Which AI model is best for hardware engineering in 2026? Six models, one real benchmark

If you pick your hardware assistant from a coding leaderboard, you are reading the wrong list. On the independent LiveCodeBench ranking, Claude Fable 5.1 is first with 90.52% (Vals). On EEBench, the only electrical-engineering benchmark any of the six big vendors cites, the same model scores 56.4% ±10.1, behind Claude Opus 5 and well behind GPT-6 Astra at 69.3% ±10.7 (eebench.org). And none of the six vendors publishes a CAD score, a PCB layout metric, or a datasheet-comprehension number. Not one.

So which of the six newest frontier models should a hardware engineer actually use? It depends on the task, and on how much evidence you will take on the vendor's word. This post separates published facts from demos, as of October 14, 2026. We will update it whenever a major model ships.

The six models at a glance

Prices are list, per million tokens, input / output, from each vendor's docs. "Hardware evidence" means anything measured on circuits, boards, parts or CAD, not general coding.

Model	Vendor, release	Price in / out	Context	Image input	MCP and tools	Hardware evidence
GPT-6 Astra	OpenAI, Sept 3-4, 2026	$10 / $50 (cached $1; prompts over 272K tokens cost more)	1,050,000 (max input 922,000), output 128K	Yes (text + image only)	mcp tool in the Responses API; ChatGPT Developer mode	EEBench 69.3% ±10.7 (independent); KiCad demo, no metric
Claude Opus 5.5 and Claude Fable 5.1	Anthropic, Sept 22 and Sept 1, 2026	Opus 5.5 $4 / $20; Fable 5.1 $10 / $50	1M, output 128K	Yes	MCP in the Claude apps and Claude Code (official)	EEBench Fable 5.1 56.4% ±10.1 (independent); nothing from Anthropic
Gemini 4 Argon	Google DeepMind, Oct 13, 2026 (general)	$2 / $10 until Nov 30, then $4 / $20; cached input 95% off	1,048,576 unified in + out; output 65,536 default	Yes, plus video, audio, PDF	Function calling; custom MCP official in Gemini Enterprise, consumer app unverified	Blender MCP demo, ERQA 87% (vendor); no CAD or PCB benchmark
Muse Wonder	Meta, Oct 12, 2026	$4 / $20, cached $0.20	~1,048,576 (secondary sources)	Yes, plus audio, video (per VentureBeat)	Tool calling listed on OpenRouter; Meta's page does not mention MCP	None published
Grok 4.7	xAI, Sept 21, 2026	$2 / $6 under 200K tokens; $4 / $12 at or above 200K	500K	Yes (JPG, PNG up to 20 MiB)	Function calling; remote MCP tools per third-party guide	EEBench 64.0% (vendor-reported)
Mistral Large 4	Mistral AI, Oct 6, 2026 (public preview)	List $1.36 / $4.18; preview $0.68 / $2.09	1M per docs (Artificial Analysis lists 524,288)	Yes	Function calling, structured outputs; MCP not mentioned	None published

Two numbers to keep in mind. The independent Artificial Analysis Intelligence Index (v4.3.2, read October 14) puts Claude Opus 5.5 at 58, Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon at 53, Muse Wonder at 52, Grok 4.7 at 46 and Mistral Large 4 Preview at 38. And cost per index task diverges from token price: Argon $1.99, Astra $3.26, Grok 4.7 $3.74, Wonder $4.52, Opus 5.5 $5.98, Fable 5.1 $7.63, because some models write 60,000 to 80,000 output tokens per answer.

How to read the evidence

There are three kinds of claim in this post, and they are not equal.

Vendor-reported numbers come from launch pages. Every vendor runs its own harness, and Artificial Analysis, Vals and ARC Prize routinely get lower numbers under neutral ones. ARC-AGI-3 is the extreme case: 99.9% for GPT-6 Astra under OpenAI's stateful harness, 62.7% on BenchLM, roughly 17-63% in ARC Prize's stateless runs.

Independent hardware benchmarks exist, but they are young and small. EEBench is 13 schematic and circuit tasks simulated in SPICE; the score is 65% technical correctness plus 35% BOM cost efficiency, and it does not test layout. It is built and funded by atopile, a tool vendor: a useful signal with a conflict of interest. Its error bars are around ±10 points, wider than most gaps between models. The academic benchmarks (HWE-Bench, BenchCAD, PCB-Bench) tested older models and have not been re-run on the 2026 releases.

Anecdotes are forum posts and vendor case studies. They show what failure looks like, not how often it happens.

GPT-6 Astra

What it is: OpenAI's flagship, in limited preview from September 3 and open to paid ChatGPT users from September 4, 2026 (Wikipedia). The API model is gpt-6-astra, also on Azure, Amazon Bedrock and OpenRouter. Context is 1,050,000 tokens with a 922,000 input cap and 128K output; input is text and image only, knowledge cutoff April 30, 2026 (OpenAI docs). The Responses API exposes web_search, file_search, code_interpreter, computer_use, mcp and hosted_shell. List price $10 / $50 per million tokens, cached input $1, Batch and Flex at 50% off.

