What’s new in the AI world — five shifts to watch in mid-2026

The AI landscape moves in quarters now, not in years. The pace of model releases has slowed slightly from 2024’s frenzy, but the ground around the models — the integration patterns, the unit economics, the procurement language, the regulatory enforcement — is shifting faster than ever. Here are the five shifts we have watched land across our client base over the last three months, what each one changes for buyers, and where the next quarter’s decisions actually matter.

1. Foundation-model vendors are now consulting firms

OpenAI’s acquisition of tomoro.ai earlier this month was the most visible signal in a pattern that has been building all year. Anthropic, Google, and AWS are building or buying delivery capacity to sit next to their API access. The reasoning is structural: foundation models are a capability without a workflow, and the gap between API access and shipped business outcome is where most enterprise AI projects still die. The vendors have noticed, and they are not going to leave that revenue to independent partners forever.

For buyers, this changes the procurement landscape more than the technology landscape. The “one throat to choke” platform-and-consulting bundle is now a real option from each major vendor. The trade-off is commercial alignment: a consulting team owned by a model vendor will, structurally, recommend solutions built on that vendor. Worth taking deliberately, not by default.

2. Coding agents are infrastructure now, not a tool

Eighteen months ago, AI coding assistants were a tool an engineer chose to use. Today they are part of the engineering stack — sitting alongside source control, CI, and the language runtime. The teams pulling ahead are not the ones who adopted them first; they are the ones who have integrated them into the operating model. Pull-request ownership policy, dependency review automation, eval suites for AI-generated code: this is the unglamorous work that separates a 60% throughput gain from a six-month review-burden crisis.

The most consequential shift inside this is governance. Enterprise tiers from the major vendors now offer zero-retention configurations, contractual exclusion from training data, and SOC 2 reporting for code processed. Companies still on consumer tiers in mid-2026 are accepting a meaningfully different risk profile than peers on enterprise. The price gap is small. The exposure gap is not.

3. The interop layer finally matured

Model Context Protocol, originally a single-vendor initiative, has quietly become the de-facto standard for connecting agents to tools and data across the major model providers. The practical effect on our work has been significant: integrations that used to require custom SDK glue for each provider now reuse a single tool definition across Anthropic, OpenAI, and several open-weight runtimes. The portability story for AI investments is finally credible.

This is also where most of the integration value sits for the next twelve months. The interesting AI work in 2026 is not the model layer — that gap has narrowed sharply between providers — it is the wiring between the agent and your business systems. The companies winning this quarter are the ones whose AI can read from the CRM, write to the ERP, and pass clean audit logs back to compliance, all over a standardised interface. If you are weighing where to spend the AI budget right now, the integration layer beats the model upgrade nine times out of ten. Cravings AI Integration engagements live at exactly this seam — RAG over your real corpus, copilots inside your existing tools, model-routed pipelines into your CRM and finance stack — with the auditability the compliance team is going to ask about in 2027.

4. Cost is no longer the bottleneck. Orchestration is.

Inference cost per token has fallen another 60–75% in the last twelve months across both the frontier and the middle tiers. For most of our 2025 builds, the model bill was the largest line item. For most of our 2026 builds, it is the third or fourth — behind engineering time, evaluation tooling, and observability infrastructure.

This has rewritten the unit economics conversation. The interesting cost question is no longer “what does this model cost per million tokens” — it is “what does it cost to run a routed fleet of models where the cheap fast one handles the easy 80% and the expensive smart one handles the genuinely ambiguous decisions.” Teams who have figured out per-call routing are running production AI at 30–60% of the cost of teams who have not, with no quality penalty.

The corollary: companies still benchmarking AI projects against a single-model cost curve from 2024 are working from an outdated business case. The numbers have moved. The proposals you got last year are worth revisiting.

5. Regulation is finally shaping buyer behaviour

The EU AI Act’s high-risk obligations have moved from announcement to enforcement. We have watched at least four clients have to rebuild parts of their AI workflows this quarter to meet the documentation, transparency, and human-oversight requirements. None of it was technically difficult; all of it required someone senior to own the response, and most companies do not have that person yet.

For buyers in regulated industries — financial services, healthcare, hiring, education — the practical implication is that the AI strategy document and the compliance posture are now the same document. They cannot be written by separate teams in separate quarters. The companies handling this best are the ones who put compliance review at the front of every AI engagement, not the back, and who have an honest map of where AI touches a regulated workflow on their roadmap.

The honest counter-argument

It is fair to ask whether any of this changes the fundamental advice we have been giving for the last two years. Mostly, it does not. Write the eval set before the prompt. Keep a human in the loop on anything that touches a system of record. Treat the runbook as a first-class deliverable. Hire for the on-call rotation, not the demo. The principles are the same. What changes quarter to quarter is which specific platform decisions have just become cheaper, riskier, or unavoidable.

If you are looking for the one thing that has actually shifted the strategic question this quarter, it is the foundation-model-vendors-as-consulting-firms move. Everything else is a continuation of trends already in motion. The vendor consolidation is a step-change.

What to do in the next 30 days

• Re-cost your in-flight AI proposals. The model-cost line item from twelve months ago is now wrong by a meaningful margin, in your favour. Whether the project gets bigger or the scope tightens, the numbers are worth refreshing.
• Audit your AI vendor contracts for portability. The interop layer is mature enough that switching is now a real option. Renewing without negotiating portability is paying for lock-in that no longer needs to exist.
• Pick which side of the platform-aligned vs independent question you are on. The vendors are forcing a decision by acquiring the alternative. Make it deliberately, document it in the AI strategy doc, and revisit it annually.
• Get the compliance officer into the AI standup. The companies still treating compliance review as a quarterly check-in are about to discover what the companies who treated it as a build-time requirement already know.
• Stop benchmarking against last year’s model. The mid-2025 baseline is not the right comparison for a mid-2026 system. Some projects that did not pencil out last year do now. Some that did then do not.

The next three months will probably bring another shift we have not flagged here. They usually do. What does not move is the discipline around shipping AI into real businesses — the readiness audit before the build, the evals before the prompt, the team owning the system after the launch. That part still beats the model upgrade, the framework switch, and the vendor consolidation, every quarter.

If your AI roadmap was scoped against a mid-2025 view of the landscape, the next thirty minutes is the cheapest time to test whether it still holds. A Cravings readiness conversation maps your in-flight projects against the new reality — model costs, interop options, regulatory exposure, and where the highest-leverage moves are this quarter. No platform lock-in built into anything we recommend.