Fusion API: Multi-Model LLMs After Fable 5

In mid-June 2026, the AI industry experienced two major events that quickly reshaped the discussion around large language models.

First, Anthropic’s flagship model Claude Fable 5, together with its high-end variant Mythos 5, was suspended globally due to U.S. export control policies. The shutdown came suddenly and caught many developers by surprise. A developer conference that was expected to feature Fable 5 also had to replace its main showcase with Claude Opus 4.8.

Soon after, OpenRouter launched the Fusion API, a multi-model collaboration solution. According to OpenRouter’s benchmark results, Fusion API can deliver performance close to Claude Fable 5 at about half the cost. This quickly became a major topic in the global developer community. Related posts reportedly gained more than 5.18 million views on social platforms.

This article analyzes the technical logic, benchmark performance, practical experiments, and industry implications of OpenRouter Fusion API. It also reviews the background of the Claude Fable 5 shutdown, attempts to replicate its capabilities, and the broader shift toward multi-model collaboration.

1. Core Logic and Workflow of OpenRouter Fusion API

For a long time, the mainstream way to improve AI model performance was to increase parameter size, expand training data, and optimize the architecture of a single model.

Fusion API takes a different path.

Instead of relying on one model to complete the entire task, it lets multiple models work together. The goal is to improve final output quality through parallel research, cross-review, and synthesis.

In essence, Fusion API is a standardized multi-model orchestration system. It breaks a complex task into several stages. Different models handle research, review, and final answer generation.

This design helps reduce several common problems in single-model outputs, such as hallucinations, incomplete analysis, and weak logical coverage.

1.1 Three-Step Collaboration Workflow

Step 1: Parallel research by panel models

After receiving a user prompt, Fusion API sends the task to several panel models at the same time.

These models work independently. They can perform web searches, collect data, organize information, and generate preliminary answers under the same tool permissions.

Each model forms its own judgment based on its training background, reasoning style, and tool-use behavior.

Step 2: Cross-review by a judge model

A dedicated judge model then collects all outputs from the panel models.

It compares their answers and identifies:

• Shared conclusions
• Conflicting views
• Missing information
• Unique insights
• Possible factual errors
• Safety or reliability risks

This step acts like a quality review layer. It helps filter weak content and highlights parts that need correction.

Step 3: Final synthesis

The final synthesizer model uses the judge model’s review results to produce the final answer.

It integrates useful points, fills information gaps, corrects errors, and organizes the content into a coherent response for the user.

The key point is that Fusion API is not simply combining several answers. Its real value lies in cross-verification and multi-angle supplementation.

1.2 Simple Integration for Developers

Fusion API is designed to be easy to access.

Developers can enable the default multi-model setup by using the model alias:

openrouter/fusion

The platform also supports customization. Developers can choose panel models for parallel research and select judge models for review and synthesis.

This makes Fusion API adaptable to different use cases. For example, a research-heavy workflow may use strong search-oriented models as panel members. A legal or medical workflow may choose more conservative models for review.

For teams that manage multiple model services, a unified API gateway can reduce integration work. For example, 4sapi can simplify access to different model services and help developers test Fusion API and other models in one technical environment.

2. DRACO Benchmark Results and Performance Data

To evaluate Fusion API, OpenRouter used the DRACO benchmark released by Perplexity AI.

DRACO is designed for deep research tasks. It contains 100 practical tasks across ten domains, including academia, finance, law, medicine, technology, and UX design.

Unlike simple question-answering benchmarks, DRACO focuses on research quality. Each task contains nearly 40 evaluation criteria. These criteria cover factual accuracy, analytical completeness, information integration, and citation reliability.

The benchmark also includes negative scoring for incorrect information and dangerous suggestions. This prevents models from gaining high scores simply by producing long answers.

2.1 Official Fusion Combination Results

OpenRouter tested a cost-effective model combination.

The panel models included:

• Gemini 3 Flash
• Kimi K2.6
• DeepSeek V4 Pro

Claude Opus 4.8 was used as the judge and synthesizer model.

