Google, Anthropic and AMD announce new AI models and chips

In May and July 2026, Google, Anthropic and AMD each made major announcements in the areas of AI models and hardware. The three companies each took a further step in their own way toward more autonomously operating AI systems, for both consumer and enterprise applications.

Google focused at Google I/O on so-called agentic AI, in which models independently execute multi-step tasks. Anthropic lowered the barrier to top-tier performance by bringing its new standard model closer to the premium level. AMD presented new AI infrastructure that it claims outperforms competing systems in compute power and memory.

Google I/O 2026: Gemini 3.5 Flash and Gemini Spark

Google held Google I/O 2026 on 19 and 20 May in Mountain View, California, with the main keynote live-streamed from the Shoreline Amphitheater. The central message was that AI models are increasingly able to handle tasks autonomously, without requiring user input at every step.

The first concrete model is Gemini 3.5 Flash, which became generally available on 21 May 2026 via the Gemini API. It is the first model in the new Gemini 3.5 family and is aimed at fast, efficient inference. Alongside Gemini 3.5 Flash, Google also announced Gemini Omni, a multimodal model, and the AI agent Gemini Spark.

Google further presented a significant overhaul of Google Search through so-called Search Agents and Generative UI, in which the search engine generates answers and can execute follow-up actions. The company also unveiled AI-powered audio glasses, developed together with Samsung, Warby Parker and Gentle Monster, with an expected launch in autumn 2026.

Anthropic: Claude Sonnet 5 as new standard model, Opus 5 for heavier workloads

Anthropic released two new models in June and July 2026 that narrow the gap between its mid-range and top tier. Claude Sonnet 5 appeared on 30 June 2026 and replaced the previous standard model for both free and paid users on Claude.ai. The model is also available via the Claude API.

Sonnet 5 is designed for what Anthropic itself describes as agentic tasks, including coding, tool use, browser automation and more extensive knowledge work. In terms of performance, Anthropic states that the model comes close to Claude Opus 4.8 on reasoning, tool use and coding. The introductory price was $2 per million input tokens and $10 per million output tokens until 31 August 2026; after that, the rates are $3 and $15 respectively.

On 24 July 2026, Claude Opus 5 followed as the new premium model. Anthropic positions it as the strongest model on Claude Pro and the default model on Claude Max. According to the company, Opus 5 scores higher than all competitors on the Frontier-Bench and achieves three times the score of the nearest competitor on the ARC-AGI 3 benchmark. On CursorBench it performs within 0.5 percentage points of the previously released Fable 5, but at half the cost. The price is set at $5 per million input tokens and $25 per million output tokens.

A notable characteristic of Opus 5 is what Anthropic calls agentic persistence: the model verifies its own output and repeats steps until a task has actually been completed. Both Sonnet 5 and Opus 5 have a context window of 1 million tokens and a maximum output size of 128,000 tokens.

Earlier, on 9 June, Anthropic also released Claude Fable 5 at $10 per million input tokens. This Mythos-class model was temporarily withdrawn on 12 June due to a US export control order and made available again on 1 July 2026.

AMD Advancing AI 2026: Helios rack with 2.9 exaflops

On 23 July 2026, AMD held its annual AI event at the Moscone Center in San Francisco. CEO Lisa Su opened the keynote and called it the largest event in the company's history.

The central announcement was Helios, an AI rack that AMD describes as fully in production and set to ship by the end of the third quarter of 2026. Helios combines 72 Instinct MI455X GPUs with sixth-generation EPYC processors and Pensando networking components via open UALink-over-Ethernet. The system delivers 2.9 exaflops of FP4 compute.

AMD claims that Helios outperforms competing systems by at least 15 percent on large models, offers 50 percent more HBM memory and achieves up to 30 percent higher throughput. These claims originate from AMD itself; independent verification was not yet available at the time of publication.

For developers and infrastructure teams considering alternatives to Nvidia hardware, Helios represents a concrete offering with known specifications. The combination of open interconnect standards and high HBM capacity addresses the need for scalable compute clusters for training and running large language models.

What these announcements mean for AI builders and investors

The three announcements each address a different layer of the AI stack. Google focuses on the application layer, with models and agents being embedded directly into consumer products such as Search and the Gemini app. With Sonnet 5 and Opus 5, Anthropic shifts the boundary of what a standard model can do: those who previously needed a premium subscription for demanding reasoning and coding tasks can now get further with the lower-cost model.

AMD's Helios announcement operates at the infrastructure level. With an alternative to Nvidia hardware in production and a delivery date in the third quarter of 2026, hyperscalers and large enterprise customers have a concrete offering to evaluate. Whether the performance claims hold up in practice will need to be confirmed by independent benchmarks.

For teams currently building AI products, the models from Anthropic and Google primarily mean that the cost per task is falling while capacity is rising. Larger context windows of 1 million tokens make it possible to process more extensive documents, codebases or conversations in a single session, opening up new applications for sectors such as legal services, software development and data analysis.