Hardware evidence: Astra has the best published electronics number of any model. On EEBench's September 4 snapshot it scored 69.3% ±10.7 at $0.83 per task in 1 minute 46 seconds, against Claude Opus 5 at 61.6% ±6.0, $2.38 and 13 minutes (eebench.org). Note the error bars overlap. The launch post includes a "Circuit board" demo in which Astra places and routes a PCB in KiCad in 2 minutes 54 seconds and OpenAI calls the result "manufacturable"; no PCB metric is published. Quilter, a layout vendor, points out that the demo shows GUI operation, not electrical validity. Cherry Blossom, another vendor, reports first-pass fab yield rising from 94% to 97% across 40 boards behind its own verification gate, raw-model yield not given. For long documents, OpenAI reports MRCR v2 8-needle retrieval at 96.3% in the 512K to 1M range (vendor-reported).

What breaks: Kilo's preview notes describe over-engineering ("a targeted fix often comes back as a large, sprawling change"), giving up earlier than competitors, and OpenAI's safety monitoring halting legitimate work mid-task. The AA-Omniscience hallucination rate is 51%, against 15% for Gemini 4 Argon (The Next Web). Hidden prompt injection in documents still gets through 8.5% of the time, versus 4.8% for Claude Opus 5 (The Decoder). A Hackaday commenter reports ChatGPT recommending discontinued parts as replacements; another calls it "moderately good at PCB design but still very bad with high frequency signals".

Who should use it: engineers iterating on schematics with a simulator in the loop, where speed matters and every output gets checked. Not for unverified BOM work, given the hallucination rate and the discontinued-part reports.

Claude Opus 5.5 and Claude Fable 5.1

What they are: Fable 5.1 shipped on September 1, 2026 (claude-fable-5-1), Opus 5.5 on September 22 (claude-opus-5-5), both on the Claude API, Bedrock, Vertex, Microsoft Foundry and the Claude apps. Anthropic's own guidance is to "start with Claude Opus 5.5 for most workloads" and reach for Fable 5.1 "for demanding reasoning and long-horizon agentic work" (models overview). Both have 1M context and 128K output, with text and image input; PDF handling is an app and API feature rather than a model modality. Opus 5.5 costs $4 / $20 with cache reads at $0.20; Fable 5.1 costs $10 / $50 with cache reads at $0.25. Opus 5.5 tops the Artificial Analysis index at 58.

Hardware evidence: Anthropic has published nothing on CAD or PCB. Third-party data: Fable 5.1 scored 56.4% ±10.1 on EEBench's September 4 snapshot; we could not verify a current entry for Opus 5.5. HWE-Bench (March 2026, board-level schematic design from 2,914 datasheets) found the best overall model was the older Claude Sonnet 4.5 at 8.15% fully correct designs. BenchCAD (May 2026) tested Claude Opus 4.7, not the 5.x line. Adjacent, Anthropic-reported: Opus 5.5 scores 58.7 on Terminal-Bench-Science against 22.4 for GPT-5.6 Sol and 81.8% partial on OSWorld 2.1 computer use. One small, vendor-adjacent test from Sketchto had six Claude Opus 5 agents build three printable parts in CadQuery and OpenSCAD, all passing a mesh check.

What breaks: Fable 5.1 is the most expensive model per task on Artificial Analysis at $7.63. Anthropic runs its benchmarks with production safeguards on and says that likely lowered scores; cyber-flavored tasks get redirected. Anthropic notes Opus 5.5 "often suspects it is being tested". For integrators: forced tool use (tool_choice: any or a named tool) returns a 400 on Fable 5.1.

Who should use it: anyone running long agentic sessions over a hardware repo, a KiCad project next to CadQuery scripts and firmware, especially in Claude Code where MCP is first-class. How to use Claude Code for hardware design covers the repo layout. Pick Opus 5.5 by default and Fable 5.1 only when Opus stalls.

Gemini 4 Argon

What it is: announced September 30 for vetted cyber defenders and released generally on October 13, 2026 as gemini-4-argon in AI Studio, the Gemini app (AI Pro and Ultra) and Google Antigravity (Google blog). It takes text, image, video, audio and PDF as input, with a unified 1,048,576-token context and 65,536 default output, extendable to 1M in "long decode" (Gemini docs). Introductory price $2 / $10 until November 30, then $4 / $20, with cached input 95% off. It is the cheapest of the six per index task at $1.99, with the lowest hallucination rate measured by Artificial Analysis (15%), although its accuracy on the same test was 50% against Astra's 63%.

