Predicting 2026: LLM/Agent Consolidation

August is coming to a close and we have 2/3 of 2026 finished. The last couple of weeks I have felt that the AI hypecycle reached a plateau in mid-2026 and we might be able to get a glimpse of how 2026 will turn out over the last four months.

My prediction: by the end of 2026 frontier-model subscriptions become optional for agentic coding — good-enough open models plus a well-tuned harness will be the default, and the next step is self-hosting.

We went through rapid model evolution in 2025 and by the end of the year the harnesses caught up to make use of the more powerful models.
When both forces combined (and models were created with agentic use cases in mind) we saw real explosion in agentic uses.
The first agent runtimes (like OpenClaw) which picked up tasks on a schedule or by an external trigger entered the scene early this year and it went bonkers for a while - everyone aimed to gain marketshare and subsidize their models. That helped

• to flex in front of investors how adoption was going hockey-stick, so they would pour more money in
• to lock users into their harnesses (Codex, Claude Code) and hoover up real-world usage data to train better models.

Price Hikes and Consequences

And here we are - (frontier) inference providers raise prices left and right. I used Anthropic’s Max5 subscription during the spring and even upgraded to Max20 for a couple of months. It used to be below $200/month - now the monthly price is 299€.
I used Opus 4.x and briefly Fable 5 but it does not feel that Fable was getting better results.

In June I read a blog post by Addy Osmani The New Software Lifecycle which quotes a paper by Google folks The New SDLC With Vibe Coding. It mentioned that 10% of agentic coding is the model, 90% is everything else - what’s called the harness.

The 10% Model

That made sense to me at the time and I explored how I can tweak my harness to my liking. It’s not perfect yet and might not be while we iterate in 2026, but what’s clear is that we can substitute frontier models with good-enough models if we have a good enough harness to power the agents.

Provider Switch

Last week my Max20 subscription expired and I dropped to Anthropic’s lowest tier (Pro - 22€/month) subscription to not go cold-turkey directly - while switching my day-to-day inference in OpenCode to OpenRouter.ai.
And I am happy to announce that my agentic coding is doing fine. My workflow using OpenCode, OpenSpec and git worktrees within a coding session directory is just fine. The first morning using OpenRouter models got the task done and did not feel substantially slower. I spent $10 on OpenRouter - mainly due to the use of GLM-5.2, which is rather expensive. I switched to deepseek-v4-pro and glm-5.3-flash (which had started out as a free preview called Ox Alpha). Mid-week glm-5.3-flash was released and I spent less than a dollar for the whole day.

One asterisk: OpenRouter was bought by Stripe a while back — not sure yet what that means long-term for their model catalogue and pricing.

iPhone X Moment

That reminded me of the times when the urge to update to a new(er) iPhone faded away because the older iPhones were still good enough. The camera was not improved much for the “point and shoot” use case, the battery was still good for a day and the other new features were no game-changers.
With AI models it seems to be the same now. Using deepseek-v4-pro and recently more glm-5.3-flash convinced me that I can stop my Anthropic subscription next month and use open models more and more. Out of convenience I still have an OpenAI subscription - I might use that for frontier models if the OpenRouter models are stuck.

Next stop: self-hosted

I suppose the next frontier for this journey is to host models locally. I got 128GB of StrixHalo VRAM. Not super fast, but for agents running in the background and over night I reckon that might be even enough.
A routine task to check my emails, aggregate news from my (TT)RSS reader or prepare a morning briefing to kick-off a day has all night to finish.

This week Xiaomi revealed a prototype of their AI Cube (me.mashable.com blog post). 1.2TB/s memory bandwidth (similar to the new $10k+ Mac Studio). The StrixHalo has 256GB/s - if that’s priced aggressively and/or the memory prices go down in 2027 self-hosted is on!

2026 AI caveat

Even though I think it becomes clearer, who knows if I need to eat my words by Christmas. But as I argued in this post: we are getting towards a more stable 2026 (at least in AI land).

I’ll report back in December how well this blog post aged and whether the open-model setup held, and if the StrixHalo got to earn its keep over night.