The State of AI 2026 Mid-Year Reality Check

0:00Matt Paige:Hey everyone, Matt Paige here, host of Talking AI, and we're doing something a little different with this special episode. So Haptrix AI just released our State of AI 2026 Mid-Year Reality Check, and it's a comprehensive look at what has fundamentally changed since January and where AI is headed in the second half of the year. And instead of only publishing it as a written report, we're using 11 laps to turn the full report into an audio experience. So the voice you're about to hear is AI-generated, but the research, the analysis, and the point of view came directly from our team at Haptrix AI.

0:26Matt Paige:The report covers everything from the latest jump in model capabilities to the growing pressure to prove ROI, the rise of open models, the new enterprise AI stack, agent security, the impact on jobs, and our predictions for the rest of 2026. You can also grab the complete written report through the link in the show notes. With that, here's the State of AI 2026 Mid-Year Reality Check. From the experts at Hatchworks AI, State of AI 2026, the mid-year reality check, a comprehensive roundup of industry stats, research, and critical insights on where artificial intelligence stands at the half, what has fundamentally changed since January, and where the second half of the year is headed.

1:05Co-authored by Brandon Powell, Matt Page, and Omar Shanti. Edited by Andy Smith. Copyright 2026. Hatchworks AI. All rights reserved. One line from the introduction frames everything that follows. The value is real. The spend is real. And the gap between the companies getting one in exchange for the other and the companies getting neither has never been wider. Chapter one. Mid-year by the numbers. Ten numbers. Ten storylines. One number per storyline. Three times. That is how much better the newest model generation holds long horizon, multi-step work compared with models from six months ago, according to Anthropic.

1:51680 times. That is the gap in monthly AI spend per employee between the top 1 % of firms at$7 ,450 and the median firm at$11. Source, the Ramp AI Index, June 2026. 41%. The share of U.S. businesses with a paid Anthropics subscription, now the most adopted AI provider, with OpenAI Flat. Also from the Ramp AI Index. $60 billion. What SpaceX paid for Cursor's parent company, AnySphere. All stock, four days after its NASDAQ debut. Reported by CNBC. Two. The number of frontier labs that confidentially filed for IPOs in June, Anthropic and OpenAI, both at valuations near$1 trillion. 18 days. How long?

2:46Fable 5, the most capable model ever released publicly, was suspended under a U.S. export control directive. We will come back to that story in Chapter 6. Number 1. DeepSeek's rank on Ramp's June trending vendor list, a live read on the appetite for price competitive open models, roughly 50%. How much Coinbase cut its AI spend while token usage kept growing through open model defaults, routing, and caching. Brian Armstrong, July 2026. 1.3 billion. The number of AI agents predicted to be operating by 2028, according to IDC, plus 10%. Headcount growth at heavy AI adopters in the two years after adoption.

3:35Entry-level roles grow 12%. Source. The Ramp Economics Lab. June, 2026. Chapter 2. The Plateau That Wasn't. In January, we ran a section titled Models Are Plateauing. Architectures Are Getting Smarter. It was the consensus view, and half of it aged very well. The other half got overtaken fast. The models that landed between late last year and the first half of this year represent a genuine step change rather than an increment. Anthropik's Opus series reset expectations for agentic work, and the fable generation went further. Anthropik's own language was uncharacteristically direct, capabilities that exceed those of any model they have ever made generally available, state-of-the-art on nearly every tested benchmark.

4:26OpenAI, for its part, has released its answer, GPT 5.6, signaling the next capability wave, and it is proving to be a truly formidable competitor to Fable. The frontier has found another gear. The receipts came fast. Stripe used Fable to run a 50 million line code migration, the kind of project measured in engineer months, in a single day. On long horizon work, the ability to hold the thread across hundreds of steps without losing the plot, the new generation performs roughly three times better than the models of six months ago. That last number matters more than any benchmark. The thing that held agents back was rarely raw intelligence.

5:08It was follow-through, memory, planning, and consistency across long tasks. A model that is three times better at not losing the plot is the difference between an assistant you supervise and a system you delegate to.

