I know, we’re not out of 2025 yet, but I wanted to get this out before it drowns in the end-of-year static, so here we go.
A funny thing happened in 2025. The hype kept screaming, yet the real story was quieter and more consequential. Costs fell. Enterprise adoption patterns matured. “Reasoning” moved from a slide-deck buzzword to visible product behavior. Video generation turned a creative niche into daily tooling. Home devices, PCs, and phones quietly refit themselves around on-device and near-device intelligence. Regulators stopped only talking and began enforcing timelines with teeth. Robotics edged out of demo purgatory. All of that matters more than any single model leaderboard.
If you only remember five AI-themes from 2025, remember these.
1. The price-performance curve bent hard. The Stanford AI Index quantified a collapse in inference cost and a narrowing gap between open-weight and closed models. The report tracked more than a 280-fold decline to reach GPT-3.5-level performance since late 2022 and noted annual efficiency gains at the hardware level, plus a shrinking performance gap on some benchmarks to within a few percentage points. Translation: what felt elite in 2023 felt commodity in 2025, and that unlocked new use cases across the stack.
2. “Reasoning” became a product setting, not a research promise. OpenAI launched GPT-5 in August, billed as a unified system that chooses between fast replies and longer chain-of-thought style analysis. Anthropic shipped Claude 3.7 Sonnet with “hybrid reasoning,” letting users and developers dial how long it thinks. Google rolled out the Gemini 2.0 family with a “Flash Thinking” line that makes the deliberate mode explicit in apps. Even xAI framed Grok 3 around reasoning agents. You could watch models spend extra compute to plan, critique, and revise, then stop when done. This was the year deliberate inference crossed into mainstream interfaces.
3. Generative video matured fast. OpenAI’s Sora 2 shipped late September with better physics, synced dialogue, and a consumer app that added reusable characters and clip stitching. Google’s Veo 3 moved to general availability, picked up vertical formats and 1080p support, cut per-second pricing, and brought image-to-video and richer audio into mainstream workflows through Gemini, Flow, Vids, and the Gemini API. Brands stopped asking “is it real” and started asking “is it on brand and on budget.”
4. AI moved into devices you live with. Google swapped the old Assistant for Gemini in Home speakers and displays and pushed a redesigned Home app built for a more contextual AI. Apple expanded “Apple Intelligence” across iPhone, iPad, and Mac with system-level writing tools, live translation, and visual understanding. Windows marched Copilot+ PCs forward, re-releasing the controversial Recall feature with stronger privacy gates and shipping more AI-assisted search and on-screen actions. The industry experimented in public, sometimes clumsily, and shipped anyway.
5. The hardware floor rose. Nvidia’s Blackwell platform began rolling into real clusters. GB200 and Blackwell Ultra set the tempo for 2025 deliveries and into 2026 pipelines, and analysts projected Blackwell to dominate Nvidia’s high-end shipments for the year. Cheaper tokens plus faster interconnects meant more teams could afford deliberate reasoning and long-context workloads without lighting money on fire.
Let’s unpack the year in more detail, then look ahead to the first quarter of 2026.
Models and modalities: Reasoning, agents, and the video moment
Reasoning as a dial. In August, OpenAI introduced GPT-5 as a “unified” system that knows when to sprint and when to think longer. That is not marketing fluff. It reflects a tactical engineering shift where post-training and inference-time methods let you trade latency for rigor. The visible “think longer” modes across vendors are downstream of that shift. Anthropic’s Claude 3.7 made this explicit with a control for how long the model should think, plus a visible extended-thinking mode so users see the tradeoff. Google’s Gemini 2.0 Flash Thinking models brought similar ideas to consumer apps and the developer platform. xAI launched Grok 3 around the same narrative of heavier test-time compute for better planning. Even if the claims sometimes outpaced benchmarks, the direction landed in products people paid to use.
Agentic coding and “computer use.” Anthropic paired 3.7 with Claude Code and agentic workflows that draft, run, and revise code with tighter loops. That echoes the broader 2024–2025 push toward models that operate tools and apps on your behalf. The line between a chatbot and a generalist operator blurred, which is why enterprises spent more energy on guardrails, approvals, and observability this year than on raw model swaps.
