Anthropic Economic Index 2026: AI Augments vs Displaces Freelance
Anthropic's January 2026 report shows 52 percent of Claude conversations are augmentation, 45 percent automation — up 5 percentage points in three months. Brookings' July 2025 freelance-market study found high-performing freelancers in AI-exposed categories lost 5 percent of monthly earnings. The two lenses disagree productively. Here is what they actually tell freelancers.
The Delivvo team· May 17, 2026 7 min read
Two datasets dropped in 2025-2026 that together give the clearest picture of where AI is actually changing freelance work — not in speculation, but in measured behaviour. They look like they contradict each other. Read carefully, they do not.
The first is Anthropic's Economic Index, with two new reports in 2026: *Economic Primitives* (January 15) and *Learning Curves* (March). The second is Brookings' July 2025 paper *Is generative AI a job killer? Evidence from the freelance market*. The Anthropic data measures conversation patterns inside Claude. The Brookings data measures earnings outcomes on a major freelance marketplace. They are different lenses on the same underlying shift.
This post combines them.
What Anthropic's data shows
The headline numbers from the January 15, 2026 Anthropic report:
52 percent of Claude conversations are augmentation; 45 percent are automation — up 5 percentage points on the augmentation side from August 2025. "Directive conversations" (the most pure-automation pattern, where the user issues a single command and accepts the AI output) fell to 32 percent, down 7 percentage points from August (Anthropic, Economic Index Report: Economic Primitives, January 15 2026).
Computer and Mathematical occupations account for 34 percent of Claude.ai conversations and 46 percent of API traffic. Educational Instruction roughly doubled from 9 percent in January 2025 to 15 percent in November 2025.
The skill profile of work Claude handles skews high: "Claude covers tasks requiring 14.4 years of education on average, while the economy's average task requires only 13.2 years." That is the equivalent of AI selectively handling the components of jobs.
Keep reading
higher-skill
The March 2026 follow-up report (*Learning Curves*) adds a methodological anchor that matters: task values are "derived from May 2024 BLS Occupational Employment and Wage Statistics (OEWS) Tables" — so when Anthropic talks about mean task value declining from $49.30/hr to $47.90/hr equivalent across that window, the dollar values are triangulated against real US labor data, not internal estimates (Anthropic, Economic Index Report: Learning Curves, March 2026).
The pattern the data tells: more augmentation, less automation. AI handling the higher-skill slice of a workflow, leaving the lower-skill slice for humans. Task values declining modestly. Computer / math / writing / educational work absorbing the largest share of usage.
For a freelancer reading this and thinking "good, augmentation means I'm safe," the Brookings data is the necessary second read.
Freelancers in AI-exposed occupations saw "a roughly 5 percent decrease in their total monthly earnings on the platform" post-ChatGPT-launch, controlled against non-exposed occupations.
They also saw "a decline of approximately 2 percent in the number of new monthly contracts."
The most counterintuitive finding: "those with stronger past performance...experience larger declines" — high-performing freelancers in AI-exposed categories are hit harder than low-performing ones.
The last finding is the one freelancers should sit with. The intuitive read of "AI commoditises low-quality work first" is wrong, according to the Brookings data. AI is competing directly with the better-paid, higher-quality work in copywriting, design, translation, and some software-engineering categories — the work where the AI output is "good enough" to substitute for a freelancer's premium offering, but where the freelancer's premium was built on relative quality rather than non-substitutable judgment.
How to reconcile the two
The lenses point at different things and that is what makes them useful together.
Anthropic measures behavior inside Claude. People are using AI more collaboratively over time. The mode of usage is shifting from "do this task for me" toward "help me think through this." That is genuine augmentation. It is also self-selecting — people who use Claude and are willing to converse with it iteratively are over-represented.
Brookings measures market outcomes on freelance work. When AI is good enough to do the task, fewer humans get hired to do the task, and the ones who do get hired earn less. That is genuine displacement. It is also marketplace-specific — Upwork-style freelance has different selection dynamics than direct-relationship freelance, and the substitution pressure shows up more visibly on the open marketplace.
The combined read: AI is augmenting work where the human keeps owning the judgment and the relationship; AI is displacing work where the human's role was largely to produce the deliverable from a clear brief.
