VC & Startups/Nov 20 edition/Reporting & analysis
ExitsValuationsFunding

ExitsIPOs, listings & liquidity

Four engineering leaders detail their coding-agent token bills as Anthropic prices its $81.3 billion IPO

Business Insider profiled engineering leaders at Grafana Labs, Theory Ventures, Harness and Menos AI. Their coding-agent token bills, much of them for Claude, run from thousands of dollars a month to tens of millions a year. The story appeared as Anthropic priced its $81.3 billion IPO.

From the Business Insider · Startups article of November 18, 2026 and cited sources: Spending varies widely, and a few heavy users stand out. [1] [2] [9]

SOURCES[1] Business Insider original businessinsider.com[2] Meet 4 engineering leaders…… jingletree.com[9] Anthropic's IPO and the B2D Strategy – EE Times eetimes.com

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
Spending varies widely, and a few heavy users stand out. [1] [2] [9]
Original source
Business Insider · Startups article, independent perspective
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
When
November 18, 2026 article published
Evidence
11 sources across 10 sites

Reach: not measured yet

Illustration from Business Insider: Four engineering leaders detail their coding-agent token bills as Anthropic prices its $81.3 billion IPO
Image: Business Insider — Original article ↗
THE CORE IDEAS4 TAKEAWAYS
01

Spending varies widely, and a few heavy users stand out. A Theory Ventures engineer spent six figures in one month, mostly on Claude. Grafana spends tens of millions a year, about two-thirds of it on Claude. Harness runs about 265 billion tokens a week across more than 500 engineers. Menos AI spends $3,000 to $4,000 a month. [1] [2] [9]

02

Teams are building controls to manage costs. Grafana now gives engineers the most powerful models only in 24-hour access windows. Harness sets a soft cap of $125 per engineer per week and uses Sonnet to save money. Menos AI picks a model for each request based on how hard it is. Theory Ventures sometimes turns to cheaper open-weight models. [2]

03

The harness, meaning the wrapper around the model, can matter more to cost than the model itself. On 492 SWE-bench Verified tasks run on the same model, harnesses changed cost per task by up to 3x but moved accuracy by only a few points. Most of the gap came from the fixed prompt being resent at every step: 16,581 tokens for Claude Code versus 829 for mini-SWE-agent. About 13% of tasks also flipped between pass and fail on simple reruns. [4]

04

The IPO priced at $83 a share, raising $81.3 billion at a valuation of about $1.5 trillion. On its first trading day the stock closed at $97.17. The S-1 shows $16.3 billion in revenue for the first half of 2026, compared with $4.6 billion for all of 2025. About one-fifth of revenue comes from two customers. [3] [5] [6] [7] [8]

WHY IT MATTERS

a few heavy coding-agent users report very large bills, and the S-1 shows steep revenue growth from concentrated customers.

Read the full assessment

Implication: controlling spend may depend on harness design and the number of agent steps, while investors face concentration and seat-loss risk.

Status: research is done, and here is the dossier.

Executive brief

One engineer at Theory Ventures spent six figures on AI tokens in a single month, mostly on Claude. He says he could pass $1 million if he ran more simulations (Jingletree/Business Insider). Business Insider profiled four engineering leaders, at Grafana Labs, Theory Ventures, Harness and Menos AI, whose coding-agent bills help explain Anthropic's revenue. The story ran on the day Anthropic priced its IPO at $83 a share, raising $81.3 billion (Yahoo Finance). The figures come from the companies' own leaders, and no one has audited them.

What changed and event timeline

  1. Paper measures what a coding "harness" costs

    Shandong University researchers found that on the same model, the harness (the wrapper around the model) changes cost per task by up to 3× but changes pass rates little. Claude Code's fixed prompt was 16,581 tokens ().

  2. Anthropic files S-1

    Anthropic plans to list on Nasdaq under the ticker "ANT," with Goldman Sachs and Morgan Stanley as underwriters (). The filing shows $16.3B revenue for H1 2026 ().

  3. Business Insider publishes the spending profiles; IPO prices

    Four engineering leaders describe their token budgets (). The IPO priced at $83 per share: 980M shares sold, $81.3B raised, about $1.5T valuation ().

  4. First trading day

    ANT opened at $111.50, hit a high of $115.10 and closed at $97.17, about 17% above the IPO price (;).

  5. EE Times explains the strategy

    It describes Anthropic's approach as "business-to-developer": winning developers first, then the enterprises that employ them. It repeats the Grafana and Theory figures ().

Capabilities and access

  • Tools in use: Claude Code, Claude through Cursor, and OpenAI's Codex. Harness also uses Sonnet to cut costs (Jingletree/BI).
  • Model routing: Menos AI sends each user request to a different model depending on how hard it is.
  • Access controls: after an accidental overspend, Grafana gives engineers the most powerful models in 24-hour access windows.
Read the full section
  • Tools in use: Claude Code, Claude through Cursor, and OpenAI's Codex. Harness also uses Sonnet to cut costs (Jingletree/BI). The coverage does not say which model versions anyone uses.
  • Model routing: Menos AI sends each user request to a different model depending on how hard it is. Theory Ventures sometimes uses open-weight models because they are cheaper (same source).
  • Access controls: after an accidental overspend, Grafana gives engineers the most powerful models in 24-hour access windows. Harness sets a soft budget of $125 per engineer per week (same source).