This Fusion setup achieved a DRACO score of 64.7.

For comparison:

The gap between Fusion API and Claude Fable 5 was less than one point. At the same time, the operating cost was about half of Fable 5.

This result suggests that multi-model collaboration can approach the performance of top single models while keeping costs lower.

2.2 Control Experiment with Homogeneous Models

OpenRouter also ran a control experiment.

In this test, two identical Claude Opus 4.8 models were used as the panel group. Another Opus 4.8 model handled review and synthesis.

The final DRACO score reached 65.5. This was 6.7 points higher than standalone Claude Opus 4.8, which scored 58.8.

This result is important.

It shows that performance gains do not only come from using different models. Even the same model can generate different reasoning paths when asked to solve the same task independently.

Different runs may focus on different details, call different tools, or make different assumptions. Cross-review can then combine strengths and correct weaknesses.

Based on these experiments, OpenRouter concluded that about 75% of Fusion API’s performance gain comes from the review and synthesis process. Only about 25% comes from model diversity.

The tests also revealed differences in tool-use behavior. Opus 4.8 tends to call tools frequently, so its advantage may shrink when tool budgets are limited. Claude Fable 5 appears to prefer planning before execution, making it less sensitive to tool restrictions.

DeepSeek V4 Pro also performed well as a standalone model, with a score of 60.3, close to GPT-5.5 and Opus 4.8.

2.3 Notes on Benchmark Interpretation

The DRACO results are useful, but they should be interpreted carefully.

First, different judge models can cause score fluctuations of 10 to 25 points. Therefore, these scores are more suitable for relative comparison within the same test setting, rather than direct comparison with academic benchmark results.

Second, Claude Fable 5 failed to complete 7 out of 100 tasks due to content filtering restrictions. Its final score was calculated using the remaining 93 tasks. This means its test conditions were not fully identical to models that completed all tasks.

Third, during early testing, some models accidentally found DRACO scoring standards through web search. OpenRouter later blacklisted the relevant pages and reran the evaluations. The published scores came from the corrected test round.

These caveats do not invalidate the benchmark. But they show that Fusion API’s results should be viewed as practical evaluation data, not as absolute proof of model superiority.

3. The Background of Claude Fable 5’s Global Shutdown

The rise of Fusion API is closely connected to the sudden shutdown of Claude Fable 5.

According to reports cited in the source material, the suspension of Claude Fable 5 and Mythos 5 was not caused by a simple technical failure. It involved enterprise security feedback and government regulatory intervention.

Amazon, Anthropic’s largest investor, reportedly played an important role in the incident.

Since 2023, Amazon has invested $13 billion in Anthropic and planned to add another $20 billion. During internal testing, Amazon’s technical team allegedly discovered an effective jailbreak method for Claude Fable 5. The method could bypass the model’s safety barriers and obtain network-attack-related information.

Amazon CEO Andy Jassy reportedly submitted the test results to U.S. regulators. After that, U.S. National Cyber Director Sean Kanes held an emergency meeting. The U.S. government then issued an export control order.

According to the same reports, Anthropic received only a 90-minute response window. Regulators required Anthropic to fix the jailbreak vulnerability, but CEO Dario Amodei refused.

Anthropic stated publicly that the vulnerability was only a minor flaw. The company also argued that similar jailbreak risks existed in other mainstream public models. However, this explanation was not accepted by the U.S. government and Amazon.

As a result, Fable 5 and Mythos 5 were banned for all non-U.S. citizens worldwide. The restriction reportedly included foreign employees working inside Anthropic, such as Andrej Karpathy.

For existing users, Anthropic later stated that Opus, Sonnet, and Haiku models would remain available. User quotas would also be reset. Users dissatisfied with the service could request refunds before June 20. Those who subscribed through Apple channels needed to follow Apple’s separate refund process.

This regulatory move triggered wide debate in the AI industry. Many observers viewed it as a targeted restriction on Anthropic’s newest flagship models rather than a general policy aimed at all AI companies.