5:24Matt Paige:Quick break in the pod. I keep hearing the same pattern with companies I talk to. Cloud's helping employees move faster, but in many companies, the business itself hasn't changed. The value's still trapped in isolated chats and experiments. And that execution gap is why forward deployed engineers have become one of AI's most talked about deployment models. They embed with your team instead of advising from the outside. It's also why the FDE model is now central to every client engagement we lead at Hatchworks AI. As an official Anthropic partner, we embed Anthropic certified FDEs to identify high value business problems, build and deploy the solution and put governance and security around it, then transfer the capability back to your team.

5:56Matt Paige:If your cloud rollout is still mostly individual usage, check out how Hatchworks AI FDEs work at hatchworks.com slash cloud dash FDE. You can also find it in the show notes. Now back to the show. And here is where January's architectures are getting smarter. Call proved exactly right. Arguably more right than we knew. The capability jump is the model multiplied by everything now wrapped around it. The agent harness, meaning verification steps, sub-agents, permissions, and guardrails. Persistent memory, skills that encode how work should be done, and tool ecosystems that let models act instead of just answer.

6:35Better models made the scaffolding more valuable. Better scaffolding extracts more from every model. The compounding of the two is what moved the ceiling. which leads to the sentence we would put on the wall if we could only keep one. If your organization concluded the technology isn't ready, that conclusion has expired. The practical consequence, there are entire categories of work that simply were not automatable 12 months ago that now are. Multi-day engineering projects, end-to-end research and synthesis, workflows that chain dozens of tools and decisions. If your organization scoped its AI ambitions in 2024 or even in mid-2025 and concluded the technology wasn't ready, that conclusion has expired.

7:25For most enterprise work, the capability question is settled. The open question is who is positioned to turn that capability into a number ACFO will accept. Which brings us to the theme of the year. Chapter 3. From Token Maxing to Show Me the ROI. The first half of 2026 had an unofficial strategy, and it was token maxing. Scale through compute. Spend as strategy. Throw the new capability at everything and sort it out later. You could see it in the numbers. Enterprise AI investment hit$37 billion in 2025, tripling in a single year. And the freshest 2026, Reed shows where the momentum concentrated.

8:11Per Ramp's June AI Index, the top 1 % of firms now spend about$7 ,450 per employee per month on AI, while the median firm spends roughly$11, a 680-fold gap. Boards approved AI budgets the way they once approved cloud budgets, on the theory that underinvesting was the bigger risk and the most aggressive boards approved them at a scale the median company has not begun to imagine. The back half of the year is shaping up differently. The phrase we hear in every boardroom now is some version of show me the ROI. Notice how much that question concedes. Does AI work? Ended as a debate a while ago. Are we getting value?

8:58Almost everyone can point to something. The live question is sharper. Is the value worth the relative spend? Token bills are now a real line item. Co-pilot seats, API invoices, and platform fees compound quietly, and the CFO has noticed. The burden of proof moved from the skeptics to the spenders. Here is the uncomfortable truth underneath that question, and it is the central argument of this report. The ROI variance between companies has very little to do with the AI. Everyone is buying roughly the same models. The same frontier capability is available to your company and your competitor for the same per-token price.

9:43Yet one of you will clear the bar, and one of you will not. The difference lives in three places, none of which come in the box. First, data connection. A model that cannot see your systems produces generic output at premium prices. Companies clearing the ROI bar connect AI to their CRM, their warehouse, their documents, and their communications, operating directly on business data instead of the Internet's average. Second, workflow embedding. Value comes from AI wired into the actual flow of work, the ticket queue, the deal desk, the close process, The deploy pipeline, usage that lives off to the side in a chat window, evaporates.

10:29Usage inside the active workflow compounds over time. Third, and this is the stubborn one, adoption. Access does not equal adoption, and seats do not equal changed behavior. The MIT finding showing 95 % of pilots failing to hit the P &L highlights that AI success has always been a people, process, and integration challenge. The same MIT research contained the tell. Purchase tools and specialized partners succeed about 67 % of the time, while internally built solutions succeed 33 % of the time. The gap comes down to where the failure modes live. in integration and adoption, exactly where outside operators with pattern recognition across dozens of deployments have an advantage and where a solo internal team is learning everything for the first time.

11:27Call this the 2026 posture shift from token maxing to relentless execution. The second half of 2026 belongs to a disciplined approach. Instrument spend, connect data, embed workflows, drive adoption, and measure relentlessly. Companies that do this will find the ROI question easy. The others will find budget season very long. Chapter 4. The lab landscape found its new equilibrium. The who's winning conversation got stale precisely because the answer stabilized. The interesting story now is what the settled board looks like and what each player's position tells you about where this goes. Anthropic, the enterprise incumbent.

