Claude vs ChatGPT for hardware design: does a 13-point gap on circuits settle it?

Ask around and you will hear that ChatGPT is better at circuits and Claude is better at code, and that this is settled. The one public number behind the first half is a 13-point gap on EEBench: GPT-6 Astra at 69.3% versus Claude Fable 5.1 at 56.4% (eebench.org). What you will not hear is that both scores carry error bars of roughly ±10 points, that the benchmark has 13 tasks, and that it is built and funded by a tool vendor.

So does that gap settle anything? For some tasks, partly. For most of a hardware engineer's week, the honest comparison looks different, and it has no single winner. Here is the task-by-task verdict as of October 14, 2026, with every number sourced and every vendor claim labeled.

The contenders

On the Claude side: Claude Opus 5.5 (claude-opus-5-5, September 22, 2026) and Claude Fable 5.1 (claude-fable-5-1, September 1), plus Claude Code as the agent. Anthropic's guidance is to start with Opus 5.5 and use Fable 5.1 for "demanding reasoning and long-horizon agentic work" (Anthropic docs). Both have 1M context and 128K output.

On the ChatGPT side: GPT-6 Astra (gpt-6-astra, September 3-4, 2026), available to paid ChatGPT users and in the API, plus Codex as the agent. Context 1,050,000 tokens with a 922,000 input cap, 128K output (OpenAI docs).

Independently, Artificial Analysis scores Opus 5.5 at 58 (the top of its index), with Fable 5.1 and GPT-6 Astra tied at 53. Neither vendor publishes a CAD score, a PCB layout metric or a datasheet-comprehension number. Our six-model roundup has the wider field.

Datasheets and long PDFs

Nobody has benchmarked datasheet reading for these models, so we are working from proxies.

Astra's proxy is retrieval: OpenAI reports 96.3% on MRCR v2 8-needle at 512K to 1M tokens (vendor-reported). That is the test of finding one line in a thousand pages, which is most of what datasheet work is. Astra's input modalities are text and image only; PDFs are handled by the ChatGPT app and file_search, not by the model directly.

Claude's proxy is context and agent behavior: 1M tokens on both models, PDF handling as an API and app feature, and Anthropic's reported 680,000-line code migration in under a day for Opus 5.5 (vendor-reported). The best datasheet anecdote we found is a Hackaday commenter who says directing Claude with datasheets halved their schematic development time. It is one person's report.

Verdict: a draw on evidence. If you ask single questions of long documents, Astra's retrieval number is the only hard figure. If you want the model to work through a datasheet inside a repo over many steps, Claude Code is the more documented path, which brings us to agents.

Schematic and netlist work

This is where the 13 points come from. EEBench's September 4 snapshot: GPT-6 Astra 69.3% ±10.7 at $0.83 per task in 1 minute 46 seconds; Claude Opus 5 61.6% ±6.0 at $2.38 in 13 minutes 1 second; Claude Fable 5.1 56.4% ±10.1 (eebench.org). Astra is faster and cheaper per task by a wide margin. The accuracy gap between Astra and Opus 5 is within the error bars; the gap to Fable 5.1 is about one error bar.

Three caveats before you change your subscription. EEBench is 13 tasks scored in SPICE, 65% on technical correctness and 35% on BOM cost, and it does not test layout. It was built and funded by atopile, a company selling an electronics design tool, so it has a conflict of interest even if the methodology is sound. And we could not verify a current entry for Opus 5.5, the Claude model Anthropic says to use by default.

On the ChatGPT side the circuit story also includes the KiCad demo from the Astra launch post: a PCB placed and routed in 2 minutes 54 seconds that OpenAI calls "manufacturable", with no metric. Quilter notes the demo shows GUI operation, not electrical validity, and a Hackaday commenter rates ChatGPT "moderately good at PCB design but still very bad with high frequency signals". On the Claude side, Anthropic has published nothing about circuits at all.

Verdict: ChatGPT, with a simulator in the loop. The numbers point to Astra, and nothing on the Claude side pushes back. Do not read "7 of 10 circuit tasks" as "7 of 10 boards work": EEBench's own example failure is a 22 µF capacitor that measures 11.4 µF under DC bias against a 545 µF requirement.

CAD code generation

Neither vendor publishes anything here, and the only independent data is from older models. BenchCAD (May 2026, 17,900 CadQuery programs across 106 part families) tested Claude Opus 4.7 and GPT-5.3. On code editing, GPT-5.3 thinking scored 0.865 and Opus 4.7 thinking 0.853. On image question-answering, Opus 4.7 scored 0.526 and GPT-5.3 0.513, both behind Gemini 3.1 Pro at 0.587. The common failures were missing threads and chamfers, the wrong workplane, sweeps and lofts replaced by extrudes, and standard-driven parameters flattened into constants.

That is a near tie on models two generations old. For the 2026 models we have only an anecdote: Sketchto had six Claude Opus 5 agents build three printable parts in CadQuery and OpenSCAD, all passing a mesh check. It is a small, vendor-adjacent test with no ChatGPT arm.

Verdict: no verdict. Use whichever model lives in your coding agent, and run a mesh or validity check on every part. The Claude Code guide has a review loop that works with either.

