latent
Every year in AI feels like a decade, but 2025 had a different smell to it. The conversation finally stopped being "look what it can do" and turned into "how do we ship this without it falling over in production." That shift, more than any single model release, is what I'll remember the year for.
Here's roughly how it played out.
January-February: o1 ships, then DeepSeek lands
OpenAI started the year by pushing o1 to general availability. After months in preview, the full model showed up with extended thinking and beat GPT-4 on math, coding, and anything that needed multi-step logic. The idea behind it isn't complicated: let the model spend more compute at inference time and you get better answers. Not a little better. A lot better.
Then DeepSeek dropped R1. A Chinese lab, working under US chip export restrictions, published a reasoning model that matched o1 on most benchmarks and claimed a training cost around $6 million. The whole industry did a double take. If frontier reasoning could be done at that price, the moat around the closed-model providers was a lot shallower than anyone wanted to admit.
R1 also reminded everyone how fast the open-source community moves once weights are out. Within weeks there were quantized versions running on consumer GPUs and distilled variants based on Qwen and Llama tuned for narrower tasks on Hugging Face.
March-April: Gemini 2.0 and tool use that works
Google shipped Gemini 2.0 with Flash and Pro variants and finally delivered on the multimodal pitch they'd been making since 1.5. Native image, audio, and video in one model. The bigger deal for me was that Gemini 2.0 Flash hit a price-performance point that made it the obvious default for high-volume production workloads.
Tool use is what really changed things. Gemini 2.0 models could call functions, search the web, and execute code in chained steps without falling apart. Google wired this into Vertex AI with managed RAG pipelines and the gap between "demo" and "production" got noticeably smaller.
Anthropic released Claude 3.5 Opus and upgraded Sonnet around the same time, leaning hard into coding and instruction-following. The model quality race had stopped being about who had the nicest chat experience. It was about which model could run a multi-step workflow with tools and not lose the plot halfway through.
May-June: agents stop being a meme
Computer-use agents moved from research demos to actual products. Anthropic shipped computer use that let Claude drive desktop apps, browsers, and terminal sessions. OpenAI followed with their own agent framework. What changed wasn't that agents could click buttons. It was that they could click buttons reliably enough for enterprise pilots to take them seriously.
Companies started deploying agents for internal workflows: filing expense reports, walking through ancient internal tools, running QA suites. The results were mixed but the failure modes told you everything. Agents with narrow, well-defined tasks and clear success criteria worked. Agents given vague goals in open-ended environments still flailed.
Coding agents also stopped being a punchline this period. Devin, SWE-agent, Claude Code — they went from "impressive demo" to something a growing number of developers actually used every day. The pattern was always the same: AI does the boilerplate, the human owns the architecture.
July-August: the cost collapse
Inference costs fell off a cliff. GPT-4-class performance that ran $30 per million input tokens in early 2024 was under $1 by mid-2025. Several things converged at once.
NVIDIA H200 deployments reached scale and AMD MI300X turned out to be genuinely competitive for inference. Speculative decoding, better quantization, and mixture-of-experts architectures cut compute per token. And competition from DeepSeek, Mistral, Meta, and others meant the major providers had nowhere to hide on pricing.
That collapse rewrote the economics. Use cases that were too expensive to consider in 2023 — real-time AI on customer support calls, continuous code review, always-on document analysis — became viable. The constraint stopped being "can we afford the API calls" and became "can we handle the integration complexity."
September-October: open source catches up
Llama 4 landed with a mixture-of-experts architecture that punched well above its parameter count. Meta kept releasing weights freely and the ecosystem responded — fine-tuned variants for medical, legal, and financial domains showed up within weeks.
Mistral shipped several models that competed head-to-head with closed-source offerings on practical benchmarks, and their focus on European enterprise compliance gave them a market position nobody else really had.
The open-source reasoning space exploded too. Building on DeepSeek's work, multiple groups published training recipes for reasoning capabilities. By Q4, any org with moderate GPU resources could train or fine-tune a reasoning model on its own data.
# The open-source stack by late 2025
Model weights: Llama 4, DeepSeek R1, Mistral Large, Qwen 2.5
Inference: vLLM, TGI, llama.cpp, Ollama
Fine-tuning: Axolotl, Unsloth, TRL
Evaluation: lm-eval-harness, HELM, custom evals
Orchestration: LangGraph, CrewAI, AutoGen
November-December: the deployment reckoning
The last quarter was the reality check. A lot of orgs that had enthusiastically adopted AI in H1 hit the deployment wall in H2. The models worked fine. The integrations were brittle. Evaluation was hard. Monitoring was harder.
A few lessons settled into the industry's collective memory.
Evals matter more than benchmarks. Teams that built custom evaluation suites for their specific use cases shipped. Teams that treated public benchmarks as a proxy for production quality didn't.
Guardrails aren't optional. As AI systems touched more production workflows, the cost of failure went up. Output validation, fallback paths, and human-in-the-loop checkpoints turned into standard architecture, not nice-to-haves.
The bottleneck is integration, not intelligence. Most teams that failed didn't fail because the model wasn't smart enough. They failed because connecting the model to existing systems, data pipelines, and business logic is genuinely hard engineering work, and it doesn't go away just because you wrote a clever prompt.
The narrative shift
The most important change in 2025 wasn't any single model or product. It was the way the industry talked about AI.
In 2024, the conversation was dominated by capability demos. Can it pass the bar exam? Can it write a novel? Can it solve PhD-level math? Those questions made for great headlines and never translated cleanly into business value.
In 2025 the questions changed. How do we evaluate this in production? What's the fallback when the model fails? How do we version our prompts? What's the total cost of ownership? Less exciting. Much more useful.
Heading into 2026
Closing out 2025, the picture looks pretty different from where we started.
Reasoning is standard. Every frontier model has some flavor of extended thinking. The open question isn't whether models can reason, it's how to allocate that reasoning compute without burning money.
Agents are real but narrow. The fully autonomous AI agent is still aspirational. The task-specific agent with good guardrails is shipping today.
Open source is competitive. The gap between open and closed models is the smallest it has ever been and still shrinking.
Cost isn't the constraint anymore. Inference is cheap. Engineering time, evaluation infrastructure, and organizational readiness are the real bottlenecks.
The year ahead will be defined by what we build on top of all this. The models are good enough. The question is whether we're ready to use them well.