Wall Street to brunch: Claude's wins, on Anthropic's word. Who checked?
Anthropic's IPO filing carries 20 customer case studies. Here is what Claude actually did inside banks, drug makers, law firms and kitchens, what it saved, and what the numbers can't tell you.
customers paying Anthropic $100,000+ a year (TTM, 30 Sept 2026)
customers paying $10 million+ a year
net revenue retention: existing customers spend about 9x more
revenue in H1 2026, up from $1.2B a year earlier
From months to minutes, in six companies
Bars are drawn to scale where the filing gives both a before and an after.
Bristol Myers Squibb
Pharma · customer since 2023Claude now queries the 30 to 40 external medical-records and claims databases BMS subscribes to, so scientists ask questions directly instead of waiting for a dashboard.
Watch "Up to" two months is the slowest case, not the average one. Cowork for all staff is still a plan.
AIG
Insurance · customer since 2023Underwriting by AIG Assist reads every broker submission and hands underwriters a risk summary with a full audit trail, work that used to take weeks of data gathering.
Watch Lexington figures are as of 31 March 2026. Claims pilots (days to hours) are early.
Jane Street
Trading · customer since 2023Researchers now supervise teams of agents that analyse code, watch training runs and process results. Experiments that needed babysitting run over a weekend.
Watch Both headline numbers are single examples or estimates, not firm-wide measures.
Legora
Legal tech · customer since 2024Claude is the reasoning layer in Legora's platform for lawyers, planning drafting, document review and research, and tying every answer to a source.
Watch The 4.3 hours is what lawyers reported in a study of 31 firms in 14 countries, and it measures Legora's product.
Mozilla
Browser security · customer since 2025Anthropic's Frontier Red Team turned Claude loose on Firefox, one of the most heavily scrutinised open-source codebases, to hunt for flaws defenders had missed.
Watch A model this good at finding bugs is also good at exploiting them. Mythos access is restricted for that reason.
Salesforce
Enterprise software · customer since 2023Claude Code is the main AI agent across Salesforce engineering, and Claude powers Slackbot, which nearly every employee uses.
Watch Days and person-days are different units, and Salesforce is also an Anthropic shareholder.
Twenty more ways Claude is earning its keep
Pick a sector. Every figure is from Anthropic's S-1 case studies or its 27 October drug-development report.
Showing 20 use cases
Enveda
Drug manufacturingreview time: over 25 hours to under 3
Checked 1,061 pages of manufacturing records. Caught all 11 issues humans found, plus 7 they missed. One regulatory draft took 4 hours instead of about 3 months.
Generate Biomedicines
Cell therapymore receptors in a lab test
Six scientists used Claude Code to redesign part of a CAR-T therapy. One design, about 70% shorter, also killed tumour cells more efficiently.
Divergence Labs
Allergy drugsantibody designs in about a month
Early tests found 20 designs that dislodge IgE; three appeared to beat the standard medicine.
BigHat Biosciences
Antibody designtighter binding in the best variants
About 1,010 variants made and tested in about 10 weeks. Claude also helped expose misleading stability metrics.
Mithril Medicine
Rare diseasefrom idea to promising mouse results
A one-person company used Claude to run work across 11 outside labs. Anthropic bought it; human trials possible as soon as late 2028.
Eli Lilly
Pharmaemployees with Claude Enterprise
First big drug maker to roll it out company-wide. More than 85% of staff in key functions use it actively.
Mercado Libre
E-commercecode submissions reviewed by agents in a year
23,000+ developers. About 9,000 services migrated. In Q2 2026 merged code more than doubled and deployments rose nearly 75%.
DoorDash
Deliverylaunch timelines for major products
One engineer did in three weeks a migration that once took four engineers more than a quarter. 7,000+ staff use Cowork monthly.
Rakuten
Internettime to market for a new feature
Claude Code coded for seven hours with little guidance at 99.9% accuracy. Major releases now ship every two weeks, not every quarter.
UST
Chip testingchip validation cycle time
Claude Code reads chip schematics, then writes and runs regression tests. Standard validation fell from 4 days to 2.
Shopify
Commerce softwaredaily active merchants using Sidekick, year on year
Its merchant assistant runs on Claude. More than 7,500 employees have Claude Code.
Verizon
Telecomsengineers building with Claude Code
Also testing Mythos through Project Glasswing to harden its network.
NASA
SpaceAI-planned drive on Mars
Engineers used Claude Code to plan and write navigation commands for the Perseverance rover.
PayPal
Paymentssaved per sales team member each week
12,000+ monthly users. A reporting tool cut report prep to 5–10 minutes and surfaces merchant growth signals.
