data supports handing over the narrow, bounded slice. It says nothing about the judgment calls sitting on either side of it.

What this means for your roadmap

The interesting engineering decision this year isn't "chat versus no chat." It's naming, in writing, which ten tasks your team is willing to let run without a person approving each step, and then building the eval that tells you the moment one of them silently breaks. Anthropic's numbers are a hypothesis about where that list should start, not a substitute for testing it against your own repository, the same discipline behind an eval harness built against your own golden set and the same gap that shows up in why agents that look great in a demo keep failing once they hit production. It's also worth reading alongside the other side of the productivity question: a separate, controlled study found experienced developers were measurably slower with AI tools despite feeling faster, which is the opposite failure mode from the one this report describes. One measures perception against reality on tasks that stayed manual. This one measures a real shift in which tasks stopped being manual at all. Both are true, and neither cancels the other out.

If your team is trying to work out which of your own ten tasks belong on that automatable list, and which ones a senior engineer still has to hold, that's exactly the scoping conversation our Silicon Valley team has with engineering leaders before a sprint gets planned around the wrong assumption. Talk to us about where staff augmentation from our build-your-team model fits around the parts of the roadmap that shouldn't move to an agent yet.

Sources

Frequently asked questions.

Anthropic's Economic Index is a recurring report that classifies roughly a million Claude conversations across Claude.ai and Anthropic's first-party API against the Department of Labor's O*NET occupational task list, tracking how usage shifts over time. The March 24, 2026 edition, subtitled 'Learning curves,' drew on a sample pulled February 5 to 12, 2026, and Anthropic publishes the underlying dataset on Hugging Face.

Automation describes a conversation where Claude completes a task with little ongoing input from the user, while augmentation covers interactions where a person stays in the loop through feedback, iteration or validation. Anthropic's March 2026 report found coding tasks moving toward the automation pattern in API traffic, while the same category shrank as a share of Claude.ai's more augmentation-heavy chat usage.

Anthropic's March 2026 Economic Index states that computer and mathematical task share rose 14% in the API and fell 18% in Claude.ai since August 2025, alongside rising concentration in the API's top ten O*NET tasks (28% to 33% over the same period). That pattern reflects coding work increasingly running through a CI pipeline, IDE plugin or agent framework instead of a live chat conversation.

No. The report measures how a task was classified, not whether the resulting code was correct or shipped safely. A rising automation share in the API shows where usage is concentrating, not that unattended AI coding has been validated for correctness, so treat it as a signal to test against your own codebase rather than confirmation that a workflow is production-ready.

Name, in writing, the narrow set of coding tasks your team is willing to let run without a person approving each step, then build an evaluation harness that flags the moment one of them silently breaks. Anthropic's numbers describe where automation is concentrating across its own product traffic; they are a hypothesis for where that list might start, not a replacement for testing it against your own repository.