Markets just had their worst day since 2020 and the Anthropic IPO is wobbling, but Menlo's latest State of Enterprise AI report tells a different story: business AI spend is accelerating from a bigger base, and agents are the reason. Budgets built for seats are colliding with a per-token reality.
Agents turned enterprise AI from a seat purchase into a token bill, and spending growth is accelerating.
The mainstream story about business AI is companies capping employee spend. Menlo's survey of roughly 400 enterprise decision-makers says the opposite: estimated enterprise AI spend jumped to almost $177 billion in 2026, and the growth rate sped up instead of slowing as the market matured. Breadth barely moved, but spend per enterprise rose 4.7x, about 2.5 times what companies budgeted. Companies planned for a per-seat world and got a per-token one. Frontier models, Anthropic and the lab coding agents captured most of the gains. Most of the money is net new budget rather than cannibalized software spend.
Menlo's estimates went from $1.7B in 2023 to $8.5B in 2024, $33B in 2025 and almost $177B in 2026. Growth had slowed from 5x to 3.9x as you'd expect from a maturing market. This year it outpaced even the early growth rate, starting from a far bigger base.
From Opus 4.5 last November onward, AI spend shifted from seats to tokens and everything that intelligence could do. Agents work independently, run in parallel and keep going after people step away. Spending now scales with work, not headcount.
The share of enterprises using AI rose only from 55% to 65%, which means 35% still report not using it at all. Average spend per enterprise using AI was up 4.7x over 2025.
The bear case: revenue concentrated in a handful of customers can drag the whole market down if any of them pulls back. The bull case, which Menlo takes: the heaviest spenders show where everyone else is headed, and the long tail has barely started.
The average company planned to grow AI spend about 1.9x this year: if you spent $10M in 2025, you expected $19M. Those plans assumed per-seat expansion. Actual spend jumped 4.7x.
This isn't necessarily runaway spend. It's people learning what all this raw intelligence can actually do. The whiplash between token-maxing and token restrictions is a symptom of the paradigm shift. AI budgeting looks somewhat like cloud compute, but even that analogy breaks down. It's a new category that needs to be planned on its own terms.
Frontier models' share of spend went from 35% to 63%, while mid-tier, small and older models all lost share. This happened even as labs cut small-model prices by more than 80%. Today's favorites, Claude Opus 5.5 ($4/$20 per million tokens) and GPT-6 Astro ($10/$60), cost more than last year's leaders.
Since Opus 4.5, businesses have been racing to catch up with genuinely differentiated agentic capabilities, and Opus 5.5 feels like another such jump. As adoption and the stack mature, cheaper and more efficient models will come in to form a more complex stack. The competition for those models still matters.
Anthropic's share went from 12% in 2023 to 24%, then 40%, and now 55%. The Opus models since Opus 4.5 have arguably been the default for long-horizon agentic coding, and that is now showing up in the numbers.
After falling from 50% to 27% over three years, OpenAI recovered to 31%. Menlo credits a strong second half: the GPT-5.6 series in July, GPT-6 in September, and an enterprise focus. Google fell from 21% to 9%, making enterprise AI a two-horse race for now. Gemini 4 Carbon's early indications look good but need more battle testing.
Open source's spend share fell from 6% to 4%, but its token share jumped from 11% to 19%. Near-frontier models like Grok 4.7, Kimi K3, DeepSeek V4 and GLM 5.3 together still take under 5% of API spend. In the medium to long term, token share is the more significant number, and that trend hasn't peaked.
Revenue for the labs' own coding agents grew 24x. That didn't come at the expense of Cursor, Cognition, Factory, Lovable and the rest, whose revenue still nearly tripled from $2.6B to $7.8B. The lab apps were simply far bigger.
Financial services apps grew 8x (to about $800M) and agentic coding grew 7.7x (to $34B). Customer success grew 4.3x, sales 3.6x and legal 2.8x, while marketing managed just 1.8x. The pattern suggests verifiable domains, where agentic-coding power can be brought to bear, benefited most.
AI-native companies take 96% of coding app spend, 95% of finance and operations spend and 81% of design spend, while sales is split 54/46. Incumbents win marketing, customer success, HR, data science and IT. In IT they take 91%, because spend flows to vendors already wired into systems of record.
Of the AI spending Menlo could trace, 55% came from net new budgets and just 21% from existing software budgets. SaaS growth slipped only from 13% to 11%. The next threat to SaaS, agents using software headlessly, has barely started.
Menlo's data broadly backs what the show has been tracking. It is also a reminder of a gap: the trends discussed in heavily enfranchised AI circles aren't yet the lived reality of the average company buying and using AI.
The Mag Seven lost a combined trillion dollars in market cap, with Nvidia down 9.3%. Micron fell 17.5% and SanDisk 22.3%. The S&P 500 fell 4.4% and the Nasdaq 6.4%, the largest single-day drop for the major indices since March 2020.
Nothing substantial changed between Wednesday and Thursday, and unrelated stocks like Walmart and ExxonMobil fell alongside tech. That points to market structure. Reports of liquidated and collapsing hedge funds suggest Wednesday's semiconductor drawdown triggered broad deleveraging. It could mark the end of the AI bull market, or a temporary blip like last year's tariff shock.
Anthropic had reportedly aimed to list before Thanksgiving, which would mean a roadshow starting next week. On Thursday the odds of that fell from around 85% to 60%. Odds of listing by year-end have dropped from 95% to 66% over the past week.
Thin secondary markets still price Anthropic around $1.6 trillion. Per the WSJ, three major prospective IPO investors say the valuation must come down, and one argues it should be closer to $1 trillion. Otherwise the reported $100 billion raise may need trimming. The same report puts October run-rate revenue at $80 billion, up from $65 billion at the end of July.
METR is Anthropic's first independent evaluator with employee-level access. Its first safety report concludes the acceleration came from AI assistance in research, though with "some signs that fully autonomous contributions will play a larger role soon." The review covered three unreleased models trained since July.
— New York Assemblyman Alex Bores, reacting to METR's Anthropic report. AI safety voices were rattled by the METR report, with some saying it accelerated their ASI timelines. Skeptics such as Nathan Lambert argue that a lack of good-enough benchmarks is still a bottleneck for an intelligence explosion.
Bringing in METR was meant partly to address public concern about cybersecurity breaches at the labs. The report doesn't touch sandboxing or other safety measures, so it is largely non-responsive to those concerns. METR can reasonably focus on recursive self-improvement, but if it's the only outside evaluator, the independent auditor suite is incomplete.
Anthropic relaunched Claude for personal finance with read-only bank access via Plaid. Claude helps with budgeting by analyzing spending but won't give investment advice. For bigger decisions, Anthropic says people are best served by Claude "working alongside the licensed professionals and regulated firms they trust."
Users can now talk to their named, personalized Muse agents over voice and video, with fully animated avatars and a range of voices. The avatars use expressions and gestures but can't see the user. The feature is live on mobile, web and Mac. It will power the forthcoming Muse device and works with the Muse Gadget SDK.
Images 2.6 lets users prompt short animations, adjust their speed and edit individual frames. It is 50% faster, with images in as little as six seconds, and accepts reference images such as rough doodles. Users now create four billion images a week, up from three billion a couple of months ago.