Hardware evidence: no CAD or PCB benchmark, but the most hardware-adjacent demo set. Google's page shows Argon driving Blender through Blender MCP, and learning a cornhole throw from video, building the simulation and transferring it to Apptronik's Apollo humanoid for seven consecutive shots. Google claims 87% on ERQA, an embodied-reasoning benchmark, and "a major improvement in spatial reasoning, mathematical precision, and visual coding" (all vendor-reported). Google reports Terminal-Bench 4 at 57, with Sonnet 5.5 at 64, Opus 5.5 at 60 and Astra at 59.

What breaks: Bloomberg reporting quotes several Google employees with access saying Argon struggles on real coding tasks, especially front-end, two calling it "benchmaxxing"; Google disputes this. The model is token-hungry at around 62,000 output tokens per index task. The introductory price doubles on December 1. We could not verify an EEBench entry. Custom MCP servers are official in Gemini Enterprise; in the consumer app they are reported but not documented.

Who should use it: engineers who feed the model PDFs and video directly (datasheets, app notes, a phone clip of a mechanism) and want the lowest cost per task. Treat the coding claims with the caution Google's own staff do.

Muse Wonder and Muse Wonder Full

What it is: Meta Superintelligence Labs' newest line, released October 12, 2026 (Meta). Muse Wonder is on the Meta Model API as muse-wonder-1.0 with OpenAI- and Anthropic-compatible interfaces; Wonder Full is only inside Muse Code subscriptions at $50 and $200 per month. No downloadable weights for either. Price $4 / $20, cached $0.20, flat regardless of context length. Artificial Analysis scores Wonder Full at 56 and Wonder at 52.

Meta-reported benchmarks are strong on software: DeepSWE 1.1 at 79.0 (Full) and 78.8 (Wonder) against 74.1 for Astra and 74.2 for Opus 5.5. On Terminal-Bench 4.0-AA it trails, 57.1 and 52.0 against 59.1 and 59.6.

Hardware evidence: none. We found no CAD, PCB, physics or spatial-reasoning results for either model.

What breaks: Muse Code defaults to "Contributor mode", in which Meta may train on submitted data and the terms prohibit confidential data; the pay-as-you-go API runs in "Standard mode" with no training. Zero-data-retention needs approval and disables server-side state, file uploads and web search. Meta's tables mix internal and third-party harnesses, Meta says the models "can make mistakes in unfamiliar multi-step tasks", and the full safety report is "available soon". Meta's page does not mention MCP.

Who should use it: teams on the OpenAI or Anthropic SDK who want a second strong coding model behind the same interface. Do not put a proprietary board design through Muse Code in its default mode.

Grok 4.7

What it is: xAI's model from September 21, 2026, available as grok-4.7 on the API, in Grok Build (a CLI coding agent with a free tier), Cursor, GitHub Copilot, OpenRouter and, from September 28, Amazon Bedrock (x.ai). As of September 29 it was not in the consumer Grok apps, which still ran Grok 4.6; we could not verify the status by October 14. Context 500K, input text and images, knowledge cutoff May 2026, reasoning cannot be switched off (xAI docs). Price $2 / $6 for prompts under 200K tokens, $4 / $12 at or above, but around 81,000 output tokens per index task makes it $3.74 per task. The API supports function calling, structured outputs, web search, code execution and, per a third-party guide, remote MCP tools. No Batch API.

Hardware evidence: xAI is the only vendor of the six that puts an electronics benchmark on its launch page. It reports Grok 4.7 at 64.0% on EEBench (xhigh effort), up from 53.0% for Grok 4.6, with Fable 5.1 Max at 56.4% for comparison (vendor-reported). The independent September 4 snapshot had Grok 4.6 at 57.1% ±9.9, when atopile called xAI "only the second lab to cross the 50% mark". No layout or CAD numbers.

What breaks: long, hard software work. xAI itself reports FrontierSWE V2 at 29.0% against 56.3% for Fable 5.1 and Terminal-Bench 4.0 at 37.6%, which Artificial Analysis measures at roughly 26% under its neutral harness, noting regressions on long-context retrieval and AutomationBench versus Grok 4.6. Hallucination rate 29%. Outputs run very long.

Who should use it: circuit-level work in an agent harness with simulation, where the EEBench numbers and the low token price matter. Not for multi-file refactors of a firmware tree.

Mistral Large 4

What it is: a public preview from October 6, 2026 (mistral-large-4-0, version 26.10) on Mistral Studio and the API (Mistral). A mixture-of-experts model with 1.05T total and 52B active parameters plus a 1.6B vision encoder, 1M context per the docs (Artificial Analysis lists 524,288), text and image input, function calling, structured outputs and document Q&A. Weights are promised for the end of October with no license stated; on-prem deployment is planned. List price $1.36 / $4.18, preview price $0.68 / $2.09, cached $0.07. Artificial Analysis scores it 38 at $1.13 per task.

