AI Weekly: Claude Opus 4.7 Rumors, New GPT Image Model, MiniMax M2.7, and More

The AI world is moving fast again this week, with major updates from Anthropic, OpenAI, and several open(-ish) source projects. From rumors of a nerfed Claude Opus 4.6 and a possible 4.7 upgrade, to a powerful new GPT image model and MiniMax’s latest release, there’s a lot to unpack.

Claude Opus 4.6 Concerns and a Possible 4.7 Upgrade

Many users have noticed that Claude Opus 4.6 doesn’t feel as sharp as it did a few weeks ago. Reports describe less consistent reasoning, more hallucinations, and weaker overall output quality. On top of that, higher-tier subscribers are hitting rate limits faster than before, even though their usage patterns haven’t changed much.

Fresh benchmark data backs up these impressions. Bridge Bench, a popular evaluation platform, recently re-tested Opus 4.6 on a hallucination benchmark. The model dropped from 2nd place with 83.3% accuracy to 10th place with 68.3% accuracy in a very short time. That’s a big swing, suggesting something meaningful has changed behind the scenes.

There’s no official confirmation from Anthropic that they’ve deliberately “distilled” or weakened Opus 4.6. However, internal API references have reportedly started mentioning a new Claude Opus 4.7 model. That’s usually a sign that a release is close. It’s possible that Anthropic is rebalancing infrastructure and costs in preparation for the next generation, which can temporarily affect current models.

Some users are also talking about an unofficial “token tax” in Claude Code. Captured HTTP traffic suggests that newer Claude Code versions may be injecting around 20,000 extra tokens server-side per request, even when the visible input is small. This would explain why people are hitting token limits much faster. A common workaround is downgrading to an older Claude Code version via a command like npx claude-code@2.1.98, which some users say avoids the extra hidden tokens.

All of this is happening against the backdrop of Anthropic’s broader roadmap, including the rumored Claude Mythos model. For more context on where Anthropic might be heading, see this deep dive into the upcoming Claude Mythos model.

Anthropic’s New Full‑Stack App Studio and Word Integration

Anthropic isn’t just shipping models; it’s building out a full developer ecosystem. The company is reportedly working on an AI Studio–style platform similar to Google AI Studio. The goal: give developers a full-stack environment where they can build, test, and deploy AI-powered apps end-to-end from a single interface.

This would put Anthropic into more direct competition with tools like Lovable and Replit, and it aligns with a broader trend of AI platforms trying to become one-stop shops for app development rather than just API providers.

On the tooling side, a new Claude Code desktop update has been spotted with a more unified interface, letting you work across multiple repositories in a single instance. That’s a big quality-of-life improvement for developers juggling several projects.

Anthropic is also moving deeper into productivity workflows with Claude for Word, now in beta. This integration lets you draft, edit, and revise directly inside Microsoft Word via a sidebar. It preserves formatting and uses track changes for edits, so it feels like a native part of Word instead of a copy-paste workaround. At the moment, it’s limited to team and enterprise users.

OpenAI: GPT Image Gen 2 and a New ChatGPT Pro Tier

On the OpenAI side, there are strong signs that a new image model, GPT Image Gen 2, is about to launch. OpenAI is reportedly A/B testing it inside the ChatGPT app right now, which usually means a public rollout is close—possibly as soon as Tuesday or Thursday, based on their typical release cadence.

Early examples show a big jump in image quality and prompt adherence. The model can generate highly detailed, consistent scenes, like a full “GPT-6” main menu UI mockup or a realistic Slack-style chat where Anthropic employees debate which lab poses a greater risk. The key improvement is not just realism, but how well the images follow complex, multi-part prompts.

OpenAI is also introducing a new $100/month ChatGPT Pro subscription tier aimed at heavy coders and power users. This plan offers roughly 5x the Codex usage of ChatGPT Plus, temporarily boosted to 10x through May 31. You still get all existing Pro features—access to higher-end models and more intensive reasoning modes—but with far more generous coding limits.

Given the current frustration around Claude Code rate limits, this new Pro tier lands at a very strategic time. For developers who live inside AI coding assistants, it may be the most attractive option on the market right now.

MiniMax M2.7: Powerful Model, Problematic License

MiniMax has released its M 2.7 model and is promoting it as fully open-source with available model weights. However, there’s an important catch: the license restricts commercial use without authorization. That means it doesn’t meet the official open-source definition from the Open Source Initiative, and many in the community have pushed back on the “open-source” label.

Licensing aside, the model itself is impressive. Users are already running MiniMax M 2.7 locally on heavy-duty setups (for example, 4× DGX systems) with vision-language capabilities and up to 200K context in BF16 precision. With tools like Open Code, the model can even monitor its own hardware in real time—reporting thermals, tokens per second, and time to first token—making it a strong candidate for advanced local experimentation.

Gem Opus 4 26B: Gemma Fine‑Tune That Thinks Like Opus

A particularly interesting community project this week is Gem Opus 4 26B from Jack Wrong. It’s a fine-tuned version of Google’s Gemma 4 26B model, designed to mimic the reasoning style of Claude Opus 4.6.

On dual RTX 3090 setups, Gem Opus 4 26B shows strong performance on data reasoning, signal filtering, and agent-style workflows. It feels fast and capable for many reasoning-heavy tasks. However, it still struggles with long, complex coding and deep debugging sessions, so it’s not yet a full replacement for top-tier proprietary coding models.

Still, the fact that a 26B-parameter model can approach Opus 4.6–like reasoning in many scenarios is a big deal for local and self-hosted setups. It’s another sign that open and community-driven models are catching up quickly to closed systems in certain domains. If you’re tracking this broader trend, you may also be interested in recent coverage of new open-source coding models.

AI and Factory Labor: Data Collection in India

Not all AI news this week is about software. In India, there are reports of factory workers wearing head-mounted cameras to record their hand movements. These recordings are then used to train AI systems for industrial tasks and robotics.

On paper, this is framed as data collection to improve automation. But it raises deeper questions: what happens when human labor is used directly to train the systems that may eventually replace it? It’s a stark example of how AI is already being woven into physical labor pipelines, not just digital workflows.

As AI continues to expand into every corner of work and industry, these kinds of deployments will become more common—and more controversial.

Between rumored Claude upgrades, new OpenAI models and pricing, powerful but imperfectly licensed releases like MiniMax M 2.7, and rapidly improving community fine-tunes, the next few weeks in AI look set to be even more intense. With major events like Google I/O and new releases from labs like DeepSeek on the horizon, this pace is unlikely to slow down.