Video gets real. Sora 2 mattered for two reasons. First, quality. Better physics, synced dialogue, and audio cues moved output beyond “clever clip” into storyboards and commercial pre-viz. Second, workflow. The app features signaled a real platform: reusable “character cameos,” clip stitching, and discovery. Google’s Veo response was not just competitive in quality. It was distribution and price. Veo 3 hit general availability in Vertex AI, the Gemini app, and Google Vids with vertical formats and a steep price cut, plus image-to-video and richer audio through Flow. Creative teams began asking for “one pass in Veo, then refine in Sora” or the reverse. Short-form teams started plugging Veo into YouTube Shorts pilots and social stacks. This was the year video became another model to call, not a special project.
Robotics inches toward deployment. Google introduced Gemini Robotics and Robotics-ER on top of Gemini 2.0 to generalize perception and control across platforms. The funding tide followed, with Apptronik raising a sizable round to scale Apollo. The claim is straightforward. If foundation models can plan and perceive in software, they can also help robots adapt faster in messy factories and warehouses. 2025 gave that claim real backing and roadmaps.
Platforms and devices: The home, the phone, and the PC stop waiting
Google’s Home pivot. In October, Google said the quiet part out loud. The “next era of Google Home” is Gemini-native. That means a more conversational layer over scenes, sensors, and routines, which is what users actually want once novelty wears off. It also means Google can ship features on its own cadence without the baggage of the old Assistant stack.
Apple’s system-wide approach. Apple extended Apple Intelligence across iOS, iPadOS, and macOS with default-on settings for supported devices. Live Translation, Visual Intelligence, and writer’s tools are integrated into the system shell rather than bolting on as a separate bot. Whether you like or dislike the creative aesthetics of Apple’s image tools, the integration story was the news. You did not need a separate app to benefit from AI. It was in the text field you were already using.
Windows and Copilot+ PCs. Microsoft finally re-shipped Recall with additional security controls and kept adding AI-assisted search and on-screen actions. The real pattern was not one feature. It was the movement of AI workloads onto NPUs and closer to the screen, with tighter privacy defaults after last year’s backlash. Enterprises have noticed that privacy posture is the gating factor for many deployments.
Hardware: Blackwell sets the pace
AI in 2025 was not only about smarter models. It was about having the compute to let those models reason, browse, plan, and generate longer outputs without bankrupting the project. Nvidia’s Blackwell platform defined that conversation. The GB200 Grace Blackwell Superchip and the Blackwell Ultra “AI factory” systems gave vendors a target to build around, with partner availability set for the back half of 2025 and into 2026. Analysts projected Blackwell to carry more than 80 percent of Nvidia’s high-end shipments for the year. If you felt like context windows got longer and “think” modes got more common, this is why. The interconnects and the throughput finally made it tolerable to ship those modes in production.
Enterprise adoption: From pilots to portfolio
McKinsey’s 2025 State of AI report captured the vibe well. Agentic AI spread, but the winners were the ones who paired that with disciplined human-in-the-loop practices and operating-model changes. The trick was not “pick the best model” but “instrument the workflow, define when humans must validate, measure value, and give the AI a budget.” That is less flashy than a demo, yet it is what moved the needle. Developer sentiment reflected the messiness. Stack Overflow’s 2025 survey showed usage rising but positive sentiment dipping, which is exactly what you would expect when the easy wins are done and teams start wrestling with edge cases and governance.
Meanwhile, Google flipped an “AI mode” into U.S. Search with a full conversational layer, plus paid tiers for deeper creative tools and heavy Gemini features. Whether you love or hate AI in search, that rollout was a line in the sand for product integration at one of the web’s choke points.
Policy and governance: The guardrails finally showed up with dates
The EU AI Act stopped being a future abstract and started being a calendar. As of February 2, 2025, unacceptable-risk systems were prohibited. On August 2, 2025, core rules for general-purpose models took effect. Most remaining obligations begin on August 2, 2026, but the direction is locked. If you are shipping models in Europe, 2025 was the year your compliance roadmap ceased to be optional.
