A 35B model just matched a 1T model on agent tasks

This week, three separate model releases landed that collectively do something the AI narrative has been resisting: they make the frontier-compute moat look smaller than the infrastructure bets assume. Today's issue is the read on what that actually means before your next model selection decision.

In today's issue:

• Main story: The model commodity crash

• Since Friday: Fable 5 returns with restrictions, Microsoft builds a 6,000-person AI deployment division, Anthropic signs a 20-year $19B data center lease, and more

Tencent released a model yesterday that uses 21 billion active parameters and matches the performance of models five times its size. A separate team published results showing a 35-billion-parameter model hitting trillion-parameter benchmarks on agent tasks. A third paper cut reasoning token usage in half. On a single day, the cost floor of frontier AI dropped further than it had in any comparable 24-hour stretch.

The AI industry has been telling itself a story for the past two years: better results require bigger models, bigger models require more compute, more compute requires the kind of capital that only a handful of companies can deploy. That story is still partially true at the discovery layer. It is no longer true at the production layer, and the gap between the two is widening faster than most operators realize.

The standard read of this week's model releases is about research competition: China catching up to American labs, open-source eating into proprietary advantage, benchmark scores trading hands. That framing is accurate as far as it goes.

What it misses is the operational consequence for the companies that are actually trying to build on top of these models. If a 35B parameter model can do what a 1T model does on agent tasks, you do not need a hyperscaler contract to run reliable AI in production. You need a good inference setup and the right model for the task. The moat that the frontier labs have been selling (access to intelligence itself) is thinning out.

Tencent's Hy3 did something that matters more than its benchmark score. Instead of leading with capability, it led with reliability: tool-call recovery, output format consistency, multi-turn constraint tracking, hallucination rate down to 5.4 percent. These are exactly the failure modes that kill production deployments. A model that follows instructions 95 percent of the time is unusable for automation at any meaningful volume. Hy3 is saying: we understand what enterprise buyers actually need, which is not the smartest agent but the most consistent one.

Meanwhile, Microsoft announced a new dedicated AI deployment division, 6,000 people strong, on the same week it was cutting 4,800 jobs from Xbox. Five gaming studios were affected. The $69 billion Activision acquisition was supposed to make Xbox a content powerhouse, and the budget is now flowing toward AI implementation instead. Xbox CEO Asha Sharma called the division "not healthy." When the choice is between entertainment and AI infrastructure, AI infrastructure wins, and it is not close.

This pattern is not unique to Microsoft. Across enterprise technology, headcount reductions in legacy divisions are being re-justified as AI investment. Companies are clearly cutting to fund AI. The more important question is whether the AI they are building with that money will actually work. A 6,000-person deployment unit is at best a bet that scale of implementation will eventually translate into outcomes, and that bet has not been won yet.

Anthropic made a different kind of bet this week. A 20-year, $19 billion lease with TeraWulf for a data center in Kentucky with approximately 400 megawatts of capacity. Twenty years. The infrastructure commitment assumes that AI compute demand will still be growing in 2046. That may well be right. It also means Anthropic is building as if the frontier-compute story is not over, even as the evidence from the model side suggests the gap between frontier and near-frontier is closing faster than anyone expected eighteen months ago.

TEAM READINESS AUDIT

Get your team’s AI readiness score. Leave with an AI readiness brief.

A 15-minute assessment that turns your answers into a real decision artifact: your readiness level, your six-axis shape, where you're strongest, where you're constrained, and what not to build yet.

Counterargument

There is a reasonable counterargument here: smaller models matching larger ones on benchmarks is not the same thing as smaller models replacing larger ones in practice. Benchmarks measure specific tasks. Production systems fail for reasons that benchmarks do not capture: context window edge cases, instruction drift over long sessions, unexpected input distributions. The hyperscalers have teams whose entire job is handling those failure modes. A COO running a 200-person company does not.