Travelers
Insuranceemployees with Claude assistants
Started with nearly 10,000 engineers, data scientists and analysts in January 2026.
Walmart
Retailaverage order value via its Sparky assistant
Sparky weekly users rose 110% in a quarter and units sold through it doubled.
Accenture
Consultingsecurity test coverage with Cyber.AI
Scans that took days now take under an hour. One broadband client reported a 7x return on its AI spend.
SAP
Enterprise softwaresaved per consultant per month
Its Joule for Consultants tool, using models such as Claude, delivers projects up to 14% faster.
Shy Bird
Restaurantshorter brunch ticket times, to 12 minutes
An owner who had never coded fed kitchen photos and sales data to Claude and redesigned the line. The tool now serves 60 restaurants.
International Rescue Committee
Humanitarianstaff hours returned each week
Field health teams lifted the share of key indicators used in planning from 20% to 100%. A week-long funding analysis now takes hours.
Where the gains are biggest
Paperwork with a right answer
Underwriting files, manufacturing records, data rooms. Output can be checked, so errors get caught.
Doing what was skipped
AIG now reads every submission; Enveda's review found 7 issues people missed. Coverage, not just speed.
Code migrations
Salesforce, DoorDash, Jane Street, Mercado Libre. Big, rule-bound rewrites are the clearest win.
More output, same team
Legora runs a platform for 1,500+ firms with about 80 engineers. Good for margins, harder for hiring.
Defenders get a head start
Mozilla fixed 293 flaws attackers never got to use. The same skill helps attackers.
Science, early days
Lab results are real but mostly preclinical. No drug from this work has reached humans yet.
Claude is building Claude
of contributions to Anthropic's codebase in September 2026 were written by Claude
more code per engineer per day than in 2024 (Q3 2026 average)
Claude's success rate on hard, open-ended engineering problems, up ~50 points in nine months (internal benchmark)
headcount growth vs revenue growth, Q4 2024 to Q3 2026
revenue per employee (TTM), with about 5,300 staff
cost per task for customers, Opus 5.5 vs Opus 5
What the case studies leave out
Who keeps the score
Every case study is the customer marking its own homework, published in a sales document. None is audited, and the S-1 warns its own metrics are unverified.
Feeling fast is not being fast
In METR's 2025 trial, 16 experienced developers took 19% longer with AI tools, yet believed they were about 20% faster. Tools have improved since, but self-reports deserve doubt.
The first rung of the ladder
Employment of 22- to 25-year-olds in the most AI-exposed jobs is about 19% below peers in less-exposed work, mostly through less hiring. The authors call it a pattern, not proof of cause.
Confident and wrong
In June a US judge cancelled a trial after lawyers on both sides filed AI-invented cases. Legora grounds answers in sources for exactly this reason.
The bill for agents
Anthropic says agentic work burns far more compute per task and that it has "significantly underestimated" compute needs before, while H1 compute costs were $8.9B.
Off switch you don't control
In June the US ordered Fable 5 and Mythos 5 cut off for foreign nationals; access returned on 1 July. Its five biggest customers also give about 25% of revenue.
Models with their own ideas
The filing warns Claude may act in "unintended, misaligned, or autonomous ways", and that a flaw could pass into future models now that Claude writes them.
Bugs cut both ways
Mozilla's 271 fixes show defence at a new scale. A model that finds flaws this well could arm attackers, which is why the strongest models are gated.
How to get the gains without the grief
- 1Start with checkable work: migrations, document review, triage. Avoid tasks where errors are hard to spot.
- 2Measure against a control group or a clock, not staff surveys. METR shows how far feelings drift.
- 3Keep a human signature on anything filed, shipped or sent to a client.
- 4Budget for usage that grows: agent tasks cost more than chat, and bills follow use.
- 5Keep a second model ready. A rule change or outage can switch one off overnight.
- 6Plan how juniors learn when the routine work that trained them goes to agents.
The receipts are real, but they were printed by the seller. Trust the stopwatch, not the brochure.
Read the originals
- Anthropic, PBC Form S-1, including customer case studies (SEC, 26 Oct 2026)
- Anthropic files registration statement for proposed IPO (Anthropic, 26 Oct 2026)
- Early progress in drug development (Anthropic, 27 Oct 2026)
- Measuring the impact of early-2025 AI on experienced open-source developers (METR, July 2025)
- Canaries in the coal mine, August 2026 update (Stanford Digital Economy Lab)
- US judge stops case after lawyers on both sides cite AI-fabricated cases (Legal Cheek, June 2026)
- US lifts restrictions on Fable and Mythos models (Al Jazeera, 1 July 2026)
