GPT‑6.1 Terra, Fable 5.5, Gemini 4 Carbon and a looming DeepSeek ban

The AI race in 2026 is accelerating again. New models from Anthropic, OpenAI, Google DeepMind, Minimax, and China’s GLM team are all lining up for release, while the US government moves to push DeepSeek out of federal agencies. At the same time, Anthropic and Elon Musk are floating new ideas for sharing AI-driven wealth with the public.

Anthropic prepares Fable 5.5 and its next flagship

Anthropic is reportedly close to releasing a new model in the Claude family, often referred to as Fable 5.5 or a next-generation Opus-level model. The focus isn’t just on raw benchmarks, but on how well the model understands what you actually want.

This upcoming model is said to be especially strong at:

• Interpreting your goals even from short, vague prompts
• Inferring intent when you don’t spell everything out
• Going beyond generic, “common sense” answers

Instead of simply repeating public information, Anthropic’s new model is designed to think more creatively and propose original approaches. For example, if you ask how to analyze a stock, today’s models usually talk about price-to-earnings ratios, historical performance, and diversification. The new Anthropic model might instead sketch out a custom algorithm or structured framework to forecast performance, rather than just summarizing what’s already online.

This shift—from regurgitating known advice to generating novel strategies—could be a major differentiator as top labs fight for leadership. If Anthropic lands this well, it increases pressure on OpenAI and Google to respond with more than incremental upgrades.

OpenAI readies GPT‑6.1 Terra as a mid‑tier workhorse

OpenAI is also gearing up for a new release: GPT‑6.1 Terra. In OpenAI’s current lineup, Astra sits at the top, followed by Soul, then Terra, then Luna. Terra is meant to be a capable, mid-tier model that balances performance and price.

From early indications and references in Microsoft documentation, GPT‑6.1 Terra is expected to:

• Be cheaper than GPT‑6.1 Soul
• Possibly outperform Soul on some benchmarks
• Serve as a high-value, general-purpose model for everyday workloads

If Terra really does beat Soul in some areas while being cheaper, it would give OpenAI a strong value play, even if it doesn’t fully match Anthropic’s top-end Fable 5.5 or a hypothetical Opus 6. The big open question is whether this is just a 0.1-style refresh or part of a broader 6.1 lineup update that also includes a new GPT‑6.1 Astra to compete directly with Anthropic’s best models.

We’ve already seen OpenAI go through similar cycles of rapid upgrades in the past, as covered in pieces like earlier GPT‑6 Soul and Gemini 4.0 leaks, so another step-change this year wouldn’t be surprising.

Gemini 4 Carbon: Google DeepMind’s next internal checkpoint

On the Google side, Gemini 4 Carbon has been leaking out of internal documents and employee chatter. It’s described as an internal checkpoint rather than a direct public replacement for Gemini 4 “Argon,” but it’s being actively tested on Google’s Jet Ski coding platform.

Key points about Gemini 4 Carbon so far:

• It’s an internal experiment, not yet a confirmed public release
• Some say it’s an upgrade to Argon; others say it’s a more efficient variant
• It’s expected to match, and possibly exceed, Anthropic’s Claude Opus 5.5 in coding tasks

Insiders also claim that recursive self-improvement (RSI) is now in play at Google DeepMind—meaning Google is using its own models to help design and improve newer models. If true, that could explain why new checkpoints like Carbon are appearing more quickly.

For now, Google is still officially backing Gemini 4 Argon as the main public model, but Carbon could be the next big step in closing the gap with Anthropic at the high end.

Minimax M3.1 Flash: frontier-level quality at a fraction of the cost

While the big three race on premium models, Minimax (often stylized as Miniax) has officially launched M3.1 Flash, a fast and cheap model aimed at real-world deployment.

M3.1 Flash evolved from an earlier experimental model nicknamed “Space Bunny.” Based on internal testing and early benchmarks, M3.1 Flash offers:

• Frontier-level intelligence on key tasks like code and reasoning
• Strong automation capabilities (around 62% on the Automation Bench)
• Deep software engineering performance around 74% on coding benchmarks
• Visual reasoning near 91% on tested tasks

The standout feature is cost and speed:

• About one-third the cost of the original M3 model
• Roughly three times faster in practice
• API pricing around $0.15 / $0.60 per 1M input/output tokens, with very cheap cache reads

Model weights are scheduled to be released, which will make it especially attractive for teams that want to self-host or heavily customize their stack.

Early user tests comparing M3.1 Flash to DeepSeek’s latest Flash model show:

• Minimax: ~40 minutes and effectively near-zero cost in one large batch test
• DeepSeek: ~50 minutes and about $0.40 for the same run

Quality-wise, both produced strong results, with differences coming down to personal preference and specific workloads. In a separate front-end coding test against Claude Opus 5.5, M3.1 Flash produced similar quality output while being around 88 times cheaper (roughly $0.01 vs. $1.22). For many practical use cases, that cost gap is hard to ignore.

