OpenRouter’s Fusion Promises Cheap Claude Fable-Level AI – Like Fable 5 Goes Dark

OpenRouter has implemented an API built around a simple bet: that a group of low-cost AI models, including the right path, can match a single high cost. And by “expensive,” he means Claude Fable 5.

The drug is called Fusion. It quickly transmits to several models in parallel, then uses a judge and synthesizer to combine the results into one stable solution.

The timing is lucky. As soon as he was released Legend 5 and Mythos 5 last week, a US antitrust law forced Anthropic to freeze these brands for any foreign nationals around the world, based on the findings. OpenRouter took the issue to X the next day, leaning straight into the gap with the promise of “Superior intelligence at half the price.”

How to get a cheap Fable

When you send a prompt to Fusion, OpenRouter fires up a group of matching models. Everyone gets to search the web with bash tools.

After that, the judge removes the points of agreement, contradiction, and blind spots from each answer. After this phase, the developer – Claude Opus 4.8 by default – writes the final standard response to the review.

Everything is done on the server side. You can change your model string to “openrouter/fusion” to be a default group, add a link tool so that your model can be called optionally, or create a group in the Fusion chat room without code.

OpenRouter tested this Image of DRACOPerplexity’s benchmark developed from real user requests. Fable 5 integrated with OpenAI’s GPT-5.5 and developed by Opus topped the chart at 69%. Solo Fable scored 65.3%, even though seven of his 100 jobs didn’t run because his filters blocked them.

The low-cost combination is what OpenRouter wants to remember: the cheap Gemini 3 Flash combined with the open Chinese versions of Kimi K2.6 and DeepSeek V4 Pro, mixed and developed by Opus, hit 64.7%—beating GPT-5.5 (60%) alone and Opus 4.8 (58.8%) for almost half the price.

Even combining Opus 4.8 with a different model alone scored 65.5%, a 6.7-point jump over the solo Opus; OpenRouter says that about three-quarters of the increase comes from the step itself, the rest from the actual models.

One wrinkle: giving a group access to the Internet allows colors to show up in DRACO’s results, a vulnerability that OpenRouter calls accidental rather than intentional. The configuration took one standard line to exclude benchmark storage areas from the search engine, and each printed number indicates the same speed.

Should I try it?

OpenRouter is up front that Fusion is not a replacement for Fable. DRACO jumps into a long career, where Fable says it still leads, and when coding, Fusion works as a call tool of choice, not a major exchange – a caveat that equates to that. Decrypt got the exam DeepClaudea cheap trade-off that keeps agent Claude Code from running smoothly but still pursues Opus on the most critical proposition.

The standard model still works on a day-to-day basis. Fusion is there for questions where one type might miss something important, and having a few thoughts on each other moves the needle.

For serious research, difficult planning, or anything where debate is important, the room seems to help.

These charts make a clear point: For this type of work, an expensive individual model is not the only way to achieve a strong synthesis. A group of models that are easy to find, put together, can be close to it in terms of results while offering a very small bill.

The opening thread was split almost two-to-one in favor of following up on ideas. AI researcher Andrew Trask he called it “way bigger than meets the eye,” the borderline contenders will no longer be limited by themselves. Suspects pushed back in the design, however, mentioning the bad effects of the codes, the tool calls, and the lack of transparency since the Fable 5 is no longer available to compare the results.

Fusion runs on all models that are managed through the OpenRouter architecture, so it doesn’t fix the problem of external routing at source. Anyone locked out of Fable 5 now has options: a Fusion team, a back swap like DeepClaude, or other heavy duty options like GLM-5.2 which may not be the best but it is adequate for the price.