A week for a device, a tenth of the coding bill
OpenAI and Revolut are attacking the same engineering bottleneck from opposite ends of the stack: OpenAI is showing how generative models can accelerate physical product design, while Revolut is showing how model hosting choices can change the economics of software development at enterprise scale. OpenAI said on November 3, 2026, that its models are now helping engineers design, simulate and build physical products across mechanical design, electronics, software and manufacturing workflows . Revolut, meanwhile, says it has cut AI coding inference costs by roughly 90% by moving about 2,000 engineers toward self-hosted GLM models instead of relying primarily on metered commercial coding services .
The shared theme is compression. OpenAI is compressing the calendar: fewer weeks lost between idea, PCB, enclosure, firmware and a physical prototype. Revolut is compressing the invoice: fewer dollars per million tokens, less bill shock and more freedom for engineers to delegate routine work. Neither story proves that AI can remove engineers from the loop. Both suggest that the shape of engineering work is changing fast.
OpenAI’s Codex pet: from software agent to physical companion
OpenAI’s most tangible example is a small “talk to Codex” device, described as a test device that lets users interact with the company’s coding agent through dedicated hardware . The company said GPT-6 was used to define the functional requirements, while Astra developed the enclosure, internal structures, component selection, PCB layout, user interface, firmware and software integrations needed to connect the device to Codex . Two team members worked with Astra on the side project for under a week, and a contract manufacturer produced a small batch of prototypes in under a month .
That is not the same as saying Astra produced a factory-ready product in one pass. OpenAI explicitly said the device still needed several rounds of review and iteration before manufacturing, because Astra did not consistently account for production limits, tolerances or assembly details . The Decoder’s report framed the same limitation clearly: the model could traverse the design chain from enclosure to firmware, but manufacturing constraints remained a real-world checkpoint .
That caveat matters. Hardware is not code with atoms attached. A board that routes correctly in software may still be awkward to assemble. A case that looks plausible may fail tolerance, heat, serviceability or supplier constraints. A useful prototype is not a reliable product. But the demonstration is still important because it moves AI assistance beyond code completion and into multidisciplinary product engineering.
Why this changes R&D economics
The cost of hardware R&D is often hidden in waiting: waiting for specialists, waiting for CAD revisions, waiting for PCB layout, waiting for firmware glue, waiting for the next prototype spin. OpenAI is trying to show that models can help collapse some of those handoffs. Its post says internal hardware and infrastructure teams use Codex and ChatGPT to create and revise CAD models, prepare manufacturing drawings, set up simulations and review designs across robotics and compute infrastructure . It also said model usage across those hardware teams, measured in tokens, has grown 23-fold since February .
OpenAI also disclosed a benchmark signal. On BenchCAD 2.0, which tests whether models can reconstruct 3D parts and assemblies and identify circuit-board connections, a newer internal model scored 0.875, compared with 0.594 for GPT-6 Astra at its highest reasoning setting . The company did not share enough detail to treat that as an independent industry benchmark, but it indicates where OpenAI is aiming: models that can work inside engineering tools, not merely talk about them.
The industrial angle is also broader than the Codex pet. OpenAI said it is expanding work with Chip Ganassi Racing on workflows that turn 3D scans of existing parts into editable CAD models and 2D manufacturing drawings, and it announced a research collaboration with Mazda on aerodynamics and crash-safety engineering . For automotive, racing and infrastructure teams, the potential prize is not a cute desk device. It is faster iteration through design spaces that are expensive for humans to explore manually.
Revolut’s lesson: the token price changes behavior
Revolut’s case study is less visual but arguably more immediately commercial. The Register reported on November 3 that Revolut says it reduced AI coding inference costs by roughly 90% in less than a year by making self-hosted models the default rather than relying mainly on commercial services such as Claude . David Abitbol, Revolut’s operations and infrastructure lead, said the rollout covered roughly 2,000 engineers and was associated with a 60% improvement in coding time and a 70% improvement in lead time .
The infrastructure detail is the point. Revolut initially adopted GLM-5.1 and then moved to GLM-5.3, which it now serves using 56 Nvidia B200 GPUs . The Register reported that self-hosting GLM cost about $0.30 per million input/output tokens, roughly one tenth of using OpenAI Codex or Anthropic Claude Code in Revolut’s comparison . RuntimeWire likewise reported that GLM became the default coding model for about 2,000 engineers, with Claude Code and Codex kept for exceptions .
This is not simply a procurement story. Revolut’s reported gain came from changing how engineers use AI once tokens no longer feel scarce. Abitbol said the biggest bottleneck was not only writing code, but the surrounding work: debugging, Jira tickets, Confluence material and finding requirements . When tokens are cheap enough, engineers stop reserving AI for “important” prompts and start applying it to the messy connective tissue of software delivery.
Self-hosting is powerful, but not free
The Revolut story should not be misread as “every company should buy GPUs tomorrow.” Serving GLM-5.3 for 2,000 engineers is a scale case. The reported setup uses 56 Nvidia B200 GPUs, roughly equivalent in count to seven eight-GPU DGX B200 systems . Revolut also had to tune the stack with NVFP4 quantization, disaggregated prefill and decode, context-length management and latency targets . In other words, the savings depend on engineering maturity, high utilization and operational discipline.
There are governance trade-offs too. Revolut’s argument is that running open-weight models on its own infrastructure can reduce data exposure compared with sending proprietary code to third-party APIs. The Register reported Abitbol’s view that the main concern with Chinese open-weight models is not the weights themselves but using hosted APIs running out of China; on-premises or regional cloud deployment changes that risk profile . That is a plausible governance argument, but it still requires security review, model evaluation, audit logging and clear rules for what coding agents can do.
The technical compromise is also visible. GLM is text-only, so Revolut still routes some vision and UI tasks to other models such as Gemini or GPT Luna, while keeping code generation on GLM . That hybrid pattern may become common: cheap self-hosted models for routine work, frontier models for exceptions, and routers deciding when to escalate.
The bigger signal: engineering cost is becoming a systems problem
OpenAI and Revolut are not telling the same story, but they rhyme. OpenAI’s story says the boundary of AI-assisted engineering is moving from code toward CAD, PCB design, simulation and physical prototypes. Revolut’s story says the economics of AI-assisted engineering may depend less on the brand of the assistant and more on whether the inference stack is cheap enough to use constantly.
The strategic lesson for engineering leaders is clear. The winning setup may not be the single best model. It may be the best loop: fast-enough models, cheap-enough tokens, strong-enough governance and workflows that keep humans in charge of review, taste, safety and manufacturability.
Tony Stark may not get his lab back. But the lab is definitely getting a model router, a GPU budget and a much shorter prototype calendar.
Comments
Be the first to comment.