12:14Anthropic is the enterprise incumbent now. That sentence would have read as provocation 18 months ago. Today, it is just the data. Menlo Ventures' year-end 2025 study, still the most recent comprehensive read of the market, put Anthropic at 40 % of enterprise large language model spend, up from 24 % a year earlier and 12 % in 2023, plus a commanding 54 % of the AI coding market on the strength of Claude Code. The fresher 2026 signals say the lead is holding. Ramp's June AI Index shows Anthropic as the most adopted paid AI provider among U.S. businesses, present in 41 % of them. The Fable launch extended the capability lead at the exact moment enterprises were consolidating vendors.

13:07The story has shifted from an upset in progress to a leader defending a position, OpenAI. The agentic comeback. OpenAI staged the comeback nobody fully priced in. After watching its enterprise share slide from 50 % in 2023 to 27%, OpenAI turned the consumer flywheel into an enterprise engine. Enterprise now represents more than 40 % of OpenAI's revenue, on track to reach parity with consumer by the end of the year, on a run rate north of$25 billion annually. The pattern to watch. OpenAI is winning the agentic workflow purchase, selling outcomes and platforms to buyers who came for chat GPT and stayed for the stack.

13:55Counting them out was a mistake in 2024, and it is a mistake now. Going public. Confidential filings and market impact. Both leaders have quietly lined up the option to go public. Anthropic filed a confidential S1 on June 1st, on the heels of a May round at a$965 billion valuation, with reports pointing to a listing window as early as October. OpenAI filed confidentially in June as well, near a$1 trillion valuation, and is reportedly weighing whether to wait until 2027 after SpaceX's bumpy debut. A confidential filing only sets up the option to list, with no date committed, so timing stays flexible.

14:40Late this year or next is the likeliest window. The significance for enterprise buyers is bigger than the tickers, because public companies disclose. If these two list, the economics of the frontier labs, their margins, their compute obligations, and their customer concentration become visible quarterly, expect that transparency to reshape pricing conversations. And expect both labs to chase enterprise revenue even harder, because that is the number public investors reward. Chapter 5. The Distribution and Price Frontier. Google and Gemini. Quiet, compounding. Google is compounding quietly. From 7 % of enterprise share in 2023 to roughly 21 % now, Gemini is doing what Google does best, showing up everywhere its cloud and workspace footprint already is.

15:35Google wins deals nobody tweets about. In the enterprise, that is often what durable looks like. NVIDIA and Nematron. The quiet giant. And then there is NVIDIA's quiet giant. Nemetron, NVIDIA's family of open models, is remarkably capable and conspicuously under-marketed. The strategic logic is almost funny. NVIDIA's best customers are the frontier labs, and aggressively pushing models that compete with your customers is bad business. So Nemetron advances steadily, wins serious deployments, and stays out of the headlines. For enterprises, it is one of the best-kept non-secrets in the market. Frontier-adjacent capability, open weights, backed by the one company nobody in AI can afford to alienate.

16:26SpaceX and Cursor, the distribution game. SpaceX bought Cursor, and distribution became the game. On June 16th, four days after its NASDAQ debut, SpaceX, which absorbed XAI in February, exercised its option to acquire AnySphere, the maker of Cursor, for$60 billion in stock. Cursor brings roughly$2.6 billion in business-to-business revenue and a beachhead on millions of developer desktops. Read the move plainly. XAI's models were not winning the enterprise on merit, so SpaceX bought the Surface developers already live in. Expect more of this. When model quality converges, the fight moves to distribution, and the checkbooks in this industry are extraordinary.

17:17The Chinese Openweight Labs, DeepSeek, Quen, and the CostTrust Calculus. The Chinese Openweight Labs stopped being a footnote. DeepSeek previewed V4 in April to serious reviews, and Alibaba's Quen passed 1 billion hugging face downloads by March. faster than any open model family in history and now accounts for over half of all global open model downloads. The capability story is real. Near frontier coding and reasoning at prices that barely register. A coding session that runs about$10 on a frontier U.S. model can cost under 50 cents on DeepSeek. For enterprises, the calculus is uncomfortable.

