My main bet for 2026 is that most of the visible progress will come from better tooling and better harnesses, not models that keep learning new facts on their own after release. People keep asking for continual learning, but true in-model updates are both technically hard and product-hostile. Reliability is the whole point. If the model’s behavior drifts day to day, you cannot build workflows on top of it, and you cannot debug anything. People already complain with every GPT update, even though we only get one every couple of months. If the model were constantly changing, all of the people who make rumors about models performing worse than they used to would have a field day. It would be crazy, so that’s definitely not going to happen for a long time.
Continual learning will mostly mean saved skills, files, and retrieval
The practical version of continual learning is already here: you save templates, decision logs, rubrics, preference files, and domain notes, and you feed them back through context in a disciplined way. Call it memory, call it context management, call it a folder full of markdown. That is what scales in 2026. We might get some experimental in-model continual learning, but I don’t think the actual model itself inherently getting smarter over time is going to happen on a large scale in 2026. This is because that wouldn’t even make for a good product. People want a model to be reliable, not acting one way and then acting totally differently the next week, even if it is usually better overall.
Agents keep improving, but the winning layer is the interface
2026 is going to be about agents that fit into your tools without friction. The strongest agent setups are not just a model with tools. They are a workflow that can observe state, take actions, verify results, and recover when something goes wrong. Opus 4.5 is already world-changing in its capabilities, and Claude Code is an incredibly capable and general harness. However, I see two main trends coming in 2026: More user-friendly interfaces for agents that are at least as powerful and more domain-specialized agents. Claude Code is great for coding and functions as a general harness for many other things. But a specialized agent that is good at composing music would require a different set of primitives, even if you could theoretically do the same thing with a general framework.
I wrote about the wrapper layer and why it still matters here: Meta’s $2-3 B Manus Bet: Why the Wrapper Still Matters. That same idea applies to agent products. The interface is not fluff. It is the product.
IDE integration beats living in a terminal
I expect Claude Code style workflows to keep growing, but I do not think the terminal is the end state. IDEs contain the context you want: code graph, open files, diagnostics, diffs, test output, and project structure. A terminal-only view forces the user to manually bridge gaps that the IDE already knows. Usually, when I am using the terminal, I am using the terminal built into the IDE. In my case, I will be using Antigravity as the IDE and then using the terminal within Antigravity with Claude Code inside it. This allows me to look at all kinds of relevant data and get the benefits of an IDE while still leveraging the strengths of Claude Code. For more on the evolution of these tools, check out GPT-5.2-Codex: Better Long-Horizon Agentic Coding.
Focus priority ranking for 2026 AI development.
The Model Landscape: Llama 5 and Naming
I think we’ll finally see Llama 5, GPT-6, Gemini 3.5, and at least Claude 5. Most of these will likely arrive in the first half of 2026, with the exception of Gemini. Gemini will probably take longer since their gaps between major updates tend to be longer. I do think Llama 5 will represent Meta getting back into the game. They really dropped the ball hard with Llama 4 and they just can’t afford to do that again. Consequently, I think they will refuse to put anything out until it’s actually viable. Labs have started doing more numeric version bumps instead of date-string soup, and that trend should continue in 2026. It is still messy, and if you want my full rant on naming, it is here: OpenAI’s Naming Nightmare: From GPT-1 to GPT-5.2 Confusion.
Benchmarks and METR Time Horizons
Most common benchmarks are nearing saturation. ARC-AGI has a better chance of staying relevant because real-time constraints break a lot of the usual benchmark hacks. All of the optimization labs have been doing for long-running reasoning to improve at benchmarks simply doesn’t work when you have to run it very quickly. I have a deeper post on ARC-AGI-2 and harness effects here: ARC-AGI-2 2025: From Sub-10% to 75%+ with the Poetiq Harness on GPT-5.2. The METR time horizon benchmark scales well, and I see that going into 2026 and being a major part of some model releases, showing the length of tasks that these models can reliably complete. This progress will go up at least to a certain point, because it gets harder to train for longer-running tasks once you get past something like 24 hours.
Media, World Models, and Cost
I expect another OpenAI omnimodal bidirectional model, though I hope they move away from names like 4o because speech-to-text always messes it up. Media generation will continue to diversify with more points along the cost-quality curve. Beyond images, we will see much better video generation and especially video editing models. I expect we’ll see a moment where we get a cheap, fast, and pretty good video editing model that goes viral. If you want a low-cost angle on video, see: Seedance 1.5: A Low Cost Angle on AI Video Generation. Regarding cost, I do not think we will see significant reductions in the cost of frontier models. The demand for the best intelligence you can get within reason will continue to remain in the $10 to $30 per million output token range for the kinds of models that people use regularly. While intelligence at this price point will rise, overall costs for the top-tier models are not expected to drop significantly. If you want to see how we got here, read: 2025 AI Timeline: The Year Reasoning, Agents, and Video All Hit at Once.