In the U.S., 2025 was noisy. The White House released “America’s AI Action Plan” in July. The Senate removed a proposed long federal preemption of state AI rules from a megabill. Congress passed a landmark “Take It Down Act” aimed at deepfake porn harms with swift takedown requirements. A separate “sandbox” bill seeking long regulatory waivers for AI firms drew criticism. None of this resolved the national framework question, but the takeaway is simple. The U.S. started passing real laws that bite in specific harm areas while the bigger AI governance debate stayed live.
The U.K.’s AI Safety Institute kept publishing evaluation work and coordinated the International AI Safety Report process, which released a “first key update” in October that echoed the industry shift from bigger pretraining to stronger post-training, longer-horizon reasoning, and agent safeguards. You can argue about the framing, but no one can argue the influence on how vendors talk about risk management today.
Science, climate, and robotics: Markers that mattered
A few 2025 markers stand out because they measure real-world impact.
* Protein design and bio. Recognition and funding continued for AI-driven protein design and prediction. The broader arc from AlphaFold to design tools kept paying off, underscored by awards and grants that made their way into mainstream coverage. The through-line is not a single paper. It is the institutionalization of AI in wet labs.
* Weather and Earth systems. Researchers publicized AI-first forecasting pipelines claiming speed and cost advantages over traditional stacks. Whether a specific named system ends up as the standard or not, 2025 validated a simple idea. AI can now stand up specialized forecasts with leaner compute, which is very relevant for developing regions.
* Humanoids and logistics. Beyond Google’s robotics models, funding and releases signaled momentum. Apptronik’s raise to scale Apollo framed a concrete set of warehouse and manufacturing tasks. Boston Dynamics kept pushing an electric Atlas forward as a platform rather than a stunt reel. If you care about supply chains, 2025 felt like the year humanoids entered the planning deck for pilot projects rather than only the keynote.
Why 2025 felt different
Put all of that together, and a pattern emerges.
* Costs fell and access widened. The AI Index data captured the reality practitioners felt. More teams could afford to run deliberate reasoning, larger context, and multi-modal jobs. Open-weight models kept closing the gap, which diversified vendor strategies and improved resilience. ([hai.stanford.edu][1])
* Vendors shipped visible tradeoffs. Sliders for “think longer,” toggles for browsing, and explicit agent permissions created rituals that users could understand. That built trust faster than another glossy demo.
* Workflows beat features. The winners were the teams that made AI disappear into the job to be done. Apple’s default-on system features, Google’s home pivot, and Windows’ richer on-device capabilities point to a world where the best AI is the AI you barely notice.
* Policy finally moved from panel to paper. The EU AI Act’s staged obligations became real. U.S. lawmakers passed a targeted deepfake law and batted around much bigger questions. The UK’s institute kept the safety discourse empirical. Compliance stopped being a future problem and became a 2025 program.
What to watch in Q1 2026
Predictions work best when they are boring, falsifiable, and useful. Here is what to expect between January 1 and March 31, 2026, plus how to prepare.
1) Blackwell deployments turn into features, not just press releases. Expect the practical upside of 2025 Blackwell buys to show up as longer contexts, more routine use of “think longer” modes, and more model-parallel creative pipelines in production tools. You will not always see a banner that says “powered by GB200,” but you will notice assistants handle multi-file repos, multi-scene videos, and messy spreadsheets without choking. If your roadmap depends on heavy agents or multi-modal fusion, reserve time in Q1 to re-bench with the new capacity and retune your quality gates.
2) Video gen becomes a creative staple across marketing and product. Sora 2’s character reuse and stitching will drive simple serialized content, tutorials, and pre-viz. Veo 3’s distribution through Vertex AI and Flow will keep pulling video generation into enterprise stacks with better control, audio, and aspect ratios. Expect early pilots with YouTube Shorts integrations to mature. If you run content ops, standardize your brand prompts, character permissioning, music rights, and watermarks before your teams go wild.
3) Home and office assistants consolidate routines. Gemini in Home is set to roll across more regions and devices. Apple Intelligence will continue to pick up everyday duties through system updates. Windows’ AI features will creep further into everyday tasks on Copilot+ hardware. The net effect is a steady shift from “ask a bot” to “everything around me is a bot.” If you support users, expect more tickets about “what exactly did the assistant change” and invest in explainability and activity logs.
