The Week AI Got Cheaper, Cars Got Smarter, and Gene Editing Hit the Newsstand

The New AI Race: Performance, Price, and Practicality

For the last two years, the AI model conversation has been dominated by scale — more parameters, bigger training runs, higher benchmark scores. This week made clear that the center of gravity is moving. The most consequential announcements were not about the largest model, but about models that are faster, cheaper, and more agentic — i.e., better at the work developers actually pay for.

The shift is structural. In 2024 and early 2025, the industry measured progress by leaderboard jumps: climbing MMLU, HumanEval, GSM8K, and their successors. That metric served its purpose, validating that training on trillions of tokens produced genuinely capable systems. But it also created a situation where developers chose between closed, expensive frontier models that passed hard tests and open, cheap models that failed them. The gap between those two poles defined the market.

This week, that gap narrowed from both ends simultaneously.

Claude Opus 4.8: The Agentic Bar Rises

On May 28, Anthropic released Claude Opus 4.8, upgrading its flagship with improvements across coding, agentic tasks, reasoning, and professional knowledge work. The headline numbers are broadly in line with what Anthropic has been saying about Opus for some time — this is a model built for long-horizon tasks where the agent must plan, execute, and self-correct multiple times before reaching a goal. But the concrete improvements are worth parsing.

On Anthropic's own Super-Agent benchmark, Opus 4.8 is the only current model to complete every case end-to-end, beating prior Opus models and matching GPT-5.5 at parity on cost. The CursorBench evaluations show meaningful gains too: tool calling is more efficient, using fewer steps for the same intelligence, and the model carries end-to-end tasks through rather than stalling partway. On the Legal Agent Benchmark, Opus 4.8 posted the highest score ever recorded, and is the first model to break 10% overall on the all-pass standard — the kind of accuracy lift that translates directly into how much attorney work can realistically be handed off.

What matters for practitioners, though, is the pricing and speed. Opus 4.8 introduces a "fast mode" (research preview on the Claude API) that runs at 2.5× standard speed. That fast mode is priced at three times less per token than it was for previous Opus generations. For teams running agents that need to reason reliably before acting, that combination of speed and improved judgment is the real upgrade, not the benchmark headline.

There are also product-level changes: users on claude.ai now have control over the amount of effort Claude puts into a task, and Claude Code gains a new "dynamic workflows" feature that allows it to tackle very large-scale problems by decomposing them. These are usability improvements layered on top of capability gains, and they reduce the operational friction of deploying agents in production.

MiniMax M3: Open Weights Meet Frontier Performance

MiniMax's M3 release on June 1 was the bigger structural story. The Chinese AI startup shipped a model with a 1-million-token context window, native multimodality (text, image, audio, video in one forward pass), and coding and agentic performance that VentureBeat reported eclipses GPT-5.5 and Gemini 3.1 Pro on selected benchmarks. Critically, MiniMax is pricing M3 at a special introductory rate of $0.30 per million input tokens and $1.20 per million output tokens — roughly 5–10% of the cost of leading U.S. proprietary models.

Even at full price — $0.60 input, $2.40 output — M3 remains at 8–20% the cost of leading proprietary U.S. models. That is not a minor discount. For teams processing millions of tokens daily, the savings compound quickly. More importantly, MiniMax announced that M3 weights will be released under an open-source license with full enterprise download rights within ten days. Until then, developers have API access; soon after, they can run the model on their own infrastructure, fine-tune on proprietary data, and ship without ongoing API dependency.

This matters because it collapses the traditional trade-off. Software developers have historically chosen between top-tier closed-source intelligence behind restrictive APIs and nimble, cost-effective open models that falter on multi-step reasoning, dense coding tasks, and massive data sequences. MiniMax-M3's MSA (Mixture of Sparse Attention) architecture is designed explicitly to handle long-context reasoning without the quadratic cost of traditional attention mechanisms, which is how it delivers both 1M-token context and competitive performance per dollar.

