Cadence claims 30% better chip-design agent results with Maestro

Cadence says its Maestro AI models can deliver up to 30% higher-quality chip-design results and five times the token efficiency of a general-purpose foundation model. The company plans to offer the first models to select early-access customers in Q4 2026, with about a dozen engagements underway.

Cadence announced Maestro on Oct. 26, 2026, just ahead of earnings. Its tests use an unnamed open-source foundation model that Maestro started from; Cadence calls it one of the leading open-source models and competitive with commercial frontier systems. The company has not defined how it measures the 30% quality gain, and these results remain preliminary.

What does Maestro change in chip-design agents?

Maestro supplies domain-specialized models to Cadence’s AI Super Agents. The portfolio contains post-trained language models, trained decision models and agent skills, blended with frontier and open-weight models. The agents interpret design intent, decompose tasks, evaluate trade-offs and coordinate that portfolio.

Reinforcement learning closes the loop: deterministic solvers score outputs against engineering objectives, constraints and ground truth, then feed results back into the models. The portfolio includes Nvidia Nemotron, while Cadence’s fine-tuning harness lets customers run post-training on their own GPUs and network. Cunningham said Nemotron was producing strong results and improving quickly.

Maestro embeds repeated reasoning while customers retain deployment and training control.

Where can customers run the models?

Cadence customers can run Maestro in Cadence’s cloud, public clouds or customer-controlled on-premises environments. They can post-train on proprietary data, designs and workflows without sending that material to another provider.

Synopsys took a different route with GPT-Synopsys, announced Sept. 30, 2026. It operates Synopsys tools directly on OpenAI-hosted infrastructure, uses OpenAI frontier models and does not train on customer data. Synopsys Autopilot still allows commercial, open-source and fine-tuned models in Synopsys-managed or customer-managed cloud and on-premises deployments.

For agent security, Cadence is working with Nvidia on an integration with the OpenShell Reference Architecture, launched the previous week and intended for sandboxing. Nvidia named both Cadence and Synopsys in the launch, and these options can help protect customer intellectual property.

Synopsys and OpenAI share revenue; Synopsys has not disclosed GPT-Synopsys pricing. The models differ mainly in data location: Cadence supports customer-controlled post-training, while GPT-Synopsys excludes customer data from training.

What do the 30% and 5x claims mean?

The comparison is narrower than the 30% headline suggests. In chip design, result quality often means power, performance and area, but Cadence did not say which factor its figure covers. It also disclosed no comparable result for GPT-Synopsys.

Cadence attributes the token advantage to balancing EDA and system-design-automation tool calls with LLM calls. Cunningham described the goal as using the “right model at the right time for the right task,” rather than invoking one general-purpose model for everything. A tool call can return a value after hours or days, while AI training loops expect minute-scale feedback.

Cunningham said agent costs have “really started to explode,” adding, “And I think the world realized so.” That pressure explains Cadence’s move to put some repeated agent iteration into model weights.

Maestro usage falls under a Cadence agent subscription with a consumption allowance and charges beyond it. Cadence did not say whether lower token use reduces a customer’s bill, so pricing savings remain unquantified.

Efficiency matters, but the figures do not prove cheaper design.

How did Cadence reach this point?

Maestro did not arrive in isolation. At Cadence’s July earnings call, CEO Anirudh Devgan said ChipStack had more than 20 customer engagements and ViraStack more than 25. Socionext was applying InnoStack, and Silicon Labs credited ChipStack with faster debugging. Both companies are watching Maestro, and Nvidia is also a partner.

Before the launch, Rob Knoth identified reinforcement-learning loops, domain-specific post-training of open-weight models and model routing as Maestro’s core ingredients.

• November 2025: Cadence acquired agentic verification startup ChipStack. Co-founder Kartik Hegde now leads Cadence’s core AI research group.
• February 2026: ChipStack AI Super Agent claimed up to 10x productivity for coding designs and testbenches, test plans, regressions and debugging. It has remained in early access since launch.
• April 2026: Cadence introduced ViraStack for custom and analog design, InnoStack for digital implementation and AgentStack as head agent. Socionext began deploying InnoStack in production; the others had no stated status.
• June 2026: Cadence claimed ChipStack had reached Level 5 autonomy with Nvidia Nemotron and OpenShell, delivering over 40x faster RTL validation cycles in Nvidia’s deployment. Cadence scheduled Level 5 and AgentStack for early access in H2 2026.
• July 2026: AuraStack targeted PCB and advanced-packaging design, with claims of up to 2x faster time to market and 15x higher productivity. It was available on Cadence’s Millennium M2000 supercomputer, and Nokia was deploying it as of Oct. 21.
• September 2026: ChipStack RTL Generation Agent showed average 24% less area and 18% less power than pure foundation-model code in early evaluations. Cadence scheduled select access for Q4 2026, with early Honda R&D collaboration.
• Oct. 26, 2026: Cadence placed Maestro under all four super agents. Initial results claimed up to 30% better quality and 5x token efficiency, with first-model access due to select customers in Q4 2026.

Cadence now has a shared model layer, but most benefits still rest on early tests or controlled deployments.

What must happen before these agents scale?

Both companies still distinguish early access from broad release. Synopsys targets year-end general availability for AgentEngineer and Autopilot, while GPT-Synopsys remains in early technology engagements. Cadence gives no general-availability date for Maestro.

Cunningham expects more case studies through Q4 and wants agents native to EDA, not EDA embedded in general-purpose agents. Customer results on real workflows must test the 30% claim.

The next competitive test is whether Q4 case studies show that Maestro’s token savings translate into lower bills or faster design cycles against GPT-Synopsys and Synopsys Autopilot.