OpenAI launches ChatGPT Work as Google unveils Gemma 4

OpenAI introduced ChatGPT Work on July 9, 2026. The agent in ChatGPT is powered by Codex and GPT-5.6 and is designed to work across apps and files. OpenAI says it can stay with a project for hours and produce documents, slides, spreadsheets, or web apps. Access on web and mobile began with a phased rollout by plan.¹²

The product brings chat and coding workflows closer together: ChatGPT Work is meant to carry out multiple steps toward a usable result rather than only provide answers or code.¹ For companies, access differs by platform. Work is available on more plans in the desktop app than it was at the start of the web and mobile rollout.¹² Whether its output suits a particular workflow still needs to be tested against that workflow’s requirements.

Google presented Gemma 4 in a technical report in early July. The report was submitted on July 2, and Google's Gemma team highlighted it on July 7.¹² The open-weight family spans 2.3B to 31B parameters. Its models handle multiple modalities and use both dense and Mixture-of-Experts architectures.¹ A model card for the 31B variant is available on Google's Hugging Face page.³

Google reports improvements in compute efficiency, inference speed, memory use, and long-context tasks. The technical report also describes gains on STEM and multimodal benchmarks.¹ For organizations evaluating the models, those claims are a starting point rather than a substitute for testing: context windows differ across variants.³ Teams should assess which model fits their workloads and available hardware.

Meta introduced Muse Spark 1.1 and the Meta Model API in a public preview for developers on July 9, 2026.¹² Meta describes the multimodal model as built for agentic tasks, with improvements in tool and computer use, coding, and multimodal understanding.¹ It is available in “Thinking” mode in the Meta AI app and on meta.ai, while external developers can access it through the new API.¹²

The release adds API access with tool and function calling to the Muse Spark family.² It also puts Meta more directly into the market for AI-assisted coding and automated workflows. Reuters describes its token prices as competitive with OpenAI and Anthropic, though not uniformly the cheapest.³ Meta reports improvements across several performance areas; companies will still need to test the model against their own tasks and tools.²

On July 8, 2026, xAI and Cursor introduced Grok 4.5, a jointly developed model for coding and other AI agent tasks.¹² At launch, it was available in Cursor on desktop, web, and iOS, as well as through its CLI and SDK.¹ xAI also offered access through Grok Build and its developer console.¹²

According to Cursor, Grok 4.5 builds on a new pretrained mixture-of-experts model that was then trained further on trillions of tokens of Cursor data.¹ The companies intend it to handle broader agent tasks, not just software engineering.¹ Published token prices for the base model and a faster variant let teams compare costs before deployment.¹ Whether either option is reliable enough for a particular workflow still requires testing against that team’s own tasks.

Tencent released its Hunyuan model Hy3 on July 6, 2026, and made the weights available on platforms including Hugging Face and ModelScope.¹² Tencent positions the mixture-of-experts model for agent workloads and integration into products.¹² Under this architecture, only a portion of the model's parameters is activated for each token.²

That can limit compute per token, but it does not guarantee lower total operating costs.² Tencent says Hy3 approaches the performance of larger open models in some comparisons; companies should test that claim against their own workloads.¹² The cost of hosted access is also separate from the license for the model weights: a free access offer on OpenRouter was time-limited.²³

Meta introduced Muse Image on July 7, 2026, calling it the first image-generation model from Meta Superintelligence Labs.¹² The rollout in the Meta AI app and on meta.ai began that day.¹ Features powered by the model are also available in Instagram Stories in the US and WhatsApp in select countries.¹³ Meta did not announce a worldwide launch for those features.¹

Adding image generation to Stories and chats puts the model inside apps people already use.¹ Meta says everyday use is free, while heavier creation falls under its subscription plans.¹ It has not published specific prices or usage limits.¹ That leaves costs for sustained, high-volume use unclear for businesses. Meta says support for Facebook and Messenger will follow later.¹

OpenAI made gpt-realtime-2.1 and gpt-realtime-2.1-mini available through its Realtime API on July 6, 2026.¹ The mini variant supports real-time voice interactions, reasoning, and tool use.¹² OpenAI positions it as a faster, lower-cost option alongside the more capable model, which it emphasizes for reasoning, tool use, and following instructions.¹

The models target voice and multimodal applications where response times and operating costs matter.¹ OpenAI says improved caching has reduced p95 latency across its Realtime voice models by at least 25%; that figure is not a separate measurement of the mini variant’s improvement.¹ The announcement does not include benchmark tables that would allow developers to compare performance in specific applications.¹

GitHub made Moonshot AI’s Kimi K2.7 Code generally available in GitHub Copilot on July 1, 2026.¹ GitHub calls it the first open-weight model users can select in Copilot’s model picker.¹ It initially rolled out to Pro, Pro+, and Max users; on July 7, GitHub extended availability to Copilot Business and Enterprise.²¹ The model is off by default for those organizational plans and requires an administrator to enable it.²

GitHub hosts the model on Microsoft Azure and charges for its use at the provider’s list prices under usage-based billing.² Open weights therefore do not mean that Copilot runs the model locally.² GitHub presents Kimi K2.7 Code as a lower-cost option with strong performance, but its announcements provide no benchmark figures to support that assessment.²¹ Companies considering it should test representative coding tasks and check their usage costs before switching models.

Google Cloud made AlphaEvolve generally available on July 9, 2026.¹ Developed with Google DeepMind, the agent is part of the Gemini Enterprise Agent Platform on Google Cloud.¹² Google says it uses Gemini to develop algorithms and optimize code. Companies can also use it to address infrastructure bottlenecks, according to the launch announcement.¹²

The release expands access beyond an earlier testing phase.¹ Google had previously described internal uses of AlphaEvolve to optimize its own systems.³ General availability now brings the agent to a broader set of cloud users.¹² It remains unclear whether the reported improvements to code and infrastructure can be reproduced reliably outside Google’s examples: the launch materials do not provide independent evidence of those results.¹³

In early July, Nvidia introduced a model intended to give AI startups access to compute through partner clouds.¹² Cloud providers buy Nvidia hardware and rent the resulting capacity to startups and other customers.³² Nvidia receives hardware sales revenue as well as a share of the cloud revenue generated by that capacity.³² Access therefore comes through cloud partners, rather than a program that gives startups free GPUs.¹³

The arrangement could give Nvidia recurring revenue alongside its existing hardware business. For customers, rented capacity can offer an alternative to building their own infrastructure.³² But the economics are difficult to assess from the public information: Nvidia has not disclosed its revenue-share percentage or full contract terms.⁴⁵ The available details do not establish that startups can use the GPUs at no cost.¹³