Super excited about HydraFusion in GitHub Copilot, and what it shows about the shift from model selection to model orchestration. By bringing together multiple models to plan, build, critique, and complete coding tasks, it can deliver outcomes at up to 67% lower cost. It’s a great example of the... value of a heterogeneous model ecosystem, and how we’re continuing to advance the cost-to-outcome frontier.show more

Try it out:

@morfmod has been delivering that for 2+ months now, we just can't launch because we have no funding :(

MSFT now:

So you invented a committee that plans, builds and critiques, except this one is 67% cheaper and nobody asks to circle back.

I spent the first 10 seconds wondering why Satya was wearing a GitHub hoodie..

After you destroyed what little trust people had in GitHub Copilot with your pricing rugpull, I doubt anyone is going to circle back for another look, me included.

The cost pressure is becoming impossible to ignore as coding workloads get more complex Using the expensive model for every step is a luxury orchestration can increasingly eliminate

67% lower cost is a serious gain. The interesting part is how orchestration turns different model strengths into one workflow.

Serious question: does ANYONE use copilot?

costing six seven percent less?

@grok - What is he saying? Summaries everything for me.

model orchestration is gonna be the big theme this year for sure. wild how fast we moved from picking one model to mixing a whole team of them.

Almost everything is 67% cheaper than SlOpus 5. Compare it against Luna or Sol :)

The 67% matters less than who controls the routing layer. One model: selection Multiple models: orchestration Next: automatic routing If the best model changes by task, model selection becomes infrastructure rather than a user decision.

Model orchestration is a big efficiency gain, but coordinating models is only half the problem. When multiple agents build on the same context, how do we ensure that context is consistent, traceable, and grounded in evidence?

60% huge

Can't just be me who did a double take of whether that was actually Satya or @mariorod1 in the video!

“67% lower cost” because you stopped sending every subtask to the same expensive model and added routing, critique and retries. Now show the baseline, pass rate, latency and retry budget before “cost-to-outcome frontier” becomes another dashboard noun.

model picking quietly becoming a solved problem, wild to watch

Right now copilot has Auto option that gives only 10% discount

Is this your ai Avatar in the Video?

That's a neat idea. I'm curious to see how orchestrating models will change things.

HydraFusion using specialized models to critique Copilot code tasks finally solves the "stuck in the middle" problem of agentic coding. Seeing GitHub treat orchestration as a feature, not just an architectural choice, is the right path to real-world scale.

I first thought he was AI generated Satya.

The numbers back this up four models neck and neck on openrouter this week, openai, zhipu, deepseek and tencent weekly rankings here

Sounds interesting. I think more harnesses will adopt this methodology.

I asked Satya for a model yesterday and he dropped one today. Cool

Honestly, I’d rather the system figure out which model should do what.

🔹 @AiVerifai 's Heterogeneous MultiLLM agent achieves 92.0% APR (Agent Passing Rate), surpassing the state-of-the-art RTL agent ACE-RTL (88.9%) across all categories. 🔹 Rather than relying on one model for all tasks, we assign different LLMs to specialized roles based on their individual strengths

Orchestration only beats a single strong model if the router doesn't burn the savings on redundant calls. 67% lower than what baseline — GPT-4-class single model, or the most expensive model in the pool doing everything? That denominator is the whole claim.

My Codex setup works like this now: one home base, with specialist models routed in for the jobs they are best at. I keep the operating context while the model changes underneath it.

orchastration lmao so we just adding more layers of shit to the same broken stack