The AI Race Is Not Slowing Down: Frontier Models, Physical Intelligence, and Robotaxis That Are Actually Here

Introduction

If you felt the last two weeks in tech were unusually dense, you are not imagining things. Between major model releases, a new open-weight architecture from an unexpected player, a physical-AI foundation model from NVIDIA, and robotaxi hardware that is finally built to scale, the signal-to-noise ratio in AI and mobility is high. At the same time, a small molecule from Oxford is quietly rewriting what immunotherapy can do for patients who have run out of options.

This edition of the tech roundup covers four stories that matter: the latest frontier model upgrades, a surprising open-weight challenger, the infrastructure behind physical AI, and the autonomous vehicles that are starting to earn real revenue. We also look at a biotech advance from ASCO 2026 that deserves the wonder pill label.

1. Claude Opus 4.8: Anthropic's Agent-First Upgrade

On May 28, Anthropic released Claude Opus 4.8, the successor to the widely adopted Opus 4.7. The pitch is not simply bigger benchmarks, but more reliable agency, faster collaboration, and cheaper throughput at the same price point.

The headline capability lift comes from agentic behavior. Early testers report that Opus 4.8 asks better clarifying questions, catches its own errors before they cascade, and carries multi-step tasks across long sessions without losing style or context. On Anthropic's Super-Agent benchmark, it is the only model to complete every end-to-end case, matching GPT-5.5 parity at the same cost. On CursorBench, tool calling is more efficient across every effort level: fewer steps, same intelligence. For legal work, it set a new high on Anthropic's Legal Agent Benchmark and became the first model to break ten percent on the all-pass standard.

Speed and Pricing Changes

Anthropic also introduced a fast mode that runs at 2.5x the speed of standard Opus 4.8, and cut its price to one-third of the previous fast-mode cost. For developers running cost-sensitive agent loops, the economics just improved meaningfully. Claude Code, meanwhile, gained a dynamic workflows feature aimed at very large-scale problems.

The browser and computer-use scores are worth calling out separately. Opus 4.8 scores 84% on Online-Mind2Web, a notable jump over Opus 4.7 and GPT-5.5. That matters for anyone building headless browsing agents, research assistants, or QA pipelines that need to verify live web content.

What is not changing is price. Opus 4.8 is available today at the same subscription and API price as Opus 4.7, which makes the upgrade effectively free for existing customers.

2. MiniMax M3: An Open-Weight Model That Brings Three Table Stakes Together

MiniMax released M3 on June 1, and the claim is specific: it is the first open-weight model that combines frontier coding, 1M-token context, and native multimodality in one package. That is a different angle from the usual we beat GPT on benchmarks announcement.

The architecture behind it is MSA, or MiniMax Sparse Attention. Instead of the quadratic complexity of standard full attention, MSA partitions the KV cache into blocks and uses an outer-gather approach that reads each block only once with contiguous memory access. At 1M tokens, MiniMax claims per-token compute falls to one-twentieth of the previous generation, with more than 9x speedup in pre-filling and 15x in decoding.

On benchmarks, M3 posts 59.0% on SWE-Bench Pro, surpassing GPT-5.5 and Gemini 3.1 Pro and approaching Opus 4.7. On SVG-Bench, it exceeds Opus 4.7. On OmniDocBench, it scores above Gemini 3.1 Pro. On Claw-Eval, an end-to-end agent evaluation framework, it achieves the highest score. None of these prove M3 is objectively best, but they show that the open-weight gap has narrowed sharply on tasks that matter most to developers: coding, document understanding, and end-to-end agent execution.

Why Open Weight Matters Here

The first and only open-weight model framing is strategic. Enterprises that cannot send code or customer data to closed APIs have been waiting for an open alternative that supports long context and native vision. If M3's performance holds under real-world load, it becomes a viable on-prem or private-cloud option for companies that previously chose between privacy and capability.

MiniMax is offering M3 through MiniMax Code, its token plan, and API services. That means there is already a path to test it without fine-tuning.

3. NVIDIA Cosmos 3: The Open Foundation Model for Physical AI

At GTC Taipei, NVIDIA launched Cosmos 3, an open world foundation model built for physical AI, including robots, autonomous vehicles, and vision agents that need to perceive, reason, and act in real environments. The headline architecture is a mixture of transformers: one branch reasons about vision and language, another generates video, audio, and action trajectories. The two are trained together on billions of text-image-video-sound-action samples.

The practical effect is that developers can use Cosmos 3 as a vision-language model, a world simulator, or an action-prediction engine, depending on the task. For robotics labs that previously had to stitch together separate perception and planning systems, Cosmos 3 offers a single pretrained starting point. NVIDIA says it reduces physical-AI training and evaluation cycles from months to days.

