A week of big AI launches: two new frontier models, one of them open, and the hardware that will run the next generation of agents.
GPT-5.5 takes on long, messy tasks and uses fewer tokens
OpenAI released GPT-5.5, which according to the company understands faster what the user wants and carries more of the work itself: coding, researching, analyzing data, creating documents and spreadsheets and operating software until the task is done. The gains are largest in agentic coding, computer use and knowledge work, with the same per-token latency as GPT-5.4. In Codex it uses far fewer tokens for the same tasks and, since April 24, it is also available in the API.
You can hand AI multi-step tasks with little specification, such as consolidating data from several systems or building reports. Higher efficiency also helps cut the cost of automations already running in production.
How Dexpert helps: Artificial IntelligenceSources
- OpenAI: Introducing GPT-5.5
DeepSeek V4 arrives open, with a 1 million token context by default
DeepSeek released the V4 preview with open weights in two versions: V4-Pro, with 1.6 trillion parameters and 49 billion active, and V4-Flash, smaller and cheaper. The company says a new form of attention greatly cuts memory and compute costs on long contexts, and 1 million tokens is now the default across all its services. The API accepts the OpenAI and Anthropic formats, and the models work with agents such as Claude Code and OpenCode.
Cheaper long contexts make it practical to analyze whole contracts, codebases and histories in a single call. Since it accepts the most common API formats, you can compare cost and quality by changing only the configuration.
How Dexpert helps: Artificial IntelligenceSources
- DeepSeek: DeepSeek V4 Preview Release
Google presents eighth-generation TPUs, one for training and one for inference
At Google Cloud Next, Google announced the eighth generation of its TPUs with two distinct chips: TPU 8t, for training large models, and TPU 8i, for low-latency inference aimed at AI agents. According to the company, both bring performance and energy efficiency gains over the previous generation. General availability is expected later this year.
Chips dedicated to inference tend to make cloud AI cheaper, which is where most companies spend. Those planning agent applications on Google Cloud can track TPU 8i's arrival to size costs.
How Dexpert helps: Cloud & ServerlessSources
Want to put one of these to work in your company?
Dexpert builds custom systems, apps, integrations and AI solutions. Tell us what you need and get an estimate, or take the free digital maturity assessment.