That distinction maps cleanly to specific freelance categories.
A workspace with a designer's tablet, pencil, and reference materials — the kind of judgment-heavy creative work that resists pure AI substitution
The freelance categories the data argues are augmenting
These are the categories where Anthropic's augmentation pattern dominates and Brookings' displacement signal is muted:
Senior software architecture and judgment-heavy engineering work. Computer and Math is Anthropic's biggest API category at 46 percent. But the work happening inside those conversations skews toward debugging, code review, architecture discussion, and pair-programming patterns — not "write me a CRUD endpoint." Senior engineers using Claude or Cursor as a thinking partner are getting faster and shipping more. The Brookings displacement signal in junior implementation work is real; the augmentation pattern in senior work is also real.
Strategic consulting where the deliverable is judgment, not document. A strategy memo or competitive analysis written by Claude is mediocre; a strategy conversation co-facilitated by a senior consultant using Claude to synthesise inputs is substantially better than either alone. The consulting market in 2026 is bifurcating: commodity research-and-write gets displaced; partner-grade judgment work augments.
Education and instructional design. Anthropic specifically called out Educational Instruction as the fastest-growing usage category. Freelance curriculum designers, tutoring services, and corporate-learning content creators are using AI to scale personalised content per learner — augmentation in the cleanest sense.
Healthcare-adjacent and legal-adjacent work where liability and judgment matter. Document review augmented by AI is faster; the human reviewer is not displaced because the accountability remains theirs. Same for medical writing, regulatory submissions, and contract negotiation support.
The freelance categories the data argues are displacing
These are the categories where Brookings' earnings signal lines up with the kinds of tasks AI handles autonomously well:
Commodity copywriting. Blog posts, product descriptions, social-media captions, landing-page drafts. The AI output is now genuinely "good enough" to substitute for the median freelance deliverable in these categories. High-performing freelancers in this segment are seeing earnings declines because their differentiation (better writing) is exactly the dimension AI has closed the most ground on.
Translation for non-specialised content. Already heavily disrupted; remaining freelance work is consolidating around cultural review, legal translation where liability matters, and editing of AI output rather than translation-from-scratch.
Junior-tier visual design. Logo variants, banner ads, simple iconography, layout exploration. Tools like Figma's AI features and dedicated design-AI products are absorbing the bottom of the design market.
Data entry, transcription, low-complexity research synthesis. Already largely displaced by 2024; the residual freelance market is small.
What freelancers should actually do
Three concrete moves, derived directly from the data.
1. Specialise in the judgment slice. Both datasets agree: AI augments where humans bring judgment, AI displaces where the deliverable is the work product itself. Reposition your service to emphasise the part of the work AI cannot underwrite — accountability, taste, relationship, contextual interpretation.
2. Package judgment as productised offerings. A "regulatory readiness opinion letter" priced at $5,000 is a defensible product. $120/hour for "writing tasks" inside an undifferentiated invoice is not. The Brookings displacement signal is sharpest in hourly time-for-deliverable arrangements; productised offerings that price the judgment, not the time, resist substitution better.
3. Use AI yourself — visibly. Anthropic's data shows the freelancers using AI well are pulling ahead. Brookings' data shows the freelancers in AI-exposed categories who fail to adopt are losing earnings. The freelancers who win in 2026 are the ones who treat AI as the augmenting layer it can be when paired with judgment — and who tell their clients that's what they're doing.
Delivvo gives freelancers a branded client portal for proposals, contracts, file delivery, and invoices — the surface that makes productised judgment work feel professional from the first client touch through the final invoice. Connect your own payment gateway; Delivvo takes zero platform fee. See how it works →
The takeaway
The Anthropic and Brookings datasets together resolve the "is AI helping or hurting freelancers" debate more cleanly than either does alone. Augmentation is real where judgment lives. Displacement is real where the deliverable is the work itself. The freelancers who do well in 2026 are the ones who restructure their offers to put judgment at the centre and treat AI as the leverage layer underneath — not the ones waiting for the picture to clarify on its own.
The picture is clear enough. The next move is the freelancer's.
