OpenRouter June 2026: New Models, Pricing and Rankings

OpenRouter added five major models in the ten days from May 27 to June 4, 2026 — a single window dense enough to read as a market signal rather than a scattered news cycle. The list spans frontier closed weights and open-weight challengers, from Claude Opus 4.8 at $5 per million input tokens down to MiniMax M3 at $0.30 on a launch promotion.

What makes the wave worth cataloging is not any single launch. It is the shape of the whole. Every tier of the price ladder now has a credible million-token-context option, and the cheapest capable long-context models cost a small fraction of the most expensive frontier ones. That is a structural change in how teams should think about model selection, not a temporary dislocation.

This guide catalogs the five additions with their OpenRouter pricing and context windows, builds a price-tier ladder for the platform as it stood in early June, separates real-time usage rankings from quality benchmarks, and lays out a practical routing posture. Vendor self-reported numbers are marked as such throughout — a launch claim is not an independent benchmark.

Treat the calendar as the story. Claude Opus 4.8 and its Opus 4.8 Fast variant landed on OpenRouter on May 27. Step 3.7 Flash followed on May 28. MiniMax M3 reached the API on May 31 with broader availability on June 1. Qwen3.7-Plus listed on June 3, and NVIDIA launched Nemotron 3 Ultra on June 4, the day this guide publishes, after announcing it at Computex in Taipei on June 1.

Most coverage treats each of these in isolation. Stacked into one window, they describe a market where new frontier-aimed capability arrives roughly every other day and where open-weight challengers now ship alongside closed-frontier releases rather than trailing them by months. The practical consequence for any team running inference at scale is that a model-selection decision made in early May was already stale by early June.

The wave is not a coincidence of timing alone; it reflects where inference demand now concentrates. OpenRouter raised a $113M Series B on May 26, 2026, led by CapitalG with participation from NVentures, ServiceNow Ventures, and others alongside existing investors Andreessen Horowitz and Menlo Ventures. The round more than doubled the company's valuation to roughly $1.3 billion in a single year.

The usage numbers explain the investor interest. Weekly token throughput has grown from about 5 trillion to 25 trillion over six months — a 5x increase — with the platform reporting more than 8 million developers building across 400-plus models. For a model provider, listing on OpenRouter is now one of the fastest routes to distribution, which is precisely why frontier and challenger labs alike push new releases there on day one.

The table below collects every model in the wave with its OpenRouter list date, pricing, context window, and open-versus-closed status — the canonical reference for the June window. Prices are per million tokens, read from OpenRouter model pages on June 4, 2026. Where a figure is a launch promotion or unconfirmed, the cell says so.

Read the third column carefully. Three of the five additions are open-weight or open-weight-committed releases, and four of the five carry a million-token context window. The most expensive model in the wave is roughly 25 times the input price of the cheapest — and the cheapest is still a capable multimodal model. That spread, at this density of launches, is the structural argument of this post.

MiniMax M3 is the model that gives the wave its narrative. MiniMax positions it as the first open-weight model to combine frontier-level coding, a 1M-token context window, and native multimodal input across text, image, and video. Its architecture, MiniMax Sparse Attention, replaces full attention with KV-block selection — the same family of long-context efficiency tricks the wider open-weight field has converged on. For the full technical treatment, see our MiniMax M3 deep-dive.

Two cautions belong up front. M3 is API-available now, but its open weights were promised on Hugging Face and GitHub within roughly ten days of the June 1 launch — they are not downloadable as of this writing. And its benchmark figures are vendor self-reported: MiniMax ran them on its own infrastructure with its own scaffolding. Treat the headline scores as a launch claim, not an independent result.

The honest counterweight to the vendor numbers comes from independent testing. In an independent multi-model benchmark published June 1, 2026, MiniMax M3 scored 78 out of 100 — a solid Tier B — while Claude Opus 4.8 landed at 95 (Tier A) and GPT-5.5 and GPT-5.4 sat at 96 to 97. M3 is genuinely strong and genuinely cheap; it is not, on independent evidence, on par with Opus 4.8 for quality. Both things are true at once, and the routing implication is to match the model to the task rather than to the headline price.

