2026 predictions: where the puck is heading

Most year-ahead pieces are vague enough that nothing in them can be graded later. The whole point of writing predictions in public is the grading; without that, the piece is mostly entertainment. This one tries to be specific enough that the next checkpoint (mid-2026, year-end-2026) can actually score it.

The takes below are calibrated against the patterns I missed in 2025 and the real shifts that actually happened. Twelve predictions in three buckets. I'll revisit them at the mid-year and year-end checkpoints.

Models and capability

1. The open-weights premium tier emerges. By end of 2026, an open-weights release will compete head-to-head with Opus 4 / GPT-5 / Gemini 2.5 on the hardest reasoning workloads. DeepSeek-R3 or equivalent. The "open weights are the workhorse tier and closed-frontier owns the premium tier" pattern from 2025 breaks.

2. The closed-frontier shops respond with capability moves, not price moves. Anthropic ships Opus 5 or equivalent; OpenAI ships GPT-5.5 or a new o-series flagship; Google ships a meaningful Gemini 3 release. None of them meaningfully cuts workhorse-tier prices. The premium tier is where the closed frontier defends.

3. Workhorse-tier price compression continues. End-2026 hosted workhorse-tier pricing is 30-50% below end-2025 levels. The DeepSeek-driven floor keeps moving. The hosted-vendor margin on workhorse-tier inference becomes a real squeeze.

4. A meaningfully better small-model (3-8B class) ships. Phi-5 or Gemma 4 or equivalent. The capability lift on the small-model end is at least as large as the workhorse-tier lift. The mass-market deployment surface (phones, browsers, edge) becomes more capable.

Infrastructure and tooling

5. MCP-based memory portability becomes a real category. Either an open-source standard or vendor-led interop ships in 2026 that lets personal AI memory port between vendors. Not perfect; meaningfully present. The lock-in dimension I keep complaining about gets partial relief.

6. The MCP routing layer consolidates around 2-3 winners. The current category fragmentation (MetaMCP, cloud-vendor gateways, custom builds) narrows. By end of 2026, there's a clear leader for open-source and a clear leader for managed-service, with one or two close competitors.

7. Multi-machine MLX training becomes routine. What was emerging-but-experimental at end of 2025 becomes the default for serious personal-AI training work in 2026. The activation energy drops to "a couple of hours" rather than "a meaningful project."

8. A serious AI-governance platform ships. From a hyperscaler or a startup. The governance gap I've been writing about gets a credible attempt at platform-level resolution. Not a complete answer; meaningfully present.

Personal AI and consumer products

9. Apple ships the bridge product. Apple Intelligence's headline missing features (the personal-context Siri, the agent surface, the cross-app workflows) ship in iOS 19 / macOS 26.x. They land imperfectly; they ship. The principled-personal-AI-for-casual-users gap closes meaningfully on the Apple platform.

10. The principled-user community grows by 5-10x. Not large in absolute terms; large enough to support a richer set of tools, tutorials, and small-vendor products targeting the population. The personal-AI hardware buying conversation becomes mainstream-adjacent.

11. A consumer competitor to Apple's bridge emerges. Either from Meta, from a Linux-side project, or from an unexpected entrant. Doesn't beat Apple on consumer reach; provides a credible alternative that keeps Apple from becoming the only path. Healthier set of options all around.

12. The hosted-AI personal-context capabilities improve faster than the privacy story does. Continuation of 2025 pattern. ChatGPT, Claude, Gemini all add more personal-context features; the privacy gap for casual users persists. Most users accept the trade-off; the principled-user population continues to opt out.

What I'm not predicting

A few things I'm declining to commit on because I don't have a reliable signal:

The agent-everywhere consumer-product reality. The keynote demos keep promising it. I'm not predicting whether it actually ships in a way that matters.

Foundation-model architectural breakthroughs. The transformer-plus-X variants keep iterating. I'm not predicting a step-change in the underlying architecture.

Geopolitical AI dynamics. US/China AI competition, EU regulation, sovereign-AI initiatives. These all matter; the timing on specific developments is too uncertain to predict.

Robotics-AI convergence. The category is real and moving; I don't have enough conviction on the 2026 pace to commit to specifics.

A major AI-related company collapse or acquisition. The category is volatile enough that this could happen; predicting which is closer to entertainment than to forecasting.

These are the takes I'm declining to take. Worth being plain about the gaps in my forecasts as much as the forecasts themselves.

How I'd grade myself

For the predictions above, I'm setting a 60% hit rate as the bar for "did acceptably." The directional takes should hit; the specific timing should be approximately right. I'll grade plainly at the mid-year (June 2026) and year-end (December 2026) checkpoints.

The pattern from grading 2025: I was directionally right on most things and specifically off on timing. The 2026 predictions are calibrated to expect the same shape, the directional takes are the high-confidence parts; the specific timing is the lower-confidence overlay.

If the hit rate comes in below 50%, I'll do a meaningful recalibration on whatever pattern produced the misses. If it comes in above 70%. I was probably too conservative. The middle is where calibration is working.

The bigger frame

2026 is the year I expect the personal-AI category to reach the "you're starting to see this in the wild" threshold for casual users. The foundation is built. The bridge product is the binding constraint. The forecasts above are calibrated to the bet that the bridge gets built; if it doesn't, the personal-AI conversation stays niche for another year.

The other big frame is governance. The gap is the same shape it was a year ago. The platform-level fix needs to ship in 2026 or the gap becomes a permanent feature of the AI infrastructure rather than a temporary one. Either way, the conversation evolves.

The model side is the loud part of the conversation; the foundation side is where the durable shifts happen. The predictions above weight the foundation more than the model; same calibration as 2025; same expectation that the foundation matters more.

The next checkpoint is mid-2026. Worth getting these on the record now so the grading later is grounded. Predictions in public, scored in public, recalibrated in public. The discipline that makes the predictions worth making.