Fable 5’s comeback, GPT‑5.6 leaks, Fusion API, and the latest high‑speed coding models

The last few days in AI have been chaotic in the best way: a major US export control drama around Anthropic’s Fable 5, fresh GPT‑5.6 leaks, new open-source coding models from China and Brazil, OpenRouter’s Fusion API, and more. Here’s a clear breakdown of what actually happened and why it matters if you care about cutting-edge AI, coding agents, and autonomous labs.

Why Fable 5 and Mythos 5 were suddenly blocked

Anthropic’s most advanced models, Fable 5 and Mythos 5, were abruptly pulled from global access after a reported US government export control directive. This wasn’t about everyday chatbot use—it was framed as a national security issue.

Officials were reportedly worried that, if safeguards were bypassed, these models could help foreign actors:

• Discover software vulnerabilities
• Accelerate cyber capabilities
• Analyze and debug code at a level seen as strategically sensitive

Anthropic pushed back, arguing that:

• The capabilities in question are not unique to Fable or Mythos
• Other public frontier models can already do similar things
• The issue was being overstated compared to the real risk

Because the directive applied to all foreign nationals—including Anthropic’s own foreign employees—the company disabled access to Fable 5 and Mythos 5 worldwide to comply. Other Claude models remained available; the focus was specifically on these latest, most capable systems being treated as “frontier tech” that might need export controls.

If you want more background on how we got here and what these models can do, check out this deeper dive into why Claude Fable 5 was blocked.

Amazon’s surprising role in the Fable shutdown

The story gets stranger when you add Amazon into the mix. Amazon is one of Anthropic’s biggest backers, but according to reports:

• Amazon researchers found a way to bypass parts of Fable 5’s safety system internally.
• CEO Andy Jassy raised these concerns directly with US officials.
• Just hours later, Anthropic was forced to pull Fable 5 and Mythos 5 globally.

The alleged “vulnerability” centered on the model helping read code, analyze bugs, and identify software issues—capabilities that every major frontier model already offers. Still, that was enough for regulators to treat Fable and Mythos as a potential national security risk.

So you now have a strange dynamic: Amazon is funding Anthropic with billions, while also reportedly triggering the alarm that led to Anthropic’s most powerful models being shut down.

Will Fable 5 come back?

Despite the drama, signs suggest Fable 5 is more paused than permanently killed.

• Senior Anthropic technical staff have already flown to Washington to meet with the Trump administration.
• Both sides reportedly want a resolution rather than a long-term ban.
• The likely outcome is a return with tighter controls, not a complete disappearance.

What might change when Fable 5 returns:

• Stricter safety filters and monitoring
• Lower rate limits or throttled usage
• Enterprise-first or US-only access before a global rollout
• Potentially more “nerfed” behavior compared to the original release

In other words, Fable 5 is likely to come back, possibly within weeks, but probably not in the exact same uncapped form that early users experienced. For context on how powerful it already was for developers, see our hands-on week with Fable 5.

GPT‑5.6 leaks: massive context and agentic coding

While Anthropic deals with regulators, OpenAI appears to be lining up a major counter-move. New leaks suggest GPT‑5.6 could launch as soon as this week or next.

Rumored highlights include:

• A huge 1.5 million token context window
• Lower pricing than Fable 5
• Stronger agentic coding workflows (planning, tool use, multi-step tasks)
• Direct competition with Claude-style, long, structured prompts

The timing is interesting: if Fable users are facing pauses, caps, or friction, a well-timed GPT‑5.6 release could pull many of them back into the OpenAI ecosystem.

None of this is officially confirmed yet, but prediction markets like Polymarket are already pricing in a high probability of a June release.

The Next N2 Pro and the Rio 3.5 “remix” controversy

On the open-source side, one of the most interesting new coding models is The Next N2 Pro, an open-weight model from China focused on complex software and agentic workflows.

Key traits of The Next N2 Pro:

• Built for long-horizon coding tasks and tool use
• Strong at planning, executing, and debugging multi-step app builds
• Uses an “adaptive thinking” system that can make it slower, but more deliberate
• Among the best open models tested so far for agentic coding behavior

Right after this model gained attention, Brazil’s government released Rio 3.5 Open, which was heavily hyped as beating Qwen, DeepSeek, Kimi, and other open models.

Researchers quickly noticed something odd:

• Rio 3.5’s weights looked like a linear interpolation (a blend) of The Next N2 Pro and Qwen 3.5.
• In plain language, it looked less like a brand-new model and more like a remix of two existing ones with a new name slapped on.

After this was called out, the Rio team claimed they had uploaded the wrong file. But the community consensus is that Rio 3.5 was essentially The Next N2 Pro “wearing a Brazil jersey.”

It’s a perfect snapshot of open-source AI right now: huge breakthroughs one day, and the next day, a “new” model turns out to be a clever blend of yesterday’s releases.

Kimi Code 2.7 and the new high-speed variant

Chinese lab Moonshot has released Kimi Code 2.7, the latest version of its open-source coding model—and it’s a serious upgrade over Kimi Code 2.6.

