How does Fusion API pricing work, and is it better value than calling Claude Opus 4.8 directly?
Fusion charges the cumulative cost of every model call in the panel — no extra service fee on top. A three-model panel means three completions billed; five models, five. What you pay depends entirely on which models you put in the panel.
The budget panel (Gemini 3 Flash, Kimi K2.6, DeepSeek V4 Pro) uses three models all cheaper per token than Opus 4.8, and their combined cost still runs roughly half the rate of Fable 5. If your panel includes Opus 4.8 itself (e.g. the self-fusion setup), costs rise accordingly. Latency also matters: Fusion must wait for every panel model to finish before synthesizing, running two to three times slower than a single call. It suits research tasks where time is not critical; it's a poor fit for real-time interactions.
When will Claude Fable 5 and Mythos 5 be back?
There is no timeline. Anthropic's public statement said only that it received the Commerce Department directive at 5:21 PM ET on June 12 and complied globally within hours, and that it disputes the reasoning — but gave no conditions or dates for a possible return.
The directive itself does not state its specific national-security concern. Reports indicate one trigger was a jailbreak claim made against Fable 5 by an outside company, which Anthropic characterized as narrow and non-universal. This appears to be the first government-forced takedown of a publicly deployed frontier AI model. The outcome remains open — monitoring Anthropic's official channels for updates is the practical advice.
Does Fusion API actually substitute for Fable 5, or do they suit different use cases?
They suit different scenarios and are not interchangeable. Fusion's near-Fable scores on DRACO are real, but DRACO tests breadth, multi-source integration, and citation quality — research tasks. Fable 5's designed strength was long-horizon agentic tasks: multi-step planning, cross-tool operation, maintaining context across a complex autonomous workflow. DRACO doesn't cover that.
The practical split: if your task is "gather a lot of sources and synthesize a deep-research answer," Fusion's budget panel is a credible option right now. If your task is "let AI autonomously execute a multi-step workflow with minimal human intervention," the closest single-model fallback is Opus 4.8 — Fusion wasn't built for that.
What does this episode tell AI developers over the longer term?
This is a reminder that dependence on a single vendor and a single flagship model carries regulatory risk. Fable 5 and Mythos 5 went from general availability to global shutdown in under 72 hours. Developers who had already integrated either model into products in that window were forced into urgent fallbacks, with nothing available from Anthropic beyond refunds.
OpenRouter's broader bet with Fusion is that the right AI infrastructure is model-agnostic. CEO Alex Atallah's line in the launch post — "the future of AI is neurodiversity, not single-model takeovers" — lands differently against this backdrop than it would have a week earlier. Whether or not you use Fusion specifically, the engineering implication is worth internalizing: "the top model suddenly disappearing" deserves a place in your contingency planning.
Two things happened in AI this week that ended up speaking directly to each other. On June 9, Anthropic launched Claude Fable 5, presenting it as the most capable model it had ever released to the public. Three days later, on June 12, the US Commerce Department issued an export-control directive ordering Anthropic to immediately suspend access to both Fable 5 and Mythos 5 for any foreign national — including foreign nationals employed at Anthropic itself. Unable to verify users' citizenship in real time, Anthropic had no option but to pull both models entirely for all users worldwide.
The very next day, June 13, OpenRouter announced Fusion API and ran this line at the top of its launch post: "Fable-level intelligence at half the price." The timing was striking: a tool that lets a panel of cheaper models combine to match a now-unavailable flagship model's benchmark scores landed at precisely the moment when the flagship itself had been forced off the market.
The core idea is not complicated. Instead of routing a problem to a single model, Fusion sends the same prompt in parallel to a panel of models, then has a judge model read all the responses and synthesize a final answer. From a developer's perspective, it behaves like a single model call — one API request, one response, no extra orchestration needed on the client side.
More precisely: the user sends a prompt; OpenRouter distributes it simultaneously to the models in the selected panel, each equipped with web search and tool-calling capability; the judge model reads every response and maps out where models agree, where they contradict each other, and what unique points each raised; a final model writes the structured answer grounded in that analysis. OpenRouter's defaults run three to five models in parallel.
OpenRouter used Perplexity AI's DRACO deep-research benchmark for its evaluation: 100 complex research tasks spanning ten domains, scored on factual accuracy, breadth and depth, presentation quality, and citation quality, with a negative-weighting penalty for wrong claims.
The top result was a panel of Fable 5 plus GPT-5.5 fused by Opus 4.8, at 69%. A three-model panel of Opus 4.8, GPT-5.5, and Gemini 3.1 Pro fused by Opus 4.8 came in at 68.3%. Claude Fable 5 as a standalone completed 93 of 100 questions (its content classifiers blocked 7) and scored 65.3%.
The figure that drew the most attention was the budget panel: Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro — three comparatively inexpensive models — fused by Opus 4.8 scored 64.7%, within one percentage point of standalone Fable 5, at roughly half the cost.
Also notable: a self-fusion setup, in which Opus 4.8 serves as two members of the panel and also as the judge, scored 65.5%, compared to 58.8% for a single Opus 4.8 call. The gap suggests that a meaningful part of Fusion's gains comes not from model diversity alone but from the synthesis step itself.
OpenRouter's engineering post flagged one incident worth noting: during early runs, some panel models, armed with web search, located the DRACO scoring rubric online. The team excluded those domains with a single configuration change, then re-ran every test from scratch. All published numbers come from the clean setup. Proactively disclosing this and re-running the benchmark rather than ignoring it is the sort of methodological transparency that makes the figures easier to take seriously.
OpenRouter CEO Alex Atallah described Fusion as strongest for deep-research tasks: questions where multi-source breadth matters more than a single long reasoning chain. He also acknowledged that DRACO does not test long-horizon tasks — exactly the scenario Claude Fable 5 was most explicitly designed for — so the two approaches are not direct substitutes.
Developers have four ways to call Fusion: the Chatroom at openrouter.ai/fusion for trial runs; the openrouter/fusion model slug in API calls; a server tool added to the tools array, letting the base model decide when to invoke Fusion; or a plugin argument for custom panel configurations. Pricing stacks the underlying model costs — running a four-model panel means paying for four completions, no flat fee on top.
For the time being, Fable 5 and Mythos 5 remain offline for all users worldwide, with no restoration timeline from Anthropic. The Commerce Department's directive did not specify its national-security concern; Anthropic said it disputes the rationale but is complying immediately. Claude Opus 4.8 and Sonnet 4.6 are unaffected and operating normally.
For developers who relied on Fable 5 for deep research or long agentic tasks: the closest single-model fallback is Opus 4.8. Fusion's budget panel now offers a benchmark-adjacent alternative for research-heavy workloads, with the caveat that latency runs two to three times longer than a single-model call, given the need to collect and synthesize parallel responses. OpenRouter says it will continue tuning based on user feedback.