Hardware evidence: none. The closest vendor-reported number is Dense 200 visual grounding at 42% against 41% for GPT-6 Astra. Mistral also reports Terminal-Bench 4 at 28.3% and a Surge AI blind coding score of 3.74 against 4.22 for Claude Opus 5.

What breaks: Mistral says the RL run is "still in progress", its own materials disagree on specs, and the model sits well behind the top five on independent indices. MCP and a code interpreter are not mentioned.

Who should use it: teams with an on-prem or EU-only requirement who can wait for the weights. Today, a cheap first pass; verify everything.

What no benchmark measures

Four things decide whether an AI-generated design works, and no public benchmark covers them for the 2026 models.

The physics check. EEBench's own example failure is a 22 µF hold-up capacitor that measures 11.4 µF at 4.7 V of DC bias, against a requirement of 545 µF (eebench.org). The netlist was fine; the part was wrong. HWE-Bench found models reach roughly 72% static connectivity and 59% SPICE pass on average, but only 8% fully correct designs, and concluded they "lack physical intuition" (arXiv 2603.18102). A fully connected board can still fail on return paths, impedance, coupling or assembly.

Footprints and standards. BenchCAD lists the recurring CAD failures: missing threads and chamfers, the wrong workplane, sweeps and lofts replaced with extrudes, standard-driven parameters flattened into fixed numbers (its example is a DIN 2095 spring). Text2CAD-Bench found invalid-code rates from 17% (Gemini 3 Flash) to 81% (MiniMax M2.1), on older models.

Discontinued parts. Models recommend parts that no longer ship, because their training data does not know. The only fix is a live stock lookup at the moment of choosing.

Spatial reasoning. PCB-Bench (ICLR 2026) reports "significant shortcomings in spatial layout reasoning, rule constraint following, and engineering drawing interpretation". On GeoGramBench, 19 frontier LLMs scored below 50% on the hardest tier (via Quilter). Google's ERQA 87% for Argon is the only 2026 claim of progress, and it is vendor-reported.

Recommendation by task

Task	First choice today	Why	Always verify with
Reading datasheets and long PDFs	Gemini 4 Argon, or GPT-6 Astra	Argon takes PDFs natively at the lowest cost per task; Astra reports 96.3% long-range retrieval (vendor). No datasheet benchmark exists.	The datasheet itself, page and table cited
Generating CadQuery or build123d	Whatever runs in your coding agent	Only older-model data: on BenchCAD code edits, GPT-5.3 thinking 0.865, Opus 4.7 thinking 0.853; a specialist model beat Gemini 3.1 Pro on vision-to-code, 0.601 to 0.397.	A mesh or validity check, threads and chamfers inspected
Schematic and netlist review	GPT-6 Astra, then Grok 4.7	EEBench: Astra 69.3% ±10.7 (independent), Grok 4.7 64.0% (vendor), Opus 5 61.6% ±6.0.	SPICE, then a human
BOM work: extraction, alternates	Gemini 4 Argon or Claude Opus 5.5 in an agent	Hallucination rates: Argon 15%, Grok 29%, Astra 51%. Discontinued parts are the common failure.	Live stock and lifecycle lookup before ordering
Ordering parts via MCP	Claude (apps and Claude Code), ChatGPT Developer mode, any MCP-capable coding tool	Only Claude and ChatGPT document custom MCP for end users; Gemini and Grok are reported, Muse and Le Chat silent. Details in Which AI assistants support MCP.	The checkout page: the agent cannot pay

For the Claude versus ChatGPT question specifically, see Claude vs ChatGPT for hardware design.

Whichever model you pick, the ordering step is the same

Every model above stops at the same wall: it can write the netlist, draft the bracket and list the parts, but it does not know what is in stock today, what a landed price is, or that the regulator it likes was discontinued. That is the part Next Everything handles, for any model.

It is one remote MCP server at https://mcp.nexteverything.io/mcp, authenticated with a personal nx_live_… key. Claude, Claude Code, Cursor, Codex, VS Code, Windsurf, Gemini CLI and Zed all connect with one snippet from the docs. The tools are the same for every model: nx_scan_project extracts a BOM from a repo and resolves each line to an orderable part, nx_search_parts returns live stock and pricing with compatible alternates, nx_quote prices a full build landed, nx_quote_fabrication quotes a STEP, STL, 3MF, DXF or Gerber package, nx_checkout creates a payment link, and nx_track_order follows every sub-order. Quotes are valid 24 hours. The agent cannot pay; you open the link and confirm. No assistant? Upload a BOM, Gerbers or a STEP file on the site. MCP in five minutes walks through setup.

Next Everything is free and works with every model in this post. Sign up, connect the model you already use, and let it price the parts while you check the physics.