18:04Open weights mean you can inspect them and run them entirely inside your own infrastructure, where no data ever leaves. But the trust questions are just as real. Provenance, alignment, and increasingly, politics. U.S. companies, including Airbnb and AnySphere, have drawn congressional scrutiny simply for disclosing that they use Chinese open models. Every enterprise will land on this trade-off in the next year, deliberately or by accident. Land on it deliberately. Self-hosting neutralizes most data exposure risk, but the reputational and regulatory exposure is a board-level call, not an engineering one.

18:46The composite picture. Capability is consolidating at the top. Price performance is collapsing underneath from two directions at once. And the competitive frontier has moved from benchmarks to distribution, trust, and price. Keep that trio in mind. The next two chapters are about the trust part. Chapter 6. Fable, the 18-day ban, and the arrival of trust-tiered AI. The defining AI story of the half-unfolded in the four weeks after a product launch. Here is the short version. Because the sequence matters. Early June. Anthropic releases Fable 5, the public mythos class model. It is the same underlying model as its sibling, wrapped in safety classifiers that root the most dangerous categories of request, offensive cyber, biological risk, and model distillation, to a safer fallback model.

19:44More than 95 % of sessions never touch the fallback at all. June 12th. A U.S. export control directive citing national security requires that foreign nationals be cut off from access. Unable to reliably determine which users are foreign nationals, Anthropic suspends both Fable 5 and Mythos 5 for everyone. The world's most powerful public model goes dark. July 1st. The Department of Commerce lifts the directive. Fable comes back globally after 18 days offline. July 8th. Anthropic begins requiring identity verification, a government-issued photo ID and a live selfie processed by a third-party provider for consumer accounts that want full Fable access.

20:35API customers are exempt. July 20th. Pressured by the capability of OpenAI's GPT 5.6 sole, Anthropic makes Fable 5 part of all Max plans, with the first 50 % of usage included at no additional cost. Seven weeks. A genuine turning point in the AI arms race. Perspective and context. Sit with how much is packed into one month. A frontier lab shipped its most capable model with a safety fence it spent a quarter building. A government demonstrated that it can and will switch off a frontier model by decree. And the resolution, the thing that got the model turned back on, was identity. Proving who is on the other end of the prompt.

21:22We think this is the arrival of something enterprises should internalize quickly. Trust tiered AI. The old model was one API, one price, capability gated by your credit card. The emerging model is capability gated by who you are. Anonymous users get one tier, verified individuals another, vetted organizations another still, and sovereign or air-gap deployments sit at the top of the trust ladder. Access tiers are becoming the new pricing tiers. The rest of the frontier is already adjusting to the lesson. OpenAI's approach to GPT 5.6 is a preview-first, staged rollout rather than a big-bang launch, easing its most capable model into the world with graduated access instead of daring regulators to react.

22:12Having watched a competitor's flagship go dark for 18 days, no lab wants to ship first and negotiate second. Measured, tiered, verified rollouts are becoming the frontier default. For enterprise buyers, four implications. One. Access is now a risk surface. If your workflows depend on a frontier model, a policy decision in Washington, or wherever your provider is domiciled, can interrupt your operations. The fable ban was 18 days. Your business continuity planning should now include the line, Our model got turned off. 2. Identity and compliance are becoming product features. Expect verification, usage, attestation, and audit requirements to spread across providers and expect your procurement and legal teams to care about them.

23:06Three, the trust ladder favors the prepared. Organizations that can demonstrate governance, access control, and auditability will get earlier and broader access to frontier capability. That is a new kind of competitive advantage, and it is buildable. 4. Trust runs both directions, and your intellectual property is part of the equation. The deeper your workflows, data, and fine-tuning flow through a model maker, the more of your business it understands. And the labs are demonstrably willing to move up the stack. Anthropics product expansion this year, with Claude Design being the sharpest example, landed squarely on territory held by its own ecosystem partners, Figma most visibly.

23:54So here is the uncomfortable question every company building on a frontier platform should ask. What happens when my model vendor ships my product as a feature? Defensibility now means being deliberate about what you send upstream, keeping your differentiating data and evaluation assets in your own hands, and structuring vendor relationships with the assumption that today's platform partner is tomorrow's competitor. Chapter 7. Sovereign AI Gets Real The fable episode was the demand signal. The supply side answered in the same quarter. The air-gapped stack. In June, Palantir announced an intelligent engine built on NVIDIA Nimitron Open Models to deliver AI for U.S.

