CEO/cofounder @RaylineAI and @WorkshopAI Founded + sold @DataFleets for $100M
Remote
- Tweets 626
- Following 484
- Followers 6,252
- Likes 457
David Valerio Gilmore retweeted
We open-sourced a local gateway for Claude Code.
The goal is simple: keep the Claude Code experience you already know, but use it with the models you want.
With the open-source gateway, you can connect Claude Code to other models using your own provider keys, whether that’s directly with model providers, through something like OpenRouter, or with local models running on your machine.
This is the more configurable version of what Rayline managed gives you out of the box: smarter model selection while preserving the coding-agent workflow you already like.
Rayline managed handles routing + inference for you.
The open-source gateway gives you control and lets you bring your own model connections.
GitHub repo link in thread.
[Example video below: Claude Code running with hybrid cloud + on-device AI. The main agent runs Opus in the cloud and orchestrates an Explore subagent that Rayline Local routes to a model running fully on device (Qwen3.6-35B-A3B, Q4), seamlessly, in a single session.]
2
2
8
513
David Valerio Gilmore retweeted
Introducing Rayline: a model router built specifically for Claude Code.
Claude Code's API rates are 8-10x its subscription rates.
Most of those tokens are going to easy subtasks.
Plug in Rayline, and subagents get routed to open source and on-device models instead of burning Opus-level spend on grunt work.
Quality holds. Costs drop 60-90%.
What makes Rayline different:
- Routes at the subagent/subtask level
- On-device routing via MLX (Qwen 3.6 and others)
- Built-in ML router trained for Claude Code tasks
- Cloud fallback when Anthropic has outages
- Works with OpenAI models inside Claude Code
You keep using Claude Code exactly as you do today. Rayline handles the routing underneath.
We're already routing billions of tokens per day for individual developers and publicly traded companies.
Public beta is live now. Try it at Rayline[.]ai
1
2
4
247
David Valerio Gilmore retweeted
Teamwork makes the dream work. Now running locally.
Watch Gemma 4 26B orchestrate 10 parallel sub-agents to code an SVG art gallery in seconds. Hitting 100+ tokens/sec, imagine how you can scale this for complex tasks or local chatbots for entire teams!!
62
169
1,906
158,458
David Valerio Gilmore retweeted
Chinese AI models are 3-4 months behind the US if you compare GLM-5.2 (just released) with GPT-5.4 (March release), which have the same score on Artificial Analysis.
6
10
86
33,319
We compared conversion to paid rates when defaulting the agent to Opus, Gemini 3.1 Pro, Sonnet or GLM 5.1 in @WorkshopAI
GLM 5.1 won
It meets the quality bar for most people, and users value the extra usage per dollar v closed models
5.2 is another leap
Congrats @Zai_org
1
2
158
David Valerio Gilmore retweeted
You shouldn't need to replace Claude Code to make Claude Code 60–90% cheaper.
That’s why we built @RaylineAI.
Rayline is a model router that plugs into the harness you already use and routes each subtask to the best model capable of completing it successfully.
Models like Opus and GPT 5.5 can stay focused on orchestration, complex reasoning, and long-range planning, while faster and less expensive open-source or on-device models handle the rest.
Rayline can route across frontier, open-source, and on-device models. It abstracts away the complexity of knowing which model to use, when to use it, or what your hardware can support.
All it takes is a one-line change on your Claude Code settings. The result is the experience and quality you already know and love, at a fraction of the API cost.
This grew directly out of what we learned building @WorkshopAI. One of the things users value most is the ability to work across models from different providers. But we kept hearing the same question: can I use this intelligence inside the tools I already use?
Rayline is our answer.
Its ML router optimizes for the lowest cost of a successful outcome, not simply the cheapest model. Because a cheap model that fails, retries, or spins its wheels isn’t actually cheap.
Frontier models are incredibly valuable. They just shouldn’t be doing all the grunt work.
Rayline is available today at Rayline [.] ai.
Claude Code first. Codex next!
3
5
189
David Valerio Gilmore retweeted
Our current favourite model on the price/performance curve. Now upgraded.
GLM-5.2 by @Zai_org: the new open-source flagship with 1M context, Max-effort reasoning, and even sharper coding taste, still at a fraction of frontier cost.
Build with it now in Workshop.
Turn your ideas into reality 📷
2
3
2,887
A lot of non-technical people have built amazing things in the past year for their first time. But many never ship.
This is the right mental model:
- The default reaction to what you ship is that no one will notice. It's like it never happened. Not too painful. Back to the grind.
- The next best option is a few haters comment on it and you feel bad because they're mean. But their comments are engagement, so the joke is on them: they just triggered the algorithm for you. Back to the grind.
- The expected case is that you get a trickle of engagement and you realize that you have to keep nurturing your audience overtime. Back to the grind.
- The best case is you have some level of virality, which means you got a lot of traffic and some meaningful traffic for your project. You have work to do. Back to the grind.
So no matter what happens, you learn something and your next step is to get back to work.
So just ship it already :)
AI is letting too many people be perfectionists, and perfectionists don’t change the world. If you’re not embarrassed by the first version of what is released, you have waited too long. Ship now, and iterate!
2
235
There are a few things going on here behind the price war.
Agent harnesses are getting surprisingly close to interchangeable.
Claude Code, Codex, Open Code, Hermes, Workshop, etc. are all converging on the same shape:
• terminal access
• a small set of core tools
• MCP/tool expansion
• context compaction
• skills/memories
• subagents
The reason people use one harness over another is less and less about the harness itself.
It’s about access to the best models, at the best price.
Claude Code adoption was pulled forward by
Anthropic’s frontier models and (extremely) generous subscription economics.