It is worth noting that MiniMax is not the first to challenge U.S. model pricing — Xiaomi's MiMo-V2.5 Flash and DeepSeek's v4-flash have also competed aggressively on cost — but M3 is the first to do so while also claiming frontier-tier coding and agentic performance and announcing open weights. That trifecta is what has the industry talking this week.

NVIDIA Cosmos 3: The Physical AI Foundation

Meanwhile, at NVIDIA's GTC Taipei event running alongside COMPUTEX, the company launched Cosmos 3, which it bills as the first open frontier foundation model for physical AI — systems that interact with the real physical world, including robots, autonomous vehicles, and smart spaces. The core problem Cosmos 3 addresses is that physical AI systems need more than perceptual recognition. They need to understand cause and effect, predict how scenes evolve, and generate the control signals that make things happen.

Cosmos 3 combines vision reasoning and multimodal generation across text, video, images, ambient sound, and action in a single model. Its mixture-of-transformers architecture separates reasoning and generation into distinct blocks: a reasoning block first interprets what is happening in a scene, and a generation block uses that context to produce physically grounded outputs — synthetic video for training, robot-task data, trajectory predictions. Critically, Cosmos 3 can output numerical action data: joint angles, gripper positions, trajectory points. That makes it directly useful for robot manipulation and autonomous driving pipelines that need structured control signals, not just images.

NVIDIA is releasing Cosmos 3 as open weights on Hugging Face and through its developer portal. The company is positioning it as a world foundation model — the same category of generalist-world understanding systems that researchers have been exploring for years, but in a form that is actually usable for commercial development. For autonomous vehicle companies, Cosmos 3 offers a standardized way to simulate edge cases without expensive real-world collection. For robotics teams, it provides a generative substrate for creating training scenarios that correctly obey physical constraints.

Autonomous Vehicles: From Prototype to Product

The AI model progress above is directly enabling a shift in the autonomous vehicle space. After years of prototypes and limited pilots — and after a period in 2025 when investor patience with robotaxi timelines wore thin — the last week brought signs that autonomous fleets are moving toward scalable, revenue-generating operations. The announcements were not about breakthrough capabilities; they were about cost and partnership architecture.

Waymo's Ojai: Cost Engineering at Scale

Waymo announced on May 28 that its newest robotaxi, the Ojai — a modified Zeekr-made electric minivan — is now open to select riders in Los Angeles, Phoenix, and San Francisco. The Ojai is significant less for its autonomy stack than for its vehicle design philosophy. Waymo has explicitly built the Ojai for fleet economics: it is roomier, has a removable steering wheel (reducing manufacturing complexity and cost), and is engineered to withstand hundreds of thousands of rides with lower maintenance overhead than prior models.

The launch comes with some operational caveats. Waymo has recently suspended robotaxi service on freeways in Los Angeles, Miami, Phoenix, and San Francisco to refine how its vehicles handle construction zones — a reminder that robust highway autonomy remains a hard problem. And access to the Ojai remains limited to a small cohort of riders providing feedback, with broader expansion planned later.

Still, the Ojai is the right framing for the business: Waymo's challenge is no longer proving the technology works in demo conditions, but proving the unit economics can sustain a broad deployment. A robotaxi that costs less to manufacture and can run continuously without a safety driver is the actual product. The Ojai is that product, even if the rollout timeline is measured in years, not months.

Nuro, Lucid, and Uber: A Premium Robotaxi Stack

While Waymo scales toward cost-efficient mass-market fleets, Nuro is building a different proposition. In partnership with Lucid and Uber, Nuro is developing a premium robotaxi based on the Lucid Gravity SUV — a vehicle with genuine luxury credentials, not a purpose-built robotaxi pod. Nuro secured a California DMV driverless testing permit in May and a CPUC drivered pilot permit shortly after, clearing the regulatory path for passenger service in the Bay Area. Uber will provide the dispatch, payment, and rider network infrastructure.