4. Attempts to Replicate Fable 5’s Capabilities

After Fable 5 went offline, many developers tried to replicate its performance and style.

Developer Jamieson O’Reilly conducted one notable experiment. He wanted to test whether Fable 5’s unique behavior came mainly from model weights or from its system prompt.

He used the publicly released official Fable 5 system prompt with Claude Opus 4.8. A standard Opus 4.8 setup was used as the control group.

Both groups used the same hardware conditions and a 1 million-token context window. The only variable was the system prompt.

The test task was to generate a landing page in Apple’s design style.

The outputs showed clear differences in brand positioning, text tone, page structure, and visual style. Jamieson believed he had created a simplified version of Fable 5.

However, this approach has obvious limits.

A system prompt can imitate external style and some interaction patterns. It can influence tone, structure, and task behavior. But it cannot reproduce capabilities learned during pre-training, reinforcement learning, or long-context planning optimization.

In other words, prompt engineering can simulate surface behavior. It cannot fully recreate the internal capability of a model.

Domestic AI companies also responded to the Fable 5 incident. Zhipu AI announced that GLM-5.2 had been fully opened to all users of its Coding Plan, covering Lite, Pro, Max, and Team editions.

The company emphasized that advanced AI capabilities should be available to developers more broadly, rather than controlled or restricted by a small number of parties. This message aligned with the growing industry interest in open access and sustainable iteration.

5. Industry Trends and the Future of Multi-Model Collaboration

The launch of OpenRouter Fusion API shows that multi-model collaboration is becoming a practical alternative to single flagship models.

For years, the AI industry has focused on larger models, more parameters, and greater compute investment. Fusion API suggests another path. Strong results can also come from combining mature models in a structured way.

This has several important implications.

First, multi-model collaboration can improve output quality through cross-review. When several models analyze the same task independently, their answers can reveal missing details, contradictions, and weak assumptions.

Second, it can reduce reliance on a single model provider. This is especially important after the Fable 5 shutdown. Enterprises do not want critical workflows to depend entirely on one model that may suddenly become unavailable.

Third, it creates more flexible cost structures. Developers can choose cheaper panel models for broad research and reserve stronger models for judging and synthesis. This makes high-quality AI output more economically practical.

Multi-model collaboration is especially useful for:

• Deep research
• Academic literature analysis
• Legal document review
• Medical information analysis
• Financial research
• Technical report writing
• Enterprise knowledge synthesis

However, this architecture also has challenges.

More model calls usually mean higher latency. The workflow is also harder to monitor, debug, and maintain. Teams need to choose panel models and judge models carefully based on the task type.

Over time, single large models and multi-model collaboration systems will likely coexist.

Flagship single models will still be used for tasks that require extremely high precision and low latency. Multi-model systems will become popular in research-heavy, cost-sensitive, and risk-diversified workflows.

6. Conclusion

The shutdown of Claude Fable 5 reflects the growing influence of regulation, geopolitics, and enterprise security concerns on AI access.

The rise of OpenRouter Fusion API reflects another trend: developers and enterprises are looking for more flexible and resilient model architectures.

According to OpenRouter’s DRACO results, multi-model collaboration can approach the performance of top single models at a much lower cost. This challenges the assumption that only larger models can deliver better results.

Experiments that tried to replicate Fable 5 through prompts also show an important lesson. The strength of advanced models does not come only from visible prompt design. It also comes from training data, model weights, reasoning mechanisms, tool-use behavior, and long-context planning.

For developers and enterprises, the key takeaway is clear. Relying on one flagship model creates risk. A more robust strategy is to diversify model choices and adopt collaborative architectures where appropriate.

Fusion API provides one example of this shift. It offers a practical alternative after the Fable 5 shutdown and points toward a broader future for AI infrastructure.

In that future, model aggregation, cross-review, tool orchestration, and synthesis mechanisms may become standard components of AI systems. The industry may move away from pure single-model competition and toward more open, flexible, and cost-effective AI workflows.