GLM 5.5 “Omni”: China’s next multimodal contender

Another Chinese lab, often associated with the GLM series (from Zhipu AI), appears to be preparing GLM 5.5, possibly codenamed “GLM Omni.” A recent research paper from Professor Tang Ji accidentally hinted at a new model that can:

• Understand and generate images alongside text
• Align text and images more precisely in a single model
• Boost interleaved text-image alignment accuracy from about 12.57% to 26.86%

The key idea is that image generation won’t just be a fancy output mode. For GLM 5.5, visual generation is meant to become a real channel for knowledge acquisition—meaning the model can learn, reason, and explain using images as a core part of its understanding, not just an add-on.

Details are still early, but if GLM 5.5 delivers on this vision, it could become one of the strongest multimodal models outside the US ecosystem.

The US moves to push DeepSeek out of federal agencies

While new models keep dropping, geopolitics is once again reshaping who can use what. The US Department of Homeland Security has reportedly ordered civilian federal agencies to:

• Remove DeepSeek models from their systems
• Remove third-party tools and products that rely on DeepSeek
• Stop buying new services built on DeepSeek

This is part of a broader pattern of US–China tech decoupling. Official concerns center on data security and the risk that Chinese models could be used to collect sensitive information about US citizens, companies, and infrastructure. It also appears to be a response to China’s own restrictions on certain tech exports.

Right now, the order targets civilian federal agencies, not everyday users or private companies. But it raises the possibility that restrictions could widen in the future. If that happens, developers and organizations who rely heavily on DeepSeek might need a contingency plan involving US or allied models.

We’ve already seen how quickly regulatory and competitive pressures can reshape the landscape in earlier waves of releases from Claude, Gemini, and GPT‑6, as covered in one of the wildest weeks in AI so far. The DeepSeek situation is another sign that model choice is no longer just a technical decision—it’s also a political one.

Anthropic’s public benefit fund: sharing AI upside

Amid fears that rapid AI progress could displace workers and destabilize parts of the economy, Anthropic is taking an unusual step: committing part of its equity to a new public benefit fund.

The plan, as currently described:

• Anthropic will allocate 5% of its company shares to a new Anthropic Public Benefit Fund
• For every 50% increase in Anthropic’s company value (over a baseline they set), another 2% of shares will be added to the fund

The fund is intended to be an independent entity focused on the public good. While the exact use of the money isn’t fully defined yet, the idea is to create a mechanism that shares some of Anthropic’s upside with society, especially as AI reshapes labor markets and productivity.

Internally, this signals that Anthropic expects AI to have serious economic impact—both positive and disruptive—and wants to get ahead of the criticism that only investors and founders benefit from that shift.

Elon Musk’s SpaceX shares pledge to kids in America

In a similar spirit of “AI-era redistribution,” Elon Musk has announced that next year he plans to give one SpaceX share to every child in America. He’s framed this as the first of several efforts to spread the gains from his companies more broadly.

Musk has been one of the more vocal proponents of ideas like universal basic income (UBI) in a world where AI and automation could eliminate large numbers of traditional jobs. Giving equity directly to children is a different approach, but it follows the same logic: if technology multiplies productivity and corporate value, some of that value should flow to ordinary people, not just shareholders.

There are still plenty of unanswered questions—how eligibility will be verified, how shares will be held or managed for minors, and what this looks like legally and practically. But it’s another sign that tech leaders are starting to think more concretely about how to cushion the social impact of AI.

What this all means for builders and users

Putting it all together, the AI landscape right now looks like this:

• Anthropic is pushing the frontier on reasoning and intent understanding with Fable 5.5 / Opus-level models.
• OpenAI is shoring up its mid-tier offering with GPT‑6.1 Terra, likely focusing on price–performance sweet spots.
• Google DeepMind is iterating fast with internal checkpoints like Gemini 4 Carbon, powered by recursive self-improvement.
• Minimax is proving that “flash” models can get close to frontier quality at a tiny fraction of the cost.
• GLM 5.5 aims to push multimodal understanding, treating images as a first-class knowledge channel.
• The US government is tightening control over Chinese AI models like DeepSeek in sensitive environments.
• Major players like Anthropic and Musk are experimenting with ways to share AI-driven wealth with the broader public.

For developers, startups, and enterprises, the practical takeaway is clear: model choice will keep changing fast. Cost, latency, regulation, and data residency will matter just as much as raw benchmark scores. Having a flexible stack that can swap between providers—and a strategy for handling regulatory risk—will be increasingly important.

And for everyone else, the message is that AI isn’t just about smarter chatbots. It’s about who controls the next wave of economic value, how fast that wave arrives, and who gets included when it does.