Gemini 3.5 Flash gains computer use as OpenAI unveils Jalapeño

Google introduced computer use as a built-in tool for Gemini 3.5 Flash on June 24, 2026. The model can inspect screens, plan actions, and operate software interfaces. Developers and enterprises can access the capability through Google's platforms.¹

The update brings graphical interface control into the Flash model rather than relying on a separate computer-use model.¹² That matters for agents designed to carry out tasks in applications, not just describe the steps. Google calls it its best performance yet on agentic computer-use tasks, but the announcement does not include a detailed benchmark table.¹ The available information also does not establish pricing or regional availability.¹ Businesses will still need to test how reliably it works in their own applications.

OpenAI and Broadcom unveiled Jalapeño on June 24, 2026, describing it as OpenAI’s first custom chip for running large language models.¹ OpenAI designed the processor and worked with Broadcom to bring it to production.¹² The goal is to make ChatGPT, Codex, and the API cheaper and more efficient to operate.² Jalapeño is not yet widely deployed; initial deployment is planned by the end of 2026.¹³

The chip gives OpenAI a way to shape more of the infrastructure used to run its models.¹ Whether that leads to lower API prices remains unclear. Performance and cost claims come from the companies’ early testing, not independent comparisons.³⁴ Public benchmarks and detailed technical specifications have not been released.³⁵ It is therefore too soon to assess how Jalapeño will perform against other AI chips in production.

On June 22, 2026, OpenAI expanded its Daybreak security initiative with an updated Codex Security plugin and the full release of GPT-5.5-Cyber for trusted defenders.¹² Codex Security is intended to find plausible vulnerabilities, determine whether they are reachable, gather supporting evidence, and develop and verify targeted patches. That moves the tool beyond reporting potential issues.¹

GPT-5.5-Cyber is designed to analyze large codebases more deeply, validate findings in controlled settings, and help develop and test patches. Access is restricted to trusted defenders.²³ OpenAI also introduced Patch the Planet to support vulnerability fixes in open-source projects.⁴ For companies, the practical question is whether validated findings and tested fixes can move reliably into existing security workflows.

Google made the Interactions API generally available on June 22, 2026, calling it its primary interface for Gemini models and agents.¹ The release adds Managed Agents for orchestration and background execution for longer-running tasks.¹² The API brings model and agent interactions into a single endpoint, with support for tool use, server-side state, and multimodal generation.¹

For development teams, the shared interface offers a way to integrate model and agent capabilities without building separate connections for each.¹ General availability applies to the Interactions API, however, not necessarily to every planned capability. Google said support for Gemini Omni was still forthcoming at the time of the announcement.¹ Teams migrating existing applications should check which capabilities are already available before relying on them in production.

OpenAI announced an updated version of GPT-5.5 Instant on June 24, 2026. The company says it better recognizes the intent behind a question and adjusts its response accordingly. It should also handle requests with multiple constraints more reliably. The ChatGPT rollout is staggered across paid and free users.¹

OpenAI also says the model gives better shopping and local recommendations and is more enjoyable to talk to.¹ GPT-5.5 Instant is ChatGPT’s default model, according to Engadget.² Changes to its responses therefore affect a central part of the ChatGPT experience rather than a specialized feature. The size of the improvement cannot be independently assessed from the announcement: OpenAI provided no new benchmark results for this update.¹

OpenAI announced a limited preview of GPT-5.6 on June 26, 2026. Its flagship model, Sol, is initially available only to selected partners and organizations through the API and Codex, not to individual users through ChatGPT.¹² The family also includes Terra and Luna. Sol has an “ultra” mode that delegates difficult tasks to coordinated subagents.¹³

OpenAI positions Sol for demanding work in areas including coding, biology, and cybersecurity. Terra is the balanced option, while Luna is intended as a lower-cost choice for high-volume use.¹² This preview is not a general release: public self-service access is not yet available. OpenAI says broader availability will follow in the coming weeks but has not given a specific date.¹²

OpenAI introduced Record & Replay for the Codex desktop app on macOS on June 18, 2026.¹² Users can demonstrate a workflow, which Codex turns into a reusable skill.¹³ The resulting skill can be inspected and run again later.³ The feature is initially available only with certain ChatGPT subscriptions and is not offered in every region.⁴²

This gives users a way to capture recurring work by demonstrating it rather than relying only on written instructions.¹³ Skills can be edited after recording, allowing companies to review them before reuse.³² The description does not establish that Codex retrains its base model from a demonstration. It describes a saved, reusable workflow.¹³

Alibaba has introduced Qwen-AgentWorld, a language model designed to simulate changes in digital environments after an AI agent takes an action.¹ Weights for Qwen-AgentWorld-35B-A3B and the accompanying AgentWorldBench were released on June 24, 2026.² The model covers MCP, Search, Terminal, SWE, Web, OS, and Android.¹² Rather than serving as a conventional chatbot, it is intended to predict how an environment might respond.¹

The approach could help developers study agent behavior using simulated feedback from tools and interfaces.¹ Qwen trained the model in multiple stages using real interaction trajectories.¹ AgentWorldBench tests tasks in the supported domains, but it does not establish how reliably the simulation will behave in a particular company's workflow.² That requires testing against the company's own processes and systems.

Krea released the weights for Krea 2 Raw and Krea 2 Turbo on June 23, 2026.¹² Raw is an undistilled mid-training checkpoint intended for fine-tuning, post-training, and LoRA training.² Turbo is distilled and post-trained for fast text-to-image generation.²

Krea positions the checkpoints as complementary: developers can train on Raw and use Turbo for image generation.¹² That gives developers separate starting points for adapting an image model and running it.² According to the technical report, the weights and inference are released under a permissive license.³ It is a custom license, however, rather than a standard open-source license.²³