To push the ecosystem forward, NVIDIA also announced the Cosmos Coalition, a collaboration with Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI. The coalition is aimed at advancing open world models the way open LLM communities advanced text models: shared baselines, common benchmarks, and modular extensions.

For automotive developers, Cosmos 3 matters because it natively handles multimodal input and output while producing action trajectories. That is a closer match to how autonomous vehicles actually operate than text-only or image-only models can provide.

4. Waymo's Ojai: A Robotaxi Built for the Unit Economics of Scale

While the AI labs argue over benchmarks, Waymo is trying to solve a harder problem: building a robotaxi that costs less to manufacture and maintain than the human-driven cars it is replacing. On May 28, the Alphabet-owned company began limited public rides in Los Angeles, Phoenix, and San Francisco with the Ojai, a Zeekr-based minivan designed from the ground up for ride-hail.

The Ojai has a removable steering wheel, which is both a cost saver and a regulatory signal: Waymo can ship hardware that is cheaper without engineering around a driver-facing cockpit. The exterior is stripped of Chinese connected-car modules, and the underlying platform is Geely's SEA-M architecture, a variant built specifically for future mobility products such as robotaxis and logistics vehicles.

Waymo's sixth-gen sensor suite, 13 cameras, four lidars, six radars, and external audio receivers, is carried over from earlier platforms but is now modular. That means it can be fitted to the Zeekr minivan, the Hyundai Ioniq 5, and future vehicle types without custom redesign. Modularity is the hidden enabler of fleet scale.

Context matters: Waymo is simultaneously expanding and contracting. It suspended freeway robotaxi service in several cities to fix construction-zone behavior and paused operations in Atlanta and San Antonio because of flooding incidents. The Ojai launch is therefore both a product rollout and a credibility reset. If the unit economics improve, Waymo can justify fleet growth even if individual city rollouts require localized fixes.

5. Biotech Spotlight: GRWD5769, the Immunotherapy Enabler

Turning to biotech: at ASCO 2026 in Chicago, researchers from the University of Oxford and the Christie cancer centre in Manchester presented phase-one results for GRWD5769, a small-molecule pill taken in three-week cycles alongside the immunotherapy cemiplimab. The results were strong enough to draw genuine optimism from Cancer Research UK.

The problem the drug attacks is ERAP1, an enzyme on the surface of some tumors that effectively hides them from the immune system. Immunotherapy works by waking up the body's own defenses, but it fails in roughly two-thirds of patients because those defenses cannot see the tumor. GRWD5769 binds to ERAP1 and makes the cancer visible again, renewing and sustaining the immune response.

In a trial of 83 patients across 28 cancer centres in four countries, tumors shrank in 26 patients, with 15 achieving at least a 30% reduction. The effect was broad: cervical, liver, bladder, non-small cell lung, squamous cell carcinoma of the head and neck, and a bowel cancer subtype. The bowel-cancer cohort was particularly striking; 51% of patients saw disease progression halted for six months. In non-small cell lung cancer, 55% of patients hit the same milestone, and 38% did in head and neck cancer.

This is an early-stage study, and larger trials are required. But the breadth across hard-to-treat cancers, including in patients who had already stopped responding to prior immunotherapy, makes it one of the more compelling ASCO presentations this year.

6. What Ties These Stories Together

The common thread is not AI, cars, or medicine individually. It is the shift from general-purpose excitement to domain-specific infrastructure. MiniMax and Anthropic are competing on agent reliability and long-context efficiency, not raw parameter count. NVIDIA is shipping a foundation model that targets robots and AVs specifically, not just chatbots. Waymo is obsessing over per-unit manufacturing cost and modular sensor design. Oxford's GRWD5769 is not a broad-spectrum cure; it is a targeted fix for the precise mechanism that blocks immunotherapy in the majority of patients.

That pattern, identify the exact bottleneck, build for it, and ship, is what separates trending headlines from durable progress. The models, hardware, and drugs covered this week all passed that test. Whether they hold up at scale over the next year is the next question.

Sources

Anthropic. Introducing Claude Opus 4.8. anthropic.com, May 28, 2026.
MiniMax Research. MiniMax M3: Frontier Coding, 1M Context, Native Multimodality. minimax.io, June 1, 2026.
NVIDIA Newsroom. NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI. nvidianews.nvidia.com, May 31, 2026.
TechCrunch. Waymo's newest robotaxi is Chinese-made, built to make money, and now accepting riders. techcrunch.com, May 28, 2026.
The Independent. 'Wonder pill' found to reduce tumours by 30% in six cancer types. independent.co.uk, June 2, 2026.