May 2026 Tech Roundup: Gemini 3.5, Cohere Command A+, Autonomous Vehicles, and DNA Synthesis Breakthroughs

Introduction

The month of May 2026 has been marked by a flurry of meaningful, non‑political technology news that showcases how quickly the frontier is moving. From AI models that can reason and act autonomously to self‑driving vehicles navigating complex urban environments, and from revolutionary DNA‑synthesis techniques that bridge the gap between generative AI and biological construction, the developments reported this month offer a glimpse into a near‑future where intelligent systems are increasingly capable, accessible, and integrated into real‑world workflows.

AI Models & Providers: Leaping Toward Agentic Intelligence

Google Gemini 3.5 Flash: Frontier Intelligence with Action

On May 19, 2026, Google announced the release of Gemini 3.5 Flash, the first model in its Gemini 3.5 family. Designed expressly for agentic workloads, Gemini 3.5 Flash combines high‑level reasoning with remarkable speed, positioning itself in the top‑right quadrant of the Artificial Analysis index—where frontier‑level intelligence meets exceptional latency.

The model is available today to billions of users via the Gemini app and AI Mode in Google Search, to developers through Google Antigravity and the Gemini API in AI Studio and Android Studio, and to enterprises via the Gemini Enterprise Agent Platform and Gemini Enterprise.

Performance highlights include:

• Outperforming Gemini 3.1 Pro on challenging coding and agentic benchmarks: Terminal‑Bench 2.1 (76.2%), GDPval‑AA (1656 Elo), and MCP Atlas (83.6%).
• Leading in multimodal understanding with 84.2% on CharXiv Reasoning.
• Generating output tokens at 4× the speed of other frontier models.

These numbers translate into tangible real‑world impact. Early adopters report that Gemini 3.5 Flash can automate multi‑week workflows in a fraction of the time. For example:

• Shopify runs subagents in parallel to analyze complex merchant data for global growth forecasts.
• Macquarie Bank pilots the model to accelerate customer onboarding by reasoning over 100+‑page documents.
• Salesforce integrates Gemini 3.5 Flash into Agentforce to automate complicated enterprise tasks via multiple subagents that retain context and execute multi‑turn tool calls.
• Ramp uses the model for smarter OCR by combining multimodal invoice understanding with historical pattern reasoning.
• Xero deploys agents to autonomously manage multi‑week workflows such as supplier identification and 1099 tax‑form preparation.
• Databricks leverages agentic workflows to monitor real‑time information, reason across massive datasets, and propose fixes for data scientists.

Beyond raw performance, Gemini 3.5 Flash introduces a new personal AI agent, Gemini Spark, which runs 24/7 to help users navigate their digital lives. Built on the same 3.5 Flash foundation, Spark is currently rolling out to trusted testers and will reach Google AI Ultra subscribers in the US next week.

Safety remains a cornerstone: the model was developed under Google’s Frontier Safety Framework, with strengthened cyber and CBRN safeguards and the inclusion of interpretability tools that help audit the model’s internal reasoning before it generates a response.

Cohere Command A+: A Fully Open‑Source, Efficient MoE Powerhouse

While Google’s announcement captured headlines, the AI landscape also welcomed a significant open‑source contender. On May 20, 2026, Cohere unveiled Command A+, a 218‑billion‑parameter decoder‑only Sparse Mixture‑of‑Experts (MoE) Transformer released under a permissive Apache 2.0 license—the first fully open‑source model from Cohere.

Only 25 billion parameters are active during any generation step, dramatically reducing the compute footprint. Command A+ is available in 16‑bit (BF16), 8‑bit (FP8), and a highly compressed 4‑bit (W4A4) quantization format. The W4A4 version achieves near‑lossless compression by quantizing only the MoE experts while keeping attention pathways at full precision, supplemented by Quantization‑Aware Distillation.

Performance metrics for the W4A4 quantization at low concurrency:

• 375 tokens per second (TOPS).
• Time‑to‑First‑Token (TTFT) latency of 113 ms.
• Up to a 63% increase in output speed and a 17% reduction in latency versus the prior Command A Reasoning model.

