I examined Claude Code, Codex, and Antigravity on an actual electronics mission — just one truly completed

AI coding brokers are a dime a dozen in 2026, however not all of them are price your time. Completely different fashions reply in a different way to issues, and a mannequin that may create a flawless web site in minutes may not be nearly as good at writing bash scripts. Now there are sure tasks the place I will not vibe code, however engaged on smaller electronics tasks is the right place for these fashions to shine.

So I put Claude Code, Codex, and Antigravity to the take a look at in an actual electronics mission — a easy dashboard with an ESP32, an OLED show, and a temperature sensor that may do a number of issues without delay. And whereas all three fashions can get you a working prototype, you’d not wish to blindly decide your coding assistant.

A easy dashboard changed into a critical take a look at

Constructing an ESP32 gadget that blended APIs, sensors, shows, and Dwelling Assistant

In the event you’re working with bigger codebases or monorepos, Claude Code’s moderately restricted context window and utilization limits is usually a downside. I solved Claude Code’s reminiscence downside with a Postgres database, however continually clicking a tiny icon to see how a lot of my utilization quota is left was irritating.

I made a decision to repair this by making a small devoted dashboard that confirmed my utilization limits, so as a substitute of clicking round or typing in instructions, all it takes is a look to see the place I am at. I additionally determined so as to add a temperature sensor so the dashboard may present my workplace’s ambient temperature and humidity, which I can pipe into my Dwelling Assistant dashboard to arrange automations for my air conditioner.

It is a comparatively easy construct. All you want is an off-the-shelf ESP32 — no threat of selecting the unsuitable ESP32 board right here — a DHT22 temperature sensor, a generic monocolor OLED display (I am utilizing an SH1106 show, however another show works tremendous), and a easy push button to cycle by means of the knowledge on the show. That is it.

I spent a while in Blender to provide you with a fast enclosure for the mission, and a fast 3D print later, I had every thing I wanted. Now it was time for the coding brokers to get to work. All three brokers have been supplied the identical immediate as follows:

I wish to make an ESP32-based dashboard that does three issues:

1. It ought to present my Claude plan restrict utilization
2. It ought to present the temperature and humidity studying from an hooked up DHT22 sensor and in addition report it to my Dwelling Assistant occasion.
3. It ought to present me the climate and temperature outdoors in New Delhi. I’ve an ESP32 Devkit V1 board, a DHT22 sensor, and an SH1106 128×64 monochrome OLED.

There’s additionally a 4-pin push button that can be utilized to cycle modes between the three functionalities and replace the show accordingly.

The aim was to achieve a working prototype assembly all three aims in as few and as easy prompts as potential. All three brokers would take multiple immediate to get the mission proper, however they took considerably completely different approaches to get the job finished.

Antigravity moved quick

Spectacular pace, however extra planning than doing

First up on the record was Google’s Antigravity operating Gemini 3.1 Professional on Excessive effort, and it’s quick, simply not in a means you’d count on. As quickly because the mannequin acquired its immediate, it searched the web for the way to examine Anthropic API plan restrict utilization and offered an implementation plan with some follow-up questions.

I began all three brokers on the similar time, and whereas Codex and Claude Code each got here up with the primary model of a program able to be flashed to the ESP32, Antigravity insisted on my approving its implementation plan earlier than it wrote a single line of code. After I cleared up its doubts, it got here up with yet one more implementation plan and clearly said that it will not be capable of fetch my Claude Professional subscription’s utilization limits, as there isn’t any official API from Anthropic that permits this performance. The temperature knowledge from the sensor could be despatched over to Dwelling Assistant through its built-in REST API, which is the simpler solution to do it except you are utilizing one thing like ESPHome.

I then piped in Hermann Björgvin’s Clawdmeter GitHub repo, which is the same mission based mostly on the Waveshare ESP32-S3-Contact-AMOLED-2.16 board, to push it alongside. It was then that the mannequin was ready to determine that it may use the OAuth entry token generated by the Claude CLI and ship it to a particular beta API endpoint to fetch the boundaries. Thoughts you, Antigravity nonetheless hadn’t produced a single line of code up to now. In the event you’re of the opinion that Antigravity is not able to tackle Claude Code and Codex, this does nothing however reinforce it.

As soon as I accredited the third and remaining motion plan, it created a PlatformIO mission that labored for roughly two minutes to supply a collection of recordsdata for me to sort in my Wi-Fi credentials and OAuth token individually from the principle supply code, all able to be flashed on the ESP32. That is the fitting method when it comes to safety, however for a tool that’ll by no means go away your native community, it is unnecessarily advanced. To not point out that when you’re a newbie to improvement boards just like the ESP32 and the Arduino IDE, seeing a bunch of recordsdata while you’re solely anticipating one may be overwhelming.

Codex by no means actually understood the task

The errors that saved it from producing a usable outcome

OpenAI’s Codex, utilizing GPT 5.5 on Excessive effort, wasted no time and produced two YAML recordsdata in about three minutes from receiving the immediate, whereas moving into a very surprising course. It used ESPHome to configure the ESP32 as a sensor that may natively present up in my Dwelling Assistant occasion. This is able to work greatest if I have been working with simply the DHT22 sensor to create a devoted temperature sensor for my workplace, however for a board dealing with two different duties, it wasn’t fairly the fitting method.