This three-party arrangement is notable because each participant contributes a distinct competency without duplicating the others. Lucid handles the EV platform, Nuro writes the autonomy software, and Uber manages the consumer-facing service. That division of labor mirrors how other transportation industries matured — OEMs build vehicles, Tier 1 suppliers provide critical systems, and platform operators connect them to riders. If it proves scalable, it sets a template for how autonomous services might be structured commercially: not as a vertically integrated mega-project, but as a stack of specialized partners.

Xiaomi and the World Model Play

In China, Xiaomi EV announced its own world model to advance autonomous driving capabilities. World models — systems that construct an internal simulation of the environment to predict outcomes and plan actions — represent a significant architectural bet. Rather than training a perception model to label objects on each frame, a world model builds a generative understanding of the scene that can answer "what happens next" questions.

Xiaomi's approach is consistent with how Chinese EV makers are treating autonomous driving as a software-first competitive advantage. The company has been steadily expanding its smart-driving feature set across its vehicle lineup, and the world model announcement signals deeper investment in the causal reasoning layer. This matters for the broader ecosystem because Chinese manufacturers collectively represent the world's largest EV production base, and their software investments will shape the global benchmark for what 'mass-market autonomous' means.

Biotech: Base Editing Goes Clinical, and the Cancer Test Gets Bigger

The biotech frontier moved in two meaningful directions this week. A landmark gene-editing trial result was published in The New England Journal of Medicine, and a large-scale cancer detection study confirmed the expanding utility of liquid biopsies at population scale.

VERVE-102: Editing Cholesterol in the Body

The NEJM publication on May 25 detailed the Phase 1b Heart-1 trial of VERVE-102, an in vivo base editor developed by Verve Therapeutics and now advancing under Eli Lilly's ownership. The therapy uses a precise base editing approach — essentially swapping one DNA letter for another without creating the double-strand breaks that conventional CRISPR induces — to permanently modify the PCSK9 gene and lower production of the PCSK9 protein, which regulates LDL cholesterol levels.

The results were both statistically meaningful and clinically suggestive. In the Phase 1b study, a single intravenous infusion reduced PCSK9 protein levels by up to 88% and LDL cholesterol by up to 62%, with effects that were durable across the study period's follow-up duration. For context, current standard-of-care PCSK9 inhibitor injections require dosing every two to four weeks indefinitely. A one-time infusion with sustained lipid-lowering effect would fundamentally change the treatment paradigm for hypercholesterolemia, particularly in patients who struggle with medication adherence or who face cardiovascular risk despite maximum tolerated statin therapy.

Eli Lilly is reportedly preparing a Phase 2 trial to test VERVE-102 in a broader patient population, which would be the critical next step in establishing both efficacy and safety at scale. Base editing is still an early platform — off-target effects, immune responses to delivery vehicles, and long-term genomic stability all need careful tracking. But the NEJM publication moves the technology from 'promising in animals' to 'showing meaningful human data,' which is the threshold that separates interesting science from investable medicine.

GRAIL's PATHFINDER 2: Liquid Biopsy at Population Scale

At the 2026 ASCO Annual Meeting on May 31, GRAIL presented PATHFINDER 2 results drawn from more than 35,000 participants using its Galleri multi-cancer early detection (MCED) test. Galleri is a liquid biopsy — blood draw, no imaging — that analyzes methylation patterns in cell-free DNA to detect signals from dozens of cancer types, many of which lack established screening protocols today.

The PATHFINDER 2 data showed that Galleri substantially increased overall cancer detection rates compared to standard care pathways, while maintaining a favorable safety profile and robust specificity. The 35,000-participant scale is the story here. Earlier MCED studies —— including GRAIL's own PATHFINDER 1 — raised legitimate questions about false-positive rates and overdiagnosis in healthy populations. These results, presented at ASCO's major plenary, provide the largest real-world evidence that the signal holds up at scale without generating unmanageable false-positive cascades.