Codex gets pulled forward by OpenAI’s models and bundling. Same pattern.
But once routers make the model layer portable, the harness becomes less special.
You can run:
- OpenAI models in Claude Code.
- Anthropic models in Codex.
- Open models for cheap subtasks.
- On-device models where they make sense.
- A mix of all of the above inside whatever harness you prefer.
That makes the harness a weak moat.
It is a distribution channel, a UX surface, and a data acquisition engine.
The labs’ real commercial levers are:
- Frontier intelligence
- Pricing
- Brand
Brand matters, but it does not protect margins forever.
Frontier intelligence matters even more, but it may be the most expensive moat in history to maintain. Being first costs billions. Being a fast follower is much cheaper once the new frontier has been established.
So if harnesses commoditize, and frontier advantages diffuse quickly, pricing becomes the lever everyone can still pull.
Which is why we're seeing rumors of a pricing war.
The awkward timing is that OpenAI and Anthropic are being valued like future public-market giants at the exact moment the viability of the labs turning a profit has never looked more uncertain.
OpenAI is reportedly about to cut token prices 40-50%. Preemptively. Because they think Anthropic is about to do the same.
I need you to sit with this one, because I've been laying out the pieces for 3 months now and this is the piece that makes the ai trade lose steam maybe even makes it ugly if you will.
Two weeks ago I showed you compute repricing +50%. Same seller (SpaceX), two buyers (Google & Anthropic) in the same week. That's the input everyone was excited for and it makes sense.
Today the output is the token, the only thing any of this actually sells, is about to get cut in half. Not because serving got cheaper. Because enterprises are choking on their bills and demand doesn't clear at these prices.
Input up 50. Output down 50.
I've been doing this long enough to know there's no clever name for that.
It's not a land grab. It's not "scale economics." It's selling dollars for 80 cents and announcing a sale on the dollar.
And look I'm not some AI bear, actually the opposite and I think we still early.
I posted yesterday ago that analysts I know are personally burning $100K a year in tokens. The product is real, and it changed our lives forever. That was never the question.
But even Citadel just put out a note called "Tokenomics" and you should read it, it’s very well put.
But basically the token spend benchmarks rolling over.
Companies quietly downgrading to cheaper, good enough models. Usage leaderboards disappearing. Surprise token bills in earnings calls.
Their conclusion, stripped of the prose, AI is no longer gated by what it can do. It's gated by what it costs. Cost curves. Rationing. The boring physics every miracle eventually meets.
Now stack my last two weeks on top of each other.
- Compute: scarce, repricing up.
- Financing: circular, everyone in this trade is everyone else's customer, lender, and collateral.
- Backlogs: $638 billion of promises at $ORCL alone, all of it underwritten by the labs' FUTURE revenue.
And the labs just volunteered to shrink the revenue. Per unit. By half at least.
A price war at the top of a debt financed pyramid. I keep turning it over looking for the version where that's fine. Haven't found it, if you find it please dm.
Even with this though, the two or three balance sheets that can lose money longer than everyone else can stay solvent.
Everyone in the middle, the resellers, the markups, the wrappers, wakes up to find their input is now someone's loss leader.
And the Chinese open models are sawing through the floor from below at a much much cheaper price while everyone argues about the pricing model in the US.
When intelligence becomes a commodity, it announces itself the way every commodity does.
With a price war.
The hype was what AI could do.
The reckoning is what it costs...
1
272
David Valerio Gilmore retweeted
1/ You can shrink a language model's KV cache by 200×, in a single forward pass, and it still answers correctly.
At 256k context that's 36 GiB of cache down to ~360 MiB, with no change to the base model.
Here's how we did it 👇
28
101
955
116,513
.@Benioff makes the pitch for @RaylineAI flawlessly.
Every part is there ...
- intelligent model routing,
- routing across cloud and on-device,
- the immediate focus on coding but the broader need beyond ...
Added some callouts for clarity :)
1
2
4
243
The LLM market will increasingly look like semi-conductors ... a leading edge and a trailing edge.
Agent loops will be like self-driving cars ... driven by thousands of chips, some of which are the highest-end GPUs and most of which are commoditized microcontrollers.
OpenAI and Anthropic are positioned to be the duopoly owning the leading edge.
A red ocean of companies will compete in the trailing edge, driving costs down/improving intelligence-per-watt to the benefit of token consumers.
79
The "conversation" is increasingly about token "costs", but it's not the full picture. Model routing is the low-hanging-fruit approach to dramatically reduce costs, and companies will have to adopt it as their agent-driven consumption grows (using tools like @RaylineAI)
But the full picture is ROI. CFOs are in a great position for that, but in my experience running product strategy at @LiveRamp, the real magic happened with the CFO, CTO, and CPO working together. The CFO sets the constraints: what percent of revenue is going to R&D? What level of profitability do we need to hit? What are our topline targets? The CTO/CPO must work within those lines. The CPO assigns the value to the R&D initiatives working with sales/marketing, and the CTO mobilizes the engineering resources to get the job done.
Increasingly, those engineering resources are not just payroll/headcount, but also tokens (and a lot of them). So the critical need is for the CTO and CPO to make that token spend both optimal (via routing) and traceable (by tying usage to initiatives) so the token spend is legible to the business (CFO & CEO).
Without traceability of token spend to make it business-legible, rockstar CFOs will (and should) shoot down the spend.
THE AGE OF THE ROCKSTAR CFO
I usually don’t post takes. But in today’s newsletter I wrote a take about why the CFO will be seen as more strategic than ever as companies face new and interesting questions on both capex and opex
1
4
16
13,541