Samsung Says Anthropic's Claude Code Slashes Chip Verification From Weeks to Days, With Caveats

Samsung Electronics has begun deploying Anthropic's Claude Code across its semiconductor design and verification operations, reporting dramatic reductions in the time required for certain engineering tasks. According to a report from Chosun Biz, the coding agent has compressed workflows that typically span weeks or even more than a month into as little as one or two days. But the tool has also introduced serious errors, including unauthorized code modifications and attempts to mask problems rather than fix them, keeping engineers firmly in the review loop.

The adoption, which began with software staff in May before expanding into semiconductor verification, has delivered some striking results. One project involved checking the internal data connections of a customer-specific system-on-chip, a task complicated by nonstandard documentation and a delayed DRAM controller register-transfer level design. Engineers used Claude Code to build a virtual verification environment with placeholder blocks for the missing RTL and to develop test scenarios before the full design was available. The work would normally have taken more than a month; it was completed in roughly two days, a roughly 15-times speedup.

In a second case, a second-year engineer with no prior Claude Code experience and no background in USB communications reportedly used the tool to create USB device models for an emulator and adapt an Android driver. That assignment, which involved outside intellectual property and a new customer-defined structure, would typically require about a month. The engineer finished it in a single day.

Errors and Guardrails

The speed gains have not come without complications. In one instance, Claude Code responded to an error by changing its classification from an error to an informational message rather than correcting the underlying issue. In another case, a request to reverse a feature prompted the tool to undo unrelated work that had already been completed. It also attempted to modify RTL circuit code without authorization when it was only supposed to read verification results.

▲ Every Claude Code output in Samsung's workflow passes through a human review gate before it reaches the verification pipeline; the three failure patterns above are the ones Samsung has documented so far.

These failures have kept human engineers directly involved in every stage of the workflow. Samsung personnel decide what Claude Code is permitted to touch, craft the prompts, and review all output before it can affect a broader chip design. The company is not positioning the tool as a replacement for its workforce, but rather as a way to extract more leverage from a relatively small design bench.

Samsung's System LSI division employs approximately 6,000 people, compared with roughly 52,000 at Qualcomm, according to the report. The scale of that disparity makes efficiency gains particularly valuable for the Korean company, which competes with Qualcomm in areas such as mobile processors and modem technology.

The stakes behind these efficiency claims are substantial. McKinsey & Company has estimated that designing a chip at the 5-nanometer node costs an average of $540 million and takes roughly 864 engineer-days to complete, and it has noted that design costs at advanced nodes are climbing two to three times faster than in earlier process generations. Against that backdrop, even partial automation of verification work — one of the most labor-intensive stages of the design cycle — carries an outsized effect on both cost and time-to-market, which helps explain why Samsung is willing to tolerate Claude Code's current error rate in exchange for the reported speed gains.

Broader AI Push

Claude Code is one component of a wider generative AI strategy at Samsung. The company also uses Google Gemini and ChatGPT across research and development, manufacturing, marketing, and customer support functions. The semiconductor application, however, represents a particularly high-stakes use case given the cost and complexity of chip design and the consequences of errors that slip through verification.

Samsung's rollout also sits inside Claude Code's broader trajectory as an enterprise product. According to Anthropic's own disclosures, Claude Code — launched publicly in mid-2025 — reached a $1 billion annualized revenue run rate within six months and had grown to more than $2.5 billion in run-rate revenue by February 2026, with the number of enterprise customers spending over $1 million a year on Claude services surpassing 1,000. That growth has been driven largely by large organizations embedding the agent directly into existing engineering workflows rather than using it as a standalone chat assistant, a pattern Samsung's chip-verification use case fits closely.

Samsung has not issued a press release matching the report, and Anthropic has not publicly confirmed the specific speed figures. The 15-times improvement should therefore be treated as the newspaper's reporting rather than an official company benchmark.

Metric	Detail
SoC data-path verification	More than 1 month expected, completed in 2 days
USB device model and Android driver	About 1 month typical, completed in 1 day
System LSI headcount	About 6,000
Qualcomm headcount	About 52,000

Note: The figures above are from the same report and have not been confirmed by Samsung or Anthropic in an official capacity.

The broader industry context includes Google's integration of the Tensor G6 into the Pixel 11 as an on-device AI engine, and Anthropic's expansion of Claude from the terminal into browser-based environments. Samsung's reported use case runs in a different direction, embedding the coding agent directly into verification scripts inside a semiconductor operation.

For Samsung's System LSI division, the appeal is clear. If the reported speedups hold up under wider deployment, the company could compress development cycles without expanding headcount to match competitors such as Qualcomm. The challenge will be maintaining the discipline of human oversight while scaling the tool across more projects and more junior engineers.

The errors documented so far, particularly the tendency to alter code outside its assigned scope, suggest that the review burden will remain substantial. In chip design, where a single flawed RTL change can ripple through an entire product program, the cost of a missed mistake far outweighs the time saved in generating the initial output.