Codex additionally identified that Anthropic had no official API for fetching Claude Professional utilization limits. Curiously, it did not let that hurdle cease it. Codex created a Dwelling Assistant input_number known as input_number.claude_plan_usage_percent to primarily manually set a quantity I replace myself (or through an automation) to point utilization. Offering the choice to create an actual API-backed HA sensor if I had entry to Anthropic’s Console group or admin key. I do not.

In my follow-up, I supplied the GitHub hyperlink to the identical Clawdmeter mission I gave to Antigravity. It understood the method, however as a substitute of ditching ESPHome, it doubled down on the identical path of utilizing manually set numbers as earlier than and prompted that the following step could be including a Clawdmeter-style bridge that writes actual Claude Code utilization into the helper lessons it created beforehand. It did, nevertheless, spit out a moderately clear wiring diagram for connecting all of the elements.

Now this method can work, nevertheless it requires loads of workarounds and suboptimal practices that add complexity for no good motive. It appears as if Codex determined to give attention to the temperature sensor a part of the instruction (doubtless as a result of it’s higher documented) and bent every thing else to work in that ecosystem as a substitute of adjusting the construction to higher go well with the duty at hand.

It might take fairly a little bit of forwards and backwards to get the prototype working as anticipated. Actually, it might’ve been quicker for me to jot down the code myself as a substitute of losing time down the ESPHome rabbit gap that Codex was so persistent about.

Claude Code took its time, nevertheless it acquired the duty proper

Slower iterations, higher reasoning, and code that really match the duty

Claude Code utilizing Sonnet 4.6 on Excessive effort was the slowest of the three brokers proper out of the gate. Primarily as a result of I used to be continually exhibiting a mannequin overloaded warning — indicating excessive server load. This can be a server-side difficulty you can’t do a lot about, and it is the primary time I am seeing this within the a number of months I have been utilizing Claude Code, so I made a decision to let it slide.

After some pondering, Claude produced a single INO file (utilized by the Arduino IDE) that did all three duties proper from the primary model. Not like Antigravity’s refusal or Codex’s delusion, Claude Code solved the utilization restrict fetching downside by sending POST /v1/messages with max_tokens:1 to Haiku, returning the headers that present requests and tokens remaining within the present window. It was extraordinarily clear about what the outcomes would and wouldn’t present, and my month-to-month token consumption in opposition to my Claude Professional subscription wasn’t included.

Identical as earlier than, I threw in a hyperlink to the Clawdmeter mission to assist the agent alongside, and it instantly tailored. Claude Code up to date the Arduino sketch with the brand new headers, spit out a preview of what could be proven, and a single INO file able to be flashed to the ESP32, together with directions on the place to search out the Claude OAuth token.

Much like Antigravity, Claude Code used Dwelling Assistant’s built-in REST API to ship over temperature knowledge from the sensor, however crucially, it additionally identified that the combination will not persist by means of Dwelling Assistant server restarts except the board posts knowledge once more earlier than the restart is full. It urged utilizing MQTT for a cleaner integration, which might require some extra setup on the Dwelling Assistant entrance.

Just one agent crossed the end line

Which mannequin finally delivered a working dashboard and why

For my function, Claude Code was the mannequin that produced the perfect outcomes. It gave me a clear wiring diagram and a single INO file to work with, which diminished mission complexity. On a tool that is going to sit down on my native community, there is not any motive so as to add extra recordsdata for OAuth tokens or Wi-Fi credentials.

Moreover, since I take advantage of Claude extra, it already had my board and sensor stock saved in reminiscence, it knew the native IP deal with my HA occasion lives on, and it wired all of that straight into the code, giving me fewer issues to edit earlier than flashing the ESP32. You may deal with Clade Code as an agent, however it may be a military with the fitting context. In the event you’ve been utilizing Codex or Antigravity for some time, the identical may apply to you.

Antigravity was an in depth second. It tried to do every thing, and it tried to do it proper, however added pointless complexity, and the code needed to be manually patched and debugged earlier than it might work. Antigravity additionally did not present a wiring diagram, that means I needed to look by means of the code to determine the way to wire up my {hardware}.

Final however not least, Codex made the fitting name with ESPHome, however solely within the context of the dashboard primarily being a temperature-sensing station. Regardless of offering steering on the way to fetch utilization limits, it continued down the ESPHome path, including pointless complexity and forcing me to arrange the entire thing in YAML when it might’ve been a lot simpler to flash the board through Arduino for this sort of process. You’ll be able to ask the agent to divert it is method and create an Arduino sketch that works equally to the Antigravity and Claude Code method, however that is one other back-and-forth spent troubleshooting and fixing mission structure.

When pace issues, when accuracy issues, and the place every coding agent matches

As I discussed earlier than, all three approaches would work, however some would require extra tinkering than others. One instrument is not universally higher than the others; it is that embedded firmware exposes the hole between instruments that generate code and instruments that really perceive the issue. Relying on which instrument you most frequently use and what method you wish to take, your outcomes will fluctuate.

Antigravity is quick and spectacular with the fitting context. Codex falls again to higher documented strategies, even when they’re unsuitable for the duty at hand. However on a mission the place completely different subsystems should cooperate on constrained {hardware}, Claude Code’s slower, extra deliberate, and total easier method produced the one outcome I may truly flash to the board and begin utilizing instantly.