4) Enterprise governance goes from policy to playbook. The EU AI Act’s August 2026 date looms, which means Q1 2026 will be a busy quarter for conformity assessments, model cards, data governance, and unacceptable-risk audits. Expect more guidance from the EU AI Office and national regulators, and expect larger buyers to start requiring AI Act-aligned paperwork in RFPs months in advance. If you sell into the EU, start dry-running your documentation set now.
5) Reasoning modes get cheaper and more available across vendors. Anticipate incremental tuning of GPT-5 tiers, more Claude 3.7 availability across clouds, and expanded Gemini 2.0 options in consumer and developer plans. Watch pricing and latency tweaks as providers fight for habit. If you ship apps, build a simple abstraction so you can switch vendors per task without rewiring everything.
6) Robotics proofs of value stack up. Expect more warehouse and factory pilots with Apollo, Figure, and others, powered by generalist perception and planning models like Gemini Robotics. These will be unglamorous tasks. Unloading. Tote picking. Lineside support. That is exactly why they will stick. If you operate logistics, Q1 is a good time to define the one or two workflows where a mobile manipulator could eliminate a safety risk or labor bottleneck.
7) U.S. policy will add bite in narrow harm areas while the grand bargain lingers. Expect more targeted bills and state actions on deepfakes, kids’ safety, and authentication. The national framework fight will continue, but procurement and platform policy will move faster. If your product touches media or identity, plan for stricter provenance and takedown SLAs.
8) Search and shopping continue their AI turn. Google’s “AI mode” released to the U.S. in May, was a signpost. Expect experiments with shopping, video answers, and brand controls to accelerate after the holiday data rolls in. If you run SEO or retail media, you will want fresh tracking of AI answer surfaces, not just ten blue links.
9) Developer sentiment will look paradoxical. Usage stays high. Satisfaction stays mixed. The paradox resolves when teams stop measuring “did it produce code” and start measuring “did the cycle time and defect rate improve for this class of task.” Expect a Q1 wave of case studies that are narrow and boring. That is how this becomes infrastructure.
Practical takeaways for Q1 2026
* Budget for deliberate inference. With Blackwell capacity and efficiency gains, turn “slow think” from an exception into a standard option for tasks that justify it. Wire a budget ceiling per job. Kill over-thinking when marginal gains flatten.
* Treat video like text. Build prompt libraries, review workflows, and rights management for Sora and Veo the same way you would for copy. Your first win is not a viral ad. It is a 30-second internal demo that unblocks a product review.
* Instrument agents. If a model can click for you, it can also mis-click for you. Add approvals, logs, and the ability to replay decisions. Pick two workflows with bounded blast radius to harden your stack before you scale.
* Start your EU AI Act packet. Inventory your models, map use cases to risk tiers, draft model cards, and line up your data governance story. You will need this for buyers long before August 2026.
* Benchmark across vendors, not within a favorite. The gap between open-weight and closed narrowed on several tasks. You will find value in a mixed portfolio. Build simple switches in your code so you can send retrieval to one model, long-form writing to another, and coding to a third.
The short list of 2025 milestones that actually changed the trajectory
* GPT-5 general release in August with a deliberate reasoning posture and unified interaction model. That set a bar for assistants that decide how hard to think.
* Claude 3.7 Sonnet and Claude Code turning hybrid reasoning and agentic coding into an everyday option across Bedrock and Vertex AI.
* Gemini 2.0 “thinking” models and Gemini Robotics, plus a new Home stack that made AI the default in Google’s living-room devices.
* Sora 2 and Veo 3 GA with real workflows, better controls, and prices you can model. This did not just entertain. It shortened storyboards, prototyping, and small-team production.
* Blackwell everywhere that mattered. The hardware finally caught up with the software ambitions and enabled deliberate inference at scale.
* EU AI Act obligations started to bite, and the U.S. passed at least one meaningful harm-focused law while the larger settlement stayed in play.
If 2023 was the year many people discovered chatbots and 2024 was the year we learned how brittle they could be in the wild, then 2025 was the year the industry got serious about tradeoffs, guardrails, and getting value where it counts. The next quarter will not be a fireworks show. It will be a grind. Infrastructure, compliance, procurement, and the invisible improvements that make this tech feel normal.
That is how you know it is working.
The Roine Playground 
Lämna en kommentar