My honest answer is that this counterargument was more persuasive a year ago. What's changed is not the complexity of production AI, which is as real as ever. What's changed is that the model reliability story is moving toward the enterprise buyer faster than anyone anticipated. Hy3 is not just performing well on benchmarks; it is specifically engineering for the failure modes that kill production deployments. The question shifted from whether small models can match large ones in theory to whether the tooling and reliability stack around them has caught up, and that gap is closing faster than most operators have noticed.

What this means for you

The pressure to choose a hyperscaler partner and lock in to their AI stack is real, and some of it is legitimate: support, security, compliance, and integration support are genuinely easier with a major platform. But the model layer is no longer a reason to lock in. If a 35B model can do what a 1T model did six months ago, the model you are betting on today may not be the right model for the same task in twelve months.

Watch the reliability metrics more closely than the benchmark scores. The operators I see getting the most out of AI right now are not chasing the most capable models. They are running consistent, narrow deployments where the failure modes are understood and the success criteria are specific. The "best model" question is less useful than the "right tool-call recovery rate for this workflow" question.

The cost story is shifting too. Tesla reportedly capped employee AI spending at the organizational level. Not because AI stopped working, but because inference costs at scale are real even when per-query costs look negligible. If you are in the process of scaling AI across your organization, model-per-token pricing is the number to watch, not the number to set and forget.

From the field

I've been in a few conversations this week with operators who are at the stage I keep seeing: the pilot worked, the team loved it, and now they're trying to figure out how to make it standard practice across the company. The model question comes up every time, and my answer has shifted over the past few months. A year ago I would have said pick the best model for your budget and build around it. Today I say: before you pick a model, figure out what failure looks like in your specific workflow. That question changes the answer faster than any benchmark release does.

The model commodity crash is not a reason to slow down AI adoption. It's a reason to spend less time picking models and more time engineering the workflows they run inside.

SPONSORED BY CLUTCH

Hire secure AI teammates that work 24/7.

Hire pre-built AI teammates. Give your engineers and operators a platform to ship their own AI apps. Stop losing sleep about what is running where.

Clutch is the platform behind both: pre-built agents for the workflows your ops team should automate first, plus the integration plane your team's vibe-coded apps and Claude Code projects plug into. One platform. Real production. Visible and safe by default.

Built for ops, engineering, and security teams that are tired of the shadow-AI surface area inside their own company.

SINCE FRIDAY

• Fable 5 returned globally last week, but with a catch. After a 19-day government-ordered shutdown, Anthropic's most capable model is back with a safety classifier that reroutes flagged prompts to an older model. Coding tasks fall back to Opus 4.8 by default. If your team had workflows built around Fable 5's coding capabilities, check whether they are behaving as expected.

• Microsoft created a 6,000-person AI deployment division. The same week it cut 4,800 Xbox jobs. The first time a major cloud provider has built a professional services arm at this scale, specifically for AI deployment. How well it delivers will matter to mid-market buyers who rely on Microsoft's partner network.

• OpenAI proposed giving the US government a 5% equity stake. The proposal generated significant backlash from engineers and policy observers who argued that government equity in AI labs is a conflict of interest at the regulatory level. The structure has not been finalized.

• Anthropic is building its own custom AI chip. Reporting from The Information says Anthropic has begun work on a custom server chip alongside a 20-year, $19 billion data center lease signed this week. Building proprietary silicon means Anthropic believes it will be running inference at a scale that justifies the cost and timeline of chip design.

• Figure robots are now building cars at BMW's Spartanburg plant. Not in a demo, not in a pilot. Figure's F.03 humanoid robots are operating in commercial production for logistics tasks. CEO Brett Adcock called it the first car in the world built by a humanoid robot.

P.S. If you are currently in the process of selecting models for a production deployment and want a second opinion on the architecture, reply to this email. I read every reply, and this is one of the more consequential decisions you will make in the next six months.

Haroon