24:40government agencies and critical infrastructure operators, deployable in fully air-gapped environments. The detail that matters. Agencies fine-tune the open models on their own data and own the resulting weights outright. No external API. No data leaving the boundary. No dependency on a commercial provider's uptime or policy posture. Palantir's stack enforces authorization and isolation. Nimitron supplies the intelligence. And the agency keeps the keys. Enterprise Adoption This is sovereign AI moving from conference keynote concept to procurement reality. And the appetite extends well beyond government.

25:23The same architecture, open models, private fine-tuning, owned weights, isolated deployment is exactly what a bank, a defense contractor, a healthcare system, or any regulated enterprise increasingly wants for its most sensitive workloads. NVIDIA notes that about two-thirds of companies already use open models in some form, and cost savings are a stated driver, national scale. Nations are running the same play at larger scale, standing up national compute, national models, and national data policies. Because intelligence infrastructure is being treated the way energy and telecommunications were treated in prior eras, too strategic to rent entirely from someone else.

26:12Trust and control. The through line from the last chapter is that trust and control are becoming the product. The Frontier Labs sell capability with guardrails and identity. The Sovereign Stack sells capability with ownership and isolation. Most enterprises will end up buying some of both, which raises the obvious question. How do you decide which workloads go where? That is a portfolio question, and it is the subject of the next chapter. Chapter 8. Open source is the enterprise hedge. Strategically, the most consequential shift of the half is bigger than any single model. A two-track market has matured, and it changes how enterprises should buy.

27:00Track 1 is Frontier Models, Fable, the GPT-5 class, Gemini, at premium prices, for work where the capability ceiling is the point. Track 2 is Open Models, Nemetron, the Llama class, the Quen class, and their descendants, which have crossed good enough for a large share of enterprise workloads, at a fraction of the cost, with weights you can own and deploy anywhere. The strategic argument for taking track too seriously is about dependence rather than ideology. Concentrating your AI stack on a single frontier provider is concentration risk in four flavors. Price risk. Your unit economics are hostage to someone else's pricing decisions.

27:47Policy risk. The fable ban made this concrete. Terms change, models deprecate, access rules shift, and, now demonstrably, governments intervene. Roadmap risk Your product strategy inherits your provider's priorities, which are not your priorities. Data risk The more of your proprietary context flows through someone else's API, the more your moat commutes. The numbers behind the hedge are hard to ignore. Open models now cover roughly 80 % of proprietary use cases at per token pricing 10 to 20 times lower than Frontier rates. And on input, the gap can run 35 times or more. Above tens of millions of tokens per day, self-hosting saves 40 to 60 % versus equivalent API spend.

28:40And a blended pattern, open models for routine work with frontier escalation for hard cases, cuts total costs 70 % to 80 % against running everything on a frontier model, at a quality most users cannot distinguish. So the hedge is a portfolio. Frontier models for the hard 20 % of work where the capability gap genuinely matters. Open models for the routine 80 % where it does. Not. and, critically, your own evaluation harness and data layer so you can measure the gap and reroute workloads as the models change. The companies that build that switching capability stop being price takers. One pattern we increasingly recommend within the portfolio is the advisor model strategy.

29:28Use the frontier model as the planner and the reviewer, the expensive intelligence you consult, and route the high-volume execution to cheaper open models that it supervises. The Frontier model designs the approach, decomposes the work, and grades the output. The open models do the bulk tokens. You get most of the quality at a fraction of the spend, and the ratio improves every time the open models take another step forward, which they reliably do. The operational side of this has gotten easy, too. Routing layers like OpenRouter make it close to trivial to send each request to the right model dynamically, by task, by cost ceiling, or by latency.

30:11So a multi-model portfolio no longer means multi-vendor integration pain. Chapter 9. A Case Study. Coinbase and Five Tactics for Blended Intelligence. The clearest proof point of the half came from Coinbase. In July, CEO Brian Armstrong laid out how the company cut its AI spend nearly in half while token usage kept growing. The playbook reads like the portfolio strategy running in production, and his five tactics are worth studying. Tactic 1. Better defaults, not usage caps. 91 % of Coinbase employees never hit their usage caps. So rather than lowering caps and driving up alerts, the company is moving its defaults to open-weight models, GLM 5.2 and Kimi 2.7, through its own gateway, while engineers stay free to pick the right model for the task.