If Galleri gains broader clinical and payer adoption, it could establish a new standard: a single annual blood test that screens for multiple cancer types simultaneously, something that currently requires a patchwork of colonoscopies, mammograms, low-dose CTs, and Pap smears. That is not a replacement for targeted screening, but it fills the gap for cancers that have no good screening test today — pancreatic, ovarian, and esophageal among them.

Abbott's Dual Sensor and Adjacent Milestones

Abbott received CE Mark approval for what it describes as the world's first dual glucose-ketone sensing technology for people with diabetes. The system simultaneously tracks glucose and ketone levels from a single sensor — clinically useful because ketone monitoring helps diabetic patients manage the risk of diabetic ketoacidosis, particularly during illness or periods of low carbohydrate intake.

In related developments, Voyager Therapeutics secured FDA IND clearance for VY1706, a gene therapy designed to reduce tau production in the brain for Alzheimer's disease — the first gene therapy approach aimed at tau reduction. And Shionogi gained FDA approval for XOCOVA (ensitrelvir), a post-exposure oral prophylactic for COVID-19, becoming the first agent in that category. Taken together, these approvals and clearances show a therapeutic pipeline that is expanding beyond narrow disease-modifying indications into preventive and management use cases — a sign that the biotech industry is finding commercial models for broader patient populations, not just rare-disease orphan drugs.

The Connecting Thread: Practicality Over Hype

Read together, these developments point to a single underlying theme: the technology sector is rewarding practicality over hype. In AI, the race is no longer purely about benchmarks; it is about cost per useful token, speed of inference, and the reliability of agents operating across complex workflows. MiniMax M3 and Claude Opus 4.8 represent two answers to that question — one from the open-weights, cost-disruption angle, the other from the agentic reliability angle. Both are credible.

In autonomous vehicles, the transition from 'look, no driver' to 'look, profitable rides' is the actual milestone. Waymo's Ojai is engineered for that transition, and the Nuro-Lucid-Uber premium stack is a commercial architecture that could generate revenue before mass-market robotaxis do. Neither story is finished, but both are more concrete than the autonomous vehicle conversations of two years ago.

And in biotech, VERVE-102's NEJM data and GRAIL's 35,000-person cancer detection results are the kind of clinical evidence that moves markets. Base editing has been a laboratory curiosity and a venture capital theme; now it has a Phase 2-ready human trial result. Liquid biopsies have been a 'maybe in the future' screening tool; now they have a large-scale randomized-adjacent evidence set.

None of these represent singular revolutions. They represent an industry-wide maturation: AI getting cheaper and more reliable, autonomous systems getting closer to revenue, and gene therapies getting closer to standard care. That is not a flashy narrative for a keynote stage, but it is exactly the phase of technology development that reshapes markets over years, not weeks.

Sources

Anthropic — Claude Opus 4.8 announcement and system card
MiniMax — MiniMax M3 release and pricing
VentureBeat — MiniMax M3 benchmark and cost analysis
NVIDIA — Cosmos 3 launch announcement
NVIDIA Technical Blog — Cosmos 3 Physical AI development guide
Hugging Face — Cosmos 3 for Physical AI overview
TechCrunch — Waymo Ojai robotaxi launch
CNBC — Waymo Ojai fleet cost strategy
Nuro / Uber / Lucid — robotaxi partnership announcement
TechCrunch — Nuro driverless testing permit
Xiaomi / CnEVPost — Xiaomi world model for autonomous driving
NEJM — VERVE-102 Phase 1b Heart-1 trial
PR Newswire — Lilly VERVE-102 cholesterol reduction data
Fierce Biotech / BioPharma Dive — VERVE-102 Phase 2 preparation
Ars Technica — VERVE-102 coverage
GRAIL — PATHFINDER 2 ASCO 2026 results
Abbott — dual glucose-ketone sensing CE Mark
BioSpace — Voyager VY1706 FDA IND clearance
Shionogi — XOCOVA FDA approval