Reported benchmark gains vs. Kimi Code 2.6:

• +21.8% on Kimi CodeBench v2
• +11% on ProgramBench
• +31% on MLSBench Lite

The biggest improvements are in long-horizon and agentic coding:

• Better at following complex instructions end-to-end
• Higher success rates on full project builds
• Uses ~30% fewer reasoning tokens, so it “overthinks” less while finishing tasks faster

On top of that, Moonshot released Kimi Code 2.7 High Speed, a performance-focused variant that is reportedly:

• Up to 6× faster
• Around 180 tokens/second on coding tasks
• Up to 260 tokens/second on shorter-context workloads

Access is currently limited to:

• The Kimi Code beta program
• Kimi’s API
• Business users

For developers, this combination of higher capability, lower token usage, and much higher speed makes Kimi 2.7 one of the most compelling open coding options right now.

DeepSeek 4.1 and Qwen’s rapid-fire updates

DeepSeek, another major Chinese AI lab, may be about to ship a 4.1 update. Reports suggest DeepSeek V4.1 was scheduled for June, and with the Dragon Boat Festival on June 19, many expect a release around that time.

DeepSeek V4 already made waves by offering:

• Very low-cost access
• Strong coding performance
• Impressive long-context handling

Even a smaller 4.1 update could further shake up the “value for money” segment of the AI model market.

Meanwhile, Qwen is iterating at high speed as well. In just a few weeks, we’ve seen Qwen 3.6 Plus followed quickly by Qwen 3.7 Plus. At this pace, another Qwen update in the near future wouldn’t be surprising.

OpenRouter’s Fusion API: panels of models instead of just one

One of the most interesting infrastructure ideas to emerge is OpenRouter’s Fusion API—a system that uses multiple models together instead of relying on a single model for each request.

How Fusion works:

• Your prompt is sent in parallel to a panel of different models (for example, Gemini 3 Flash, Kimi 2.6, and DeepSeek V4 Pro).
• A separate judge model reads all their answers.
• It compares agreements, contradictions, missing details, and unique insights.
• It then fuses everything into a single, stronger final answer.

According to OpenRouter’s benchmarks on 100 hard research tasks:

• Panels of models can consistently beat individual models.
• A panel of cheaper models reportedly beat solo GPT‑5.5 and solo Claude Opus 4.8.
• These panels landed within ~1% of Claude 3.5’s performance at roughly half the cost.

This hints at a possible future where:

• The “best” AI experience isn’t a single giant model, but smart routing and orchestration across many models.
• Multiple cheaper models collaborate like an AI research team, with a judge model synthesizing their work.

Anthropic’s sky-high valuation and investor appetite

Despite the Fable 5 export drama, investor confidence in frontier AI remains intense. SpaceX reportedly crossed a trillion-dollar valuation after its IPO, reminding everyone how much capital is chasing frontier tech.

New projections suggest Anthropic itself could hit a $1.75 trillion valuation this year. If that happens, it would cement Anthropic as one of the most valuable AI companies on the planet, even as its flagship models are under regulatory pressure.

Mistral and Europe’s place in the frontier AI race

French President Emmanuel Macron recently emphasized that France is the only European country with a large language model company that can truly compete with the major US and Chinese labs. He was referring to Mistral AI.

The current landscape looks roughly like this:

• US: OpenAI, Anthropic, Google, xAI, and others
• China: DeepSeek, Qwen, GLM, Yi, MiniMax, and more
• Europe: Mistral as the clear flagship frontier lab

Social media has been full of memes about Mistral dropping a model that beats Fable 5—those are satire, not real leaks—but the broader point stands: Mistral is Europe’s main answer in the frontier model race.

China’s “Black Box” robot and AI-run laboratories

On the robotics and science side, China unveiled an AI-controlled lab instrument called Black Box, built by Beijing Dynaflux Lab Solutions.

What makes Black Box different:

• No buttons, no LCD screen, no traditional human interface.
• It’s designed to be operated entirely by AI systems, not by humans turning knobs or tapping menus.

The vision is a “lights-out” lab where:

• AI systems run experiments end-to-end.
• They generate results, adjust procedures, and optimize workflows autonomously.
• Human scientists move up the chain—focusing on supervision, validation, and high-level research decisions rather than manual operation.

It’s a small device with a big signal: China is pushing hard toward autonomous labs and AI-driven science, where instruments are built from the ground up to be controlled by AI instead of people.

What this all means for AI’s near future

Across all these stories, a few clear trends are emerging:

• Frontier models are now seen as strategic assets, not just products—governments are stepping in.
• Agentic coding and long-horizon workflows are becoming the main battleground for model quality.
• Open-source models are catching up fast, sometimes through innovation—and sometimes through remixing.
• Orchestration layers like Fusion may matter as much as individual models.
• Autonomous labs and AI-run infrastructure are moving from concept to reality.

Whether you’re a developer, researcher, or just following the space, the next few weeks—between potential GPT‑5.6, DeepSeek 4.1, and a possible Fable 5 return—are going to be very busy.