Zoom out from the wave to OpenRouter's broader price ladder as it stood in early June. The bars below chart input price per million tokens across representative models, from sub-$0.10 reference rates up to the $5.00 frontier tier. Longer bars are more expensive. The point is not the ordering — it is that capable, long-context options now exist at nearly every rung.

The takeaway is structural. A year ago, the cheap end of the ladder meant short context windows and modest capability; the long-context frontier meant premium pricing. The June ladder breaks that coupling. Step 3.7 Flash, MiniMax M3, Qwen3.7-Plus, and the Nemotron family all sit at or below $0.40 per million input tokens while carrying 256K-to-1M context windows. The decision a team faces is no longer "can we afford long context" but "which long-context model fits this specific workload at this price." Our AI transformation engagements start with exactly that comparative eval.

OpenRouter publishes a programming collection ranked by usage. In early June, its top coding models by volume ran roughly: MiMo V2.5, MiniMax M3, DeepSeek V4 Flash, MiMo V2.5 Pro, Hy3 Preview, DeepSeek V4 Pro, Claude Opus 4.7, Owl Alpha, Step 3.7 Flash, and Claude Sonnet 4.6. The critical caveat: these are real-time usage rankings, not quality benchmarks. MiniMax M3 ranking near the top reflects adoption and price, not a verified quality endorsement.

Usage rankings also move fast, and history is instructive. Tencent's Hy3 preview held OpenRouter's number-one overall usage ranking from April 27 to May 11, accumulating about 7.7 trillion tokens in 19 days at $0.063 per million input. By early June it had slipped down the programming collection. Independent analysis read the steady usage after Hy3 moved to a paid tier as organic demand from a single large application rather than promotional uptake — a reminder that a ranking can reflect one heavy consumer as much as broad preference.

The most important pattern in the data is not any single model. It is the split between the volume market and the dollar market. Chinese-origin models — from Xiaomi, Alibaba, MiniMax, DeepSeek, and Moonshot — now account for more than 45% of all OpenRouter traffic by token volume, up from below 2% a year ago. Anthropic, by contrast, holds around 12.3% of token share but retains a much higher dollar share through premium pricing. OpenRouter now routes both commodity inference and premium inference as distinct lanes.

Independent market analysis frames the dynamic bluntly: the market is following price, not quality. Sub-dollar blended pricing from open-weight challengers has captured disproportionate token volume, while premium closed models dominate dollar spend despite lower token share. Inference volume across the tracked market grew roughly 11.1x year over year. This is the interpretive core of the wave — not a temporary dislocation but a durable structure that should change how developers reason about model selection.

Projecting forward, the bifurcation looks structural rather than cyclical. If sub-dollar long-context models keep arriving every few days — as the ten-day wave suggests they will — the volume lane will keep commoditizing while the dollar lane consolidates around the handful of frontier models worth a premium for the hardest work. The strategic move for most teams is to stop treating "which model" as a single decision and start treating it as a routing policy that differs by task class. That is exactly the multi-model thesis behind OpenRouter Fusion, the synthesis layer that makes multi-model routing practical.

The wave does not crown a single winner; it rewards deliberate routing. Below is a workload-by-workload posture grounded only in the evidence above — independent benchmarks where they exist, vendor claims marked as such, and price read live from OpenRouter.

For most agencies and engineering teams, the starting point is not a switch but a measurement: benchmark two or three candidates on your own prompts, measure token spend and latency, and decide per workload. The wave makes that discipline more valuable, not less, because the cheapest capable option changes month to month. Treat the price ladder as a menu and your routing policy as the thing you actually maintain. Sibling references for the individual models — Step 3.7 Flash, Qwen 3.7 Plus, and Nemotron 3 Super — carry the per-model detail this roundup keeps brief.