Flux Is Now 70% Cheaper for AI Hardware Design

Flux is now about 70% cheaper to use. Design work that cost customers about $1,100 in late August now costs about $330.

The savings come from three things. Flux uses the latest models from every major AI lab. Adaptive sends each task to the model that does it best for the price. And we build our own tools for jobs like PCB routing. Your credits now go much further.

Alongside these savings, the agent does more work for you. Each day, it now makes more than 2.5 times as many design changes as it did in August.

We measure the cost of 1,000 design changes. A design change is one edit the agent makes to your design, such as placing a part, routing a trace, or adding a feature to an enclosure.

The work behind the line

The line does not go straight down. Lower cost is hard work, and every step has a tradeoff.

AI labs ship new models every few weeks. Each new model can be faster, cheaper, or more capable, but not always all three. We test each one on real hardware engineering tasks before it does work for you. Then we make changes step by step and compare quality and cost with real use.

In late September, cost fell to about $590 per 1,000 design changes. In October, we gave some of the most difficult work to a more capable model. That model costs more for each step, so October cost was around $850–$900. We made that choice because a wrong circuit or a bracket that does not fit costs you more than a few credits. Last week, cost fell to about $330 as we brought it back down while keeping that quality.

The tradeoffs are real. A cheaper model can make more mistakes, and each mistake costs rework. More thinking helps on a hard power, enclosure cooling, or mechanical fit decision, but it wastes credits on a simple step. Our job is to find the right balance for each task, again with every new model.

Model independence pays off

Flux uses models from OpenAI, Anthropic, Google, and other labs. When a new model ships, we test it on real hardware engineering tasks. We compare quality, speed, and cost, then use it where it improves your results.

There is no single best model for hardware engineering. Some models are strong at hard design decisions. Others read long documents well, handle spatial problems, or finish routine work quickly at low cost. A complete device needs all of these skills: a circuit and firmware, an enclosure and mounting brackets, and checks that the parts fit and work together.

Claude Code ties you to Anthropic's models. Codex ties you to OpenAI's. You can choose within that lab's model family, but not take the best value from every lab for each job. Flux can. Every lab's progress can make your next device cost less to design.

Adaptive does the model selection for you

Flux Adaptive is our best-value choice for complete-device work. We actively test and select its models. You work with one AI hardware engineer, without comparing releases or switching tools.

Hard design decisions go to frontier models. Routine and support work goes to fast, low-cost models. Thinking scales with the difficulty of the task, so a simple step does not get the same treatment as a difficult power or mechanical stress decision. You pay for the work the task needs.

Independent tasks run at the same time: part selection and enclosure design, schematic capture and mounting details. Flux combines the results and checks the device. Work that takes days or weeks by hand becomes hours in Flux.

Tools built for the job

PCB auto-routing runs on a specialized routing engine trained with reinforcement learning. It is built for the job instead of asking a general chat model to draw traces. The engine is designed to route quickly and efficiently.

AI Auto-Layout puts that engine inside your design workflow. Mechanical CAD tools build parts, enclosures, and assemblies. Simulation tools check circuits, mechanical stress, and enclosure air flow. Models decide what to change. Purpose-built tools do the geometry and analysis work. Electronics and mechanical parts stay in one project.

Less rework, more finished design

We also improve how Flux checks its results. Catching missing connections, parts that collide in an assembly, and firmware pin assignments that do not match the schematic early means less repeated work across the device. Manufacturing checks cover the board and mechanical parts, from fabrication files to enclosures for 3D printing and parts for CNC milling.

Our February update already showed how faster completion and fewer retries make credits go further. That work continues across the schematic and PCB, mechanical parts and assemblies, firmware, simulation, and manufacturing files.

Build more with the credits you have

The 70% drop in cost per design change is progress we will keep building on, week by week. New models, better tools, and fewer repeated steps make complete devices faster and more affordable to design.

Open Flux, choose Adaptive, and describe the device you want to build. Put your credits into the electronics, mechanical parts, and firmware that make it real.