The new tokenizer supports 48 languages and improves tokenization efficiency for non‑English scripts—cutting tokens by 20% for Arabic, 18% for Japanese, and 16% for Korean—directly lowering operational costs for global deployments.

Command A+ was built with agentic workflows in mind. Benchmark leaps over the previous generation are stark:

• 𝜏²‑Bench Telecom (complex reasoning): 37% → 85%.
• Terminal‑Bench Hard (agentic coding): 3% → 25%.
• AIME 25 (complex mathematics): 57% → 90%.

While the model trails the largest Chinese open‑source rivals in pure agentic coding and broad‑scale intelligence, its hardware efficiency gives it a decisive edge for enterprises that need to run frontier‑grade AI on‑premises or in air‑gapped environments.

Crucially, Command A+ introduces native citation generation. When the model retrieves information from an external tool, it outputs explicit “grounding spans” that link every factual claim to the source document or database row. This traceability is invaluable for regulated industries such as finance, healthcare, and legal, turning interesting prototypes into production‑ready applications.

The model is also fully multimodal, capable of processing text and images within a 128K‑token context window, making it suitable for complex document tasks like analyzing scanned invoices, charts, or technical manuals.

By releasing Command A+ under Apache 2.0, Cohere offers true vendor independence: organizations can download the weights, fine‑tune them on classified internal data, and deploy them on private servers without being tethered to Cohere’s infrastructure, pricing, or API uptime—a realization of the “sovereign AI” thesis.

Autonomous Vehicles: From Regulatory Wins to Real‑World Challenges

Regulatory Milestones: Estonia Approves Fully Driverless Operations

In a notable step toward widespread driverless deployment, the Estonian transport authority granted approval to Bliq.ai for fully driverless road operations on public roads. Announced on May 22, 2026, the permit allows Bliq.ai’s autonomous vehicles to operate without a safety driver, marking one of the first such approvals in Europe.

The decision follows rigorous safety assessments and demonstrates that regulators are beginning to accommodate higher levels of autonomy when robust safety cases are presented.

Series‑Produced Robotaxis: XPeng’s First Off‑the‑Line Vehicle

On May 26, 2026, Chinese EV maker XPeng rolled the first series‑produced robotaxi off the assembly line. Unlike purpose‑built robotaxi platforms, XPeng’s vehicle is based on its flagship SUV, reducing development time and leveraging existing vehicle architecture.

The robotaxi is equipped with XPeng’s latest sensor suite and AI‑driving stack, positioning the company to compete in the rapidly growing robotaxi market.

Tesla Finally Launches FSD in China

After years of delay, Tesla confirmed on May 23, 2026 that its Full Self‑Driving (FSD) software is now available in China. Chinese rivals have already held Level 3 certifications and operated robotaxi fleets for some time, but Tesla’s entry signals intensifying competition in the world’s largest automotive market.

Initial reports indicate that the Chinese FSD package includes adaptations for local road conditions, signage, and driving behaviors, though regulatory approval for driverless operation remains pending.

Strategic Partnerships: ECARX and May Mobility Sign $750 M Robotaxi Deal

Also in May, ECARX** (the automotive tech arm of Geely) and May Mobility announced a ~$750 million agreement to supply thousands of purpose‑built robotaxi vehicles for deployment in the United States. The vehicles will be manufactured outside China to satisfy US compliance requirements, highlighting how global supply chains are adapting to geopolitical considerations.

Challenges and Setbacks: Weather, Construction Zones, and Safety Scrutiny

Despite progress, autonomous vehicle deployments continue to face real‑world hurdles:

• Waymo paused service in four Texas cities and Atlanta over flood‑related risks after robotaxis were observed driving into flooded roads (May 21, 2026).
• In another incident, Waymo halted freeway rides after its vehicles struggled in construction zones (May 21, 2026).
• Tesla Robotaxi crashes were blamed on human teleoperators, with backup‑system failures appearing twice in NHTSA data (May 26, 2026).

These episodes underscore that while the technology is maturing, robust operational design, continuous monitoring, and fallback mechanisms remain essential for safe at‑scale deployment.

