OpenRouter’s Fusion Promises Claude Fable-Level AI for Cheap—Right as Fable 5 Goes Dark – Decrypt

OpenRouter has launched an API constructed round a easy wager: {that a} panel of low-cost AI fashions, mixed the appropriate method, can match a single costly one. And by “costly,” they imply Claude Fable 5.

The product is known as Fusion. It sends a immediate to a number of fashions in parallel, then makes use of a choose mannequin and a synthesizer to merge the outcomes into one grounded reply.

The timing is fortuitous. Shortly after releasing Fable 5 and Mythos 5 final week, a U.S. export management directive compelled Anthropic to droop these fashions for each overseas nationwide worldwide, citing a disputed jailbreak discovering. OpenRouter took the information to X the subsequent day, leaning straight into the hole with a promise of “Fable-level intelligence at half the worth.”

How one can get an inexpensive Fable

If you ship a immediate to Fusion, OpenRouter fires it off to a panel of fashions in parallel. Every one will get net search and bash instruments.

Then, a choose mannequin extracts consensus factors, contradictions, and blind spots from each response. After this section is over, a synthesizer—Claude Opus 4.8 by default—writes the ultimate reply grounded in that evaluation.

The entire thing occurs server-side. You may swap your mannequin string to “openrouter/fusion” for a default panel, add a fusion software so your individual mannequin calls it selectively, or construct a customized panel within the Fusion chatroom with no code.

OpenRouter examined this on DRACO, Perplexity’s benchmark constructed from actual person deep analysis requests. Fable 5 paired with OpenAI’s GPT-5.5 and synthesized by Opus topped the chart at 69%. Solo Fable scored 65.3%, although seven of its 100 duties by no means ran as a result of its personal content material filters blocked them.

The cheaper mixture is the one OpenRouter needs remembered: A budget Gemini 3 Flash mixed with the open-source Chinese language fashions Kimi K2.6 and DeepSeek V4 Professional, fused and synthesized by Opus, hit 64.7%—beating solo GPT-5.5 (60%) and solo Opus 4.8 (58.8%) outright and touchdown inside a degree of Fable at roughly half the associated fee.

Even pairing Opus 4.8 with a separate occasion itself scored 65.5%, a 6.7-point leap over solo Opus; OpenRouter says roughly three quarters of that raise comes from the synthesis step itself, the remainder from real mannequin variety.

One wrinkle: giving the panel dwell net entry lets fashions floor DRACO’s personal grading rubric in search outcomes, a contamination threat that OpenRouter calls coincidental fairly than deliberate. The repair took one config line to exclude the benchmark’s internet hosting domains from the search instruments, and each printed quantity displays that cleaned-up run.



Value a strive?

OpenRouter is upfront that Fusion is not a full Fable alternative. DRACO skips long-horizon work, the place Fable reportedly nonetheless leads, and for coding, Fusion works as a software a coding mannequin calls selectively, not a wholesale swap—a caveat that echoes what Decrypt discovered testing DeepClaude, a less expensive backend swap that retains Claude Code’s agent loop intact however nonetheless trails Opus on the toughest reasoning duties.

The common mannequin nonetheless handles the day-to-day stuff. Fusion is there for the questions the place one mannequin may miss one thing necessary, and having a couple of views cross-check one another really strikes the needle.

For deep analysis, advanced planning, or something the place contradictions matter, the room appears to assist.

The charts make the essential level clear sufficient: On this sort of work, the costly solo mannequin is now not the one method to get robust synthesis. A bunch of fashions which can be nonetheless simple to get, fused collectively, can sit proper subsequent to it on the outcomes whereas delivering a a lot smaller invoice.

The launch thread break up roughly two-to-one optimistic in sentiment monitoring. AI researcher Andrew Trask known as it “a method larger deal than it appears,” arguing frontier labs won’t ever once more personal the frontier alone. Skeptics pushed again on the framing, nonetheless, citing dangerous coding outcomes, poor software calling, and a scarcity of transparency since Fable 5 isn’t out there anymore to check outcomes.

Fusion runs fully on fashions routed by means of OpenRouter’s personal infrastructure, so it does not repair the export-control drawback on the supply. Anybody locked out of Fable 5 now has choices: a Fusion panel, a backend swap like DeepClaude, or open-weight options comparable to GLM-5.2 that is probably not higher however are adequate for the worth.