Field note

Frontier AI in 2026, what actually changed and what did not

A working note on the shifts in the AI frontier through mid 2026. Long context, agentic capability, open weights, and the parts the headlines got wrong.

By Arihant Deva 6 min read

A year ago the question that animated the field was whether the scaling curve would keep delivering. Today the question is what to build on top of capabilities that were science fiction in 2023. This note is a personal accounting of what changed at the frontier through the first half of 2026, written from the perspective of someone who has been shipping AI-driven systems into a production research stack.

The summary up front: capability went up roughly as expected. The shape of the curve is what moved. Long context, tool use, and agentic behavior compounded faster than anyone publicly modeled in 2024. Open weights closed the gap further than the labs predicted. The cost curve also bent, but the bend went both ways depending on which segment you measure.

Long context became the dominant axis

The most underrated capability shift in the past twelve months is long context. Claude 4.7 ships a 1M token effective window. GPT-5.5 sits in the same neighborhood. Gemini’s long-context tier extends further still, with claims (and corresponding caveats) about multi-million token windows.

What changed is not the headline number. The headline number was achievable a year ago through retrieval gymnastics. What changed is that the model now actually uses the context. The needle-in-haystack benchmarks that were the standard in 2024 looked impressive but did not predict downstream behavior. A model could find a sentence in 1M tokens and still fail to reason across the document.

Through 2026 that gap closed. Modern frontier models can hold a full codebase, a portfolio’s worth of 10-Ks, or a year of legal correspondence in a single window and reason coherently across the whole thing.

The practical implication: retrieval-augmented generation as the dominant architecture for knowledge-intensive tasks is on its way out for any context smaller than roughly 5M tokens. Stuff the documents in, let the model attend. The simplicity gain over a vector database stack is substantial.

This is not universal. RAG is still the right call for truly massive corpora (think the entire SEC EDGAR history). But for the workflows that animated 2023’s RAG boom (a few hundred PDFs, a single company’s filings, a multi-year email thread) the architecture of choice now is “just put it in the prompt.”

Agentic capability passed the demo threshold

The second shift is harder to measure but easier to feel. Models in 2024 could do single-turn tool use credibly. Models in 2026 can hold a goal across dozens of tool calls, recover from errors, replan, and decide when to ask for help.

The benchmark that captured this is SWE-bench Verified, which measures whether a model can resolve real GitHub issues end to end. The state of the art in early 2024 sat near 13 percent. By mid 2026 frontier models with tool harnesses cross 70 percent. That is not a benchmark crawl. That is a regime change.

The downstream effect on engineering work is now legible. Coding agents take real cognitive load off a working engineer. The pattern that emerged across multiple teams I have watched: senior engineers use agents to skip the parts they could have done but did not enjoy (boilerplate, glue code, scaffolding, repetitive refactors), and spend their saved time on the parts that still require judgment (architecture, debugging across systems, performance work).

The mistake I see junior engineers make is letting the agent skip parts they have not yet learned to do. The skill atrophies. Then the agent fails on a non-standard case and the engineer has no reference for what the right answer looks like. The correct frame is agents as a force multiplier on existing competence, not as a replacement for the learning curve.

Anthropic’s Building Effective Agents post from late 2024 is still the cleanest writeup of the production patterns that hold up. The patterns it identified (orchestrator-worker, evaluator-optimizer, routing, prompt chaining) all survived the capability bump.

Open weights closed the gap further than expected

The forecasts in 2024 that the open-weight ecosystem would stay 12-18 months behind the frontier did not survive contact with reality. By mid 2026 there are at least four open-weight models within striking distance of GPT-5.5 on MMLU-Pro and HumanEval. Qwen 3, Llama 4.1, Gemma 3 27B, and Mistral’s latest tier all sit close enough that for most production workloads the choice is now about latency, cost, and trust, not raw capability.

The qualifier matters. “Striking distance” does not mean “matched.” The frontier labs are still ahead on the hardest reasoning and tool-use benchmarks. They are pulling further ahead on multimodal capability. But for the median production workload, the difference between Claude 4.7 and a well-tuned Qwen 3 235B inference is no longer the deciding factor.

The shift this enabled is routing. A serious production stack in 2026 routes calls across a portfolio of models:

  • Cheap open-weight calls for high-volume, well-bounded subtasks.
  • Frontier model calls for the steps that require judgment.
  • Specialized fine-tunes for narrow domains where you can amortize the training cost.

The teams winning on unit economics have built this routing layer. The teams paying frontier prices for every call are paying for capability they are not using.

What did not change

Three things the discourse expected to change in 2026 mostly did not:

Hallucination is still a structural problem. Frontier models hallucinate less than they did, but they still hallucinate. Any production system that depends on factual accuracy still needs grounding and verification layers. There is no model whose unstructured output you can ship without checking.

Training cost kept rising. Public estimates from Epoch AI put frontier training runs above 500 million USD. The cost curve at the very top is rising, not falling. The cost curve for inference fell dramatically, which is why open weights closed the gap. But the cost of being a frontier lab keeps climbing.

Evaluation is still the bottleneck for serious deployment. Capability outpaced the field’s ability to evaluate it. Most teams shipping AI-driven systems are flying with primitive evaluation harnesses. The ones with real evaluation infrastructure ship faster and break less. This was true in 2024 and is more true now.

What I am betting on for the second half of 2026

A few directional bets:

  • Multimodal becomes the default expectation. Text-only is going to feel limited within twelve months. Vision, audio, and structured data all become first-class.
  • Agent infrastructure becomes a recognized layer of the stack. The way databases and observability are now recognized infrastructure categories, agent orchestration will be. Today it is a wild west of one-off harnesses.
  • The “AI does my job” narrative collapses into something more nuanced. What actually happens is that AI does the boring parts of a lot of jobs, the interesting parts get more interesting, and the jobs that were mostly boring evaporate. This is a worse story than either the doomers or the boosters want to tell. It is what happens.
  • Equity research, my own corner of finance, gets restructured around capability rather than headcount. A small team with agents now does what a midsize team did in 2023. The midsize teams that do not adapt do not have a good 2027.

The bet that anchors the rest: the frontier keeps moving, the gap between demo and production stays wide, and the value accrues to the teams that build infrastructure for using the capability rather than the teams that wait for the capability to arrive in a form that solves their problem directly.

We will see how it ages. The honest version of any forecast in this field has a half-life of about six months.