Technical analysis for researchers and developers

  • Why the harness drives the bill: across 492 SWE-bench Verified tasks, Claude Code and mini-SWE-agent finished within five points of each other on accuracy.
  • Noise between runs: about 13% of tasks flip between pass and fail when the identical setup is simply rerun.
  • Limits of the paper: it uses one benchmark, and the closed frontier model was run only in Claude Code (same source).
Read the full section
  • Why the harness drives the bill: across 492 SWE-bench Verified tasks, Claude Code and mini-SWE-agent finished within five points of each other on accuracy. Cost differed by up to 3×, mostly because the fixed prompt (829 vs 16,581 tokens) is resent at every step (arXiv).
  • Noise between runs: about 13% of tasks flip between pass and fail when the identical setup is simply rerun. That is roughly the same as the difference between harnesses, so single-run comparisons can mislead (same source).
  • Limits of the paper: it uses one benchmark, and the closed frontier model was run only in Claude Code (same source).
  • Practical implication: shrinking the prompt overhead and cutting the number of agent steps may lower spend more than switching models.

Claims and evidence

  • Grafana: "tens of millions" a year in tokens, about two-thirds of it on Claude. This comes from CTO Tom Wilkie (EE Times).
  • Harness: 265B tokens a week across 500+ engineers, per Kiran Gangadharappa (Jingletree/BI).
  • Menos AI: $3,000–$4,000 a month, per cofounder Chris Yang (same source).
Read the full section
  • Grafana: "tens of millions" a year in tokens, about two-thirds of it on Claude. This comes from CTO Tom Wilkie (EE Times).
  • Harness: 265B tokens a week across 500+ engineers, per Kiran Gangadharappa (Jingletree/BI).
  • Menos AI: $3,000–$4,000 a month, per cofounder Chris Yang (same source).
  • Anthropic's financials: H1 revenue of $16.3B against $4.6B for all of 2025, $9B spent on infrastructure, and about one-fifth of revenue from two customers (Yahoo Finance). These are numbers from the company's SEC filing, not an outside audit.
  • Customer tiers (per BI): 11,900 customers spend $100K+ a year and 250 spend $10M+. Typical non-Claude Code users average under $75 a month (Jingletree/BI).
  • Valuation conflict: BI described a target of about $1.4T, while live IPO coverage reports about $1.5T at pricing.

Context and prior work

  • Business mix: EE Times reports that 74% of Anthropic's revenue in the 12 months to September 2026 came from customers paying more than $100K a year, and that Claude Code reached a $1B run rate within six months of launch (EE Times).
  • Comparison with OpenAI: for Q2 2026, Anthropic had $11.6B in revenue against OpenAI's $6.7B (same source).
  • Prior spend tracking: teams were already tracking Claude Code costs with Grafana dashboards before this story (Grafana).
Read the full section
  • Business mix: EE Times reports that 74% of Anthropic's revenue in the 12 months to September 2026 came from customers paying more than $100K a year, and that Claude Code reached a $1B run rate within six months of launch (EE Times).
  • Comparison with OpenAI: for Q2 2026, Anthropic had $11.6B in revenue against OpenAI's $6.7B (same source).
  • Prior spend tracking: teams were already tracking Claude Code costs with Grafana dashboards before this story (Grafana).

Limitations, safety and contested findings

  • Sample: four leaders, all quoting their own spend, is anecdote, not a market sample. One of them, Bischof, works at a venture firm.
  • Lost seats: EE Times notes that one Microsoft division cancelled most of its Claude Code licenses (EE Times).
  • Losses: the S-1 shows an $80.7B net loss through June, mostly from accounting charges, and up to $637B in compute commitments (Yahoo Finance).
Read the full section
  • Sample: four leaders, all quoting their own spend, is anecdote, not a market sample. One of them, Bischof, works at a venture firm.
  • Lost seats: EE Times notes that one Microsoft division cancelled most of its Claude Code licenses (EE Times).
  • Losses: the S-1 shows an $80.7B net loss through June, mostly from accounting charges, and up to $637B in compute commitments (Yahoo Finance).
  • Risk factors: the filing lists misuse, loss of human control, authoritarian abuse and economic disruption (CNN).

Business and practitioner implications

  • Who drives revenue: a small group of heavy users matters most.
  • Budget controls: expect per-engineer caps like Harness's $125 a week, time-limited model access like Grafana's, and model routing.
  • Customer concentration: anyone buying ANT stock should weigh the share of revenue from two customers and the risk of losing developer seats.
Read the full section
  • Who drives revenue: a small group of heavy users matters most. BI reports that the top 5% of coding-tool users average about $1,500 a month, against under $75 for typical users.
  • Budget controls: expect per-engineer caps like Harness's $125 a week, time-limited model access like Grafana's, and model routing.
  • Customer concentration: anyone buying ANT stock should weigh the share of revenue from two customers and the risk of losing developer seats.
  • Cost levers: the choice of harness and the number of agent steps matter as much as the price per token.
FOLLOW THE EVIDENCE

The source trail.

Sources (11)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.