Biotech Breakthroughs: Bridging AI‑Generated Design and Biological Construction

Sidewinder: A Revolutionary DNA‑Synthesis Method

Perhaps the most striking biotech development of the month is the introduction of Sidewinder, a new DNA‑synthesis technique described in IEEE Spectrum on May 26, 2026. Sidewinder addresses a long‑standing bottleneck: while generative AI models such as Evo 2 can design novel genetic sequences at extraordinary speed, physically constructing those sequences in the lab has remained slow, expensive, and error‑prone—especially when testing dozens of designs simultaneously.

The core innovation lies in tagging each oligonucleotide (oligo) with a unique molecular barcode, analogous to page numbers in a manuscript. When two bar‑coded fragments meet, they form a fleeting three‑way junction that locks the pieces in alignment before being cleanly removed, leaving a seamless strand. This approach eliminates the need for elaborate separation steps that plague conventional synthesis.

Key performance metrics:

• Error rate: approximately one incorrect junction per 10 million assembly events.
• Conventional methods misfire roughly once every 10 to 30 joins.
• Enabled the error‑free assembly of a 12,500‑letter E. coli genome redesign in a few days—a process that would likely take over a month with legacy methods.

The method was co‑developed by researchers at Caltech and Stanford, and later commercialized through a spin‑out called Genyro. While Genyro aims to serve pharmaceutical and biotech clients, the founders emphasize a desire to keep the platform broadly accessible to the academic research community.

Sidewinder’s impact is already being felt in downstream applications: it accelerates drug discovery, enables high‑density DNA data storage, and facilitates the construction of synthetic organisms. By marrying the speed of AI‑driven design with a reliable, low‑cost synthesis pipeline, Sidewinder effectively closes the loop between in‑silico generation and real‑world biological fabrication.

Gene Therapy Shows Promise Against TDP‑43‑Related Neurodegeneration

In neuroscience, a study published May 26, 2026 demonstrated that a novel gene‑therapy approach can shield the brain from TDP‑43‑mediated damage—a pathology implicated in amyotrophic lateral sclerosis (ALS) and frontotemporal dementia. The therapy, delivered via an AAV vector, reduced toxic TDP‑43 aggregates and preserved neuronal function in animal models, offering a potential avenue for future clinical translation.

Other Notable Biotech Advances

May 2026 also saw progress in several additional areas:

• Anti‑clotting medication: A global trial found that the investigational drug asundexian reduced recurrent stroke risk after non‑cardioembolic stroke or transient ischemic attack without increasing major bleeding risk.
• Liquid biopsy: The FDA approved Guardant Health’s new Guardant360 Liquid CDx panel, the largest FDA‑approved liquid‑biopsy assay to date, enabling comprehensive tumor profiling from a simple blood draw.
• PCSK9 base editor: A single dose of Eli Lilly’s VERVE‑102 base editor reduced PCSK9 by up to 88% and LDL‑C by up to 62%, with durable effects supporting its potential as a one‑time treatment for hypercholesterolemia.
• Stem‑cell therapy for Crohn’s disease: NeuroScientific Biopharmaceuticals reported an 80% clinical response rate in its stem‑cell program, clearing the way for Phase 2 trials.

Conclusion

May 2026 illustrates how disparate technology sectors are converging toward a future where intelligent systems are not only smarter but also more agentic, more autonomous, and more deeply intertwined with the physical world. AI models like Gemini 3.5 Flash and Cohere Command A+ are pushing the envelope of reasoning, tool use, and efficiency—while open‑source licensing choices are reshaping who can access and deploy frontier AI.

In transportation, regulatory approvals for driverless operations, series‑produced robotaxis, and Tesla’s FSD launch in China signal that self‑driving technology is moving beyond pilot projects into limited commercial rollouts, even as weather, construction zones, and safety concerns remind us that robustness remains a work in progress.

Meanwhile, biotech breakthroughs such as the Sidewinder DNA‑synthesis method are finally giving researchers the ability to turn AI‑generated genetic designs into tangible biological constructs at scale, accelerating fields ranging from drug discovery to data storage and synthetic biology.

Together, these advances paint a picture of a technological landscape where the barriers between conception, execution, and real‑world impact are rapidly dissolving. For developers, enterprises, and scientists, the message is clear: the tools to build the next generation of intelligent, autonomous, and biologically integrated systems are arriving now—and they are more accessible, powerful, and versatile than ever before.