Three June updates to lower your inference bill

Route to the cheapest provider that clears your bar, escalate only when a model gets stuck, and fuse a panel only when one would miss. Plus a new Image API, an MCP server, and GLM 5.2.  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ 

June 2026 Cut the cost of getting it right. Hi Vv, The cheapest token and the cheapest correct result are not the same number. Most of what shipped on OpenRouter this June moves the second one. Here is the short version. Three new server tools. Fusion fuses a panel when one model would miss, Advisor escalates only when a cheap model gets stuck, and Subagent hands routine work to a smaller one. Ten new models, led by Sonnet 5. Anthropic's most capable Sonnet-class model, plus a dedicated Image API and an MCP server that prices a model before you commit. New research on what to actually run. The open-weight models that closed the gap on the proprietary frontier, and where the price difference is still real. Browse all models Platform in numbers Models 400+ One API for all of them Providers 70+ Routed automatically Monthly tokens 100T Processed on OpenRouter Server tools Three new tools that keep expensive models off the work a cheaper one can do. 01 Fusion When one model would miss a hard question, fuse a panel instead. Send one prompt to openrouter/fusion, a panel answers in parallel with web search, and a judge synthesizes one grounded answer. Frontier-grade research without frontier rates on every call. Read the benchmark → 02 Advisor Run a fast, cheap model as your executor and let it consult a stronger one only when it gets stuck, instead of paying premium rates on every token. You pick which model executes and which one advises. Advisor docs → 03 Subagent Keep your expensive model off the grunt work. A large model breaks a task into self-contained subtasks and hands them to a smaller, faster worker model mid-generation, then assembles the result. Subagent docs → Also shipped 04 Unified Image API One endpoint for 30+ image models across 8 providers, with the exact cost returned on every render. Read the launch → 05 OpenRouter MCP Server Compare price and benchmarks and test a model from inside your editor, on a dedicated $10-capped key. Read the announcement → 06 Models API Filter the catalog by price, modality, context, and provider to find the right model for the price, programmatically. Models API docs → New models New models landed across text, image, and code in June. One API key reaches all of them, and your code stays the same when you swap the model slug. Prices are per million tokens, verified live. Featured anthropic/claude-sonnet-5 Sonnet 5 is Anthropic's most capable Sonnet-class model, with frontier performance across coding, agents, and professional work. 1M-token context at $2.00 in and $10.00 out. Also new on OpenRouter z-ai/glm-5.2 $0.93 / $3.00 · 1M moonshotai/kimi-k2.7-code $0.74 / $3.50 · 262K qwen/qwen3.7-plus $0.32 / $1.28 · 1M nvidia/nemotron-3-ultra-550b-a55b $0.50 / $2.20 · 1M google/gemini-3-pro-image $2.00 / $12.00 · image google/gemini-3.1-flash-image $0.50 / $3.00 · image cohere/north-mini-code:free free · 256K nex-agi/nex-n2-pro $0.25 / $1.00 · 262K sakana/fugu-ultra $5.00 / $30.00 · 1M See all new models Latest from the blog The open-weight models that matter: June 2026 Where open weights now match the proprietary frontier, and where the price gap is real. A field guide to spending less without giving up quality. Read the post → A robot is sprinting toward you: Claude or Grok? Eleven models, 30 rounds of a battle royale. The cheapest model won the most games, and the one you would actually ship was a different one. Why benchmark rank and real-task fit come apart. Read the post → How OpenRouter model routing works How every request spreads across 70+ providers: default load balancing, the :nitro and :floor shortcuts, model fallbacks, and the Auto Router that picks for you. Read the post → More on the blog → The number on your invoice is the cost of getting it right. Everything that shipped in June pulls on it from a different direction. The server tools route, fuse, escalate, and delegate so the spend lands where the work is. The Image API prices each render the moment it returns, the MCP server lets you price a model before you commit, and the new launches add cheaper capable options. Across 400+ models from 70+ providers and 100 trillion tokens a month, that is the number OpenRouter keeps driving down, on the same API call you already make. openrouter.ai·Discord·X·Docs 169 Madison Avenue, New York, NY 10016 Manage preferences|Unsubscribe