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

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 decide 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 international 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 value.”

Easy methods to get an affordable Fable

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

Then, a decide 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 possibly can swap your mannequin string to “openrouter/fusion” for a default panel, add a fusion device so your personal 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 desires 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 some extent 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 carry comes from the synthesis step itself, the remainder from real mannequin variety.

One wrinkle: giving the panel stay net entry lets fashions floor DRACO’s personal grading rubric in search outcomes, a contamination danger 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 revealed quantity displays that cleaned-up run.

Price a strive?

OpenRouter is upfront that Fusion is not a full Fable substitute. DRACO skips long-horizon work, the place Fable reportedly nonetheless leads, and for coding, Fusion works as a device 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 would possibly 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 fundamental level clear sufficient: On this sort of work, the costly solo mannequin is not the one technique to get robust synthesis. A gaggle of fashions which can be nonetheless straightforward 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 constructive in sentiment monitoring. AI researcher Andrew Trask called it “a method greater deal than it appears,” arguing frontier labs won’t ever once more personal the frontier alone. Skeptics pushed back on the framing, nevertheless, citing dangerous coding outcomes, poor device calling, and an absence of transparency since Fable 5 isn’t obtainable anymore to match outcomes.

Fusion runs totally on fashions routed by means of OpenRouter’s personal infrastructure, so it would not repair the export-control downside on the supply. Anybody locked out of Fable 5 now has choices: a Fusion panel, a backend swap like DeepClaude, or open-weight alternate options equivalent to GLM-5.2 that will not be higher however are adequate for the value.