Agent tooling and MCP

Both companies ship a terminal agent (Claude Code, Codex CLI) and both document custom MCP servers in it. The differences are in the chat apps and the API.

In the apps, Claude supports custom connectors to remote MCP servers on every tier including Free, with one connector on Free (support.claude.com). ChatGPT has a beta Developer mode: full MCP with write actions on Business, Enterprise and Edu, but read and fetch only for individual Pro users (and Plus, per secondary sources) (help.openai.com). For ordering parts that distinction is the whole game; Which AI assistants support MCP spells it out.

In the API, Astra's Responses API lists mcp, computer_use, hosted_shell, apply_patch and code_interpreter as built-in tools. Claude offers tool use, server-side web search and code execution, programmatic tool calling and computer use. Anthropic reports Opus 5.5 at 81.8% partial on OSWorld 2.1 and Terminal-Bench 4.0 at 66.4; OpenAI reports Astra at 72.6% on OSWorld 2.0 and 57.7 on Terminal-Bench 4.0. Those are vendor-reported figures on different harnesses and, for OSWorld, different versions, so treat them as directional. One integration gotcha: Fable 5.1 returns a 400 on forced tool use (tool_choice: any or a named tool).

Verdict: Claude, narrowly, for agentic hardware work. Opus 5.5 leads the terminal and automation numbers both vendors report, and Claude's app gives individual users write-capable MCP without a business plan.

Cost per task

Token prices: Opus 5.5 $4 / $20, Fable 5.1 $10 / $50, Astra $10 / $50 per million tokens, with cache reads at $0.20, $0.25 and $1 respectively. Per task is what matters, and Artificial Analysis measures it on its index: Astra $3.26, Opus 5.5 $5.98, Fable 5.1 $7.63 (Artificial Analysis). On EEBench specifically, Astra was $0.83 per task to Opus 5's $2.38.

Verdict: ChatGPT. Astra does the same index work for about half the cost of Opus 5.5, and Fable 5.1 is the most expensive frontier model per task on the leaderboard.

Safety layers getting in the way

Both vendors' guardrails sometimes stop legitimate engineering. Kilo's preview of Astra reports OpenAI's safety monitoring pausing or halting work mid-task, alongside over-engineering ("a targeted fix often comes back as a large, sprawling change") and giving up earlier than competitors. Anthropic says its benchmarks were run with production safeguards on, which likely lowered scores, redirects cyber-flavored tasks, and notes that Opus 5.5 "often suspects it is being tested". A firmware engineer poking at a bootloader may hit either.

Verdict: a draw, with different failure modes. Astra interrupts; Claude declines or hedges.

Hallucination and injection

On the AA-Omniscience test, Astra's hallucination rate is 51% (The Next Web); we did not find a published Omniscience figure for Opus 5.5 or Fable 5.1, so we cannot make the direct comparison. On hidden prompt injection in documents, which matters when you feed a model third-party PDFs, Astra still fails 8.5% of the time against 4.8% for Claude Opus 5 (The Decoder). A Hackaday commenter reports ChatGPT recommending discontinued parts as replacements.

Verdict: Claude on injection resistance; no verdict on hallucination for lack of a matched number. In both cases the fix is the same: never let the model's memory choose a part number, make it look the part up live.

The scorecard

Task	Verdict	The number behind it
Datasheets, long PDFs	Draw	Astra MRCR v2 96.3% at 512K-1M (vendor); no Claude equivalent; no datasheet benchmark
Schematic, netlist	ChatGPT	EEBench Astra 69.3 ±10.7 vs Opus 5 61.6 ±6.0 vs Fable 5.1 56.4 ±10.1
CadQuery, CAD code	No verdict	BenchCAD on older models: GPT-5.3 0.865 vs Opus 4.7 0.853
Agents and MCP	Claude, narrowly	Terminal-Bench 4.0 Opus 5.5 66.4 vs Astra 57.7 (both vendor); MCP on every Claude tier
Cost per task	ChatGPT	AA: Astra $3.26 vs Opus 5.5 $5.98 vs Fable 5.1 $7.63
Safety layers	Draw	Reports on both sides, no measurement
Injection resistance	Claude	4.8% (Opus 5) vs 8.5% (Astra)

A plausible setup today is both: Astra for fast schematic passes in a simulated loop, Claude Code for the long sessions over the repo. We will update this post when either vendor publishes a hardware number or when EEBench posts Opus 5.5.

Both connect to the same parts server

Whichever side you land on, the step after design is identical. Next Everything is one remote MCP server at https://mcp.nexteverything.io/mcp that Claude, Claude Code, ChatGPT in Developer mode, Codex, Cursor and VS Code all reach with one snippet from the docs. The model scans your repo for a BOM (nx_scan_project), searches live stock and pricing with compatible alternates (nx_search_parts), quotes the whole build landed with shipping and duties (nx_quote) or a single STEP, STL or Gerber package (nx_quote_fabrication), and prepares a payment link (nx_checkout). It cannot pay; you open the link, tick what you need, untick what you have, and confirm.

It is free to use, and it works the same from either camp. Sign up, connect the assistant you already argue about, and let the part numbers come from live stock instead of memory.