OpenRouter Fusion
OpenRouter's compound-model route: a model alias, plugin, and server tool that fans hard prompts to a panel of models, has a judge compare their answers, then returns structured synthesis through the normal OpenRouter API. Source: OpenRouter docs, 2026-06-17
Fusion is not a new base model. It is a bounded deliberation pipeline exposed as openrouter/fusion, the openrouter:fusion server tool, and a fusion plugin. The pipeline sends the prompt to 1-8 panel models in parallel with OpenRouter web search and web fetch enabled, has a judge model extract consensus, contradictions, partial coverage, unique insights, and blind spots, then lets the caller model write the final answer from that structured analysis. Source: OpenRouter docs, 2026-06-17
As of the 2026-07-02 source check, the live docs and model page still present the same three entry points: openrouter/fusion model alias, openrouter:fusion server tool, and fusion plugin. The model page describes the default panel as the Quality preset, with a cheaper Budget preset and explicit analysis_models / model overrides for controlled panels. Source: OpenRouter Fusion docs and model page, 2026-07-02
The reviewed launch image is a benchmark artifact, not a live leaderboard. It shows a DRACO-task comparison where Fusion panels beat solo models in the launch sample, including Fable 5 + GPT-5.5, Opus 4.8 + GPT-5.5 + Gemini 3.1 Pro, Opus 4.8 + GPT-5.5, Opus self-fusion, and a budget Gemini/Kimi/DeepSeek panel. Preserve it as dated evidence for the panel thesis, not current model-ranking authority. Source: local X artifact wiki/assets/x-bookmarks/2065856853989270011/image-01.jpg, reviewed 2026-07-02
When It Wins
Use Fusion for research, expert critique, architecture choices, competitive comparisons, and other questions where the cost of being wrong is higher than the cost of several model calls. OpenRouter explicitly says it is overkill for short tactical prompts. Source: OpenRouter docs, 2026-06-17
The launch benchmark used 100 DRACO deep-research tasks. OpenRouter reports that panels consistently beat individual models, that a Fable 5 + GPT-5.5 panel scored 69.0% versus Fable 5 alone at 65.3%, and that a budget panel using Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro came within 1% of Fable 5 at roughly half the cost. Source: OpenRouter blog, 2026-06-12
Limits
Fusion is not a drop-in replacement for a coding model. OpenRouter's own FAQ says DRACO does not cover long-horizon tasks, and recommends giving a coding model access to Fusion selectively for questions that deserve multiple perspectives, such as architecture decisions or best-practice research. It is also slower when invoked because the request waits on multiple panel completions plus judge synthesis. Source: OpenRouter blog FAQ, 2026-06-14
This limit is why Fusion belongs in the Long Live Hard SaaS frame: panels can raise answer quality on bounded research tasks, but they do not replace lived product QA, regression testing, or the maintenance loop that SaaS users pay to outsource. Source: raw/notes/saas-bug-moat-slack-thread-2026-06-22.md; OpenRouter blog FAQ, 2026-06-14
Pricing is additive: OpenRouter's model page says a request is priced as the sum of the underlying panel completions and judge call, not as a single model call. Do not quote a flat token price without checking the live OpenRouter activity/model page. Source: OpenRouter model page, 2026-06-17
Routing
| Need | Route |
|---|---|
| One-off deep research or critique | openrouter/fusion alias |
| Coding agent needs occasional second opinions | Add { "type": "openrouter:fusion" } as a server tool to the base coding model |
| Controlled panel experiment | Use the fusion plugin with explicit analysis_models and model |
| Routine implementation | Use the normal harness model; Fusion only for checkpoints |
In Kevin's stack, Fusion belongs after an The Eval Loop (Slop Is an Output Problem) says a task benefits from multiple model perspectives. It complements GLM-5.2 for open-weights coding, Claude Fable 5 for long-horizon agent runs, and Vercel AI SDK / AI Gateway for app integration. It should not replace local verification, tests, or browser checks. Source: User routing directive, 2026-06-17
Timeline
- 2026-07-02 | Deep-reviewed X bookmark
2065856853989270011; repaired and inspected the benchmark chart image, checked the live Fusion docs/model page, and recorded the current Quality/Budget/override routing surface. Source: X/@OpenRouter, 2026-06-13; OpenRouter docs/model page, 2026-07-02 - 2026-06-13 | OpenRouter launched Fusion publicly: model alias
openrouter/fusion, server toolopenrouter:fusion, and configurable plugin surface. The launch tweet reported Fable-level deep-research performance at lower cost via panel + judge synthesis. 14,782 likes and 13,438 bookmarks. Source: X/@OpenRouter, 2026-06-13 - 2026-06-14 | OpenRouter added FAQ guidance: Fusion is not a Fable replacement, not a coding-model replacement, and DRACO excludes long-horizon tasks. Use it as a selective multi-perspective tool. Source: OpenRouter blog FAQ, 2026-06-14
- 2026-06-17 | Wiki page created during full wiki/tool-routing audit; wired into model-routing hierarchy and X bookmark absorption. Source: User, 2026-06-17
- 2026-06-22 | Framed as the productized LLM Council (panel + judge) when capturing the EXM7777 "council of cheaper models" thesis and the AnatoliKopadze "best loop wins" reframe. Source: User, 2026-06-22
- 2026-06-22 | Cross-linked to Long Live Hard SaaS as a bounded harness: useful for multi-perspective research and critique, not a substitute for long-horizon product maintenance. Source: raw/notes/saas-bug-moat-slack-thread-2026-06-22.md