OpenRouter July 2026: New Models, Prices, and Rankings

OpenRouter's July 2026 intake reads like a full quarter of releases compressed into three weeks. Between July 1 and July 22, ten new model families reached the catalog — roughly one every 2.2 days — spanning OpenAI's renamed GPT-5.6 trio, xAI's Grok 4.5, Moonshot's Kimi K3, Meta's first paid-API model, Thinking Machines' debut, two coding specialists from Kuaishou, Meituan's LongCat 2.0, Google's triple Gemini launch, and poolside's open-weight Laguna line.

The spread is the story. The cheapest paid model in the wave, poolside's Laguna XS 2.1, lists at $0.06 per million input tokens; the most expensive, GPT-5.6 Sol, lists at $5.00 input and $30.00 output. That is roughly an 83x gap on input and a 250x gap on output — and a free Laguna S 2.1 variant sits below both at $0 with a 256K context window. A month ago, our June OpenRouter roundup tracked a 25x spread across five models in ten days. July widened the range and sustained the cadence for an entire month.

This tracker catalogs all fifteen July listings with live-verified pricing, computes the listing-lag column nobody else publishes — the gap between when a vendor shipped a model and when OpenRouter's created field says it arrived — and closes with a workload-by-workload routing posture. Every benchmark figure this month is vendor-reported, and the post labels each one as such.

Walk the calendar. poolside opened the month quietly, listing Laguna XS 2.1 on July 2 at the lowest paid price of the wave. xAI shipped Grok 4.5 on July 8. OpenAI took GPT-5.6 to general availability on July 9 — just under two weeks after its June 26 preview — and all six OpenRouter listings (Sol, Terra, Luna, and their pro-mode twins) went live the same day. Kuaishou's Kwaipilot released KAT-Coder V2.5 in Pro and Air tiers around July 10–11. Then the back half: Kimi K3 and Muse Spark 1.1 listed July 16, Thinking Machines' Inkling July 17, LongCat 2.0's delayed listing July 20, and a four-model July 21 that combined Google's triple announcement with poolside's Laguna S 2.1.

Count the distinct families that hit the catalog between July 1 and July 22 and you get ten — one roughly every 2.2 days. June's wave, by comparison, packed five major models into a single ten-day burst; July held nearly the same density for an entire month. The practical consequence is the same one we drew in June, amplified: any model-selection decision made at the start of July was stale by the end of it.

The table below is the canonical reference for the July window: all fifteen listings with pricing per million tokens, context window, the date OpenRouter's created field reports, and the vendor's own launch date — plus the lag between the two. Prices were read from the live OpenRouter models API at publication and cross-checked against each vendor's primary announcement. No other roundup we have seen publishes the lag column, and it is the column that prevents the most mistakes.

Two footnotes keep the table honest. GPT-5.6's pro reasoning modes (Sol Pro, Terra Pro, Luna Pro) list at the same base rates as their standard tiers — pro mode bills through reasoning tokens rather than a separate price — so the trio's three rows cover all six listings. And the KAT-Coder pair shows OpenRouter's created date one day before the stated vendor date — almost certainly a timezone or rounding artifact, not a leak. Treat July 10–11 as the release window.

July's top shelf belongs to three launches. OpenAI's GPT-5.6 went GA on July 9 with its new two-axis naming — generation number plus a durable tier name, so Sol is always the flagship, Terra the balanced tier, Luna the high-volume tier — and its announcement promised "full availability over the next 24 hours" across ChatGPT, Codex, and the API. On OpenAI's own preview benchmark tables (vendor-stated, no independent audit at GA), Sol posts roughly 88.8% on Terminal-Bench 2.1 with an ultra mode near 91.9% — while the same tables show Anthropic's Claude line still leading several agentic-coding rows, including SWE-Bench Pro, by a wide margin.

Grok 4.5 landed a day earlier at half of Sol's input price. xAI's vendor-reported numbers put it at 83.3% on Terminal-Bench 2.1 — just behind Claude Fable 5's 84.3% on the same table — and 64.7% on SWE-Bench Pro against Fable 5's 80.4%, while claiming roughly 4.2x fewer output tokens than Opus 4.8's max mode on that benchmark. The efficiency claim matters as much as the score: at $2/$6 with a $0.50 cached-input rate, Grok 4.5 is priced as a workhorse, not a trophy. Its 500K context is the smallest of the frontier three, and EU availability had not arrived at launch.

Kimi K3 is the wave's open-weight wildcard. Moonshot shipped it July 16 as a 2.8T-parameter flagship with native multimodal input and a 1M-token window at $3/$15 — premium pricing by open-lane standards — with weights promised for July 27. If that date holds, K3 becomes the largest open-weight release of the summer; until the files land, it is a closed API model with an announced commitment.

The mid-tier is where July got structurally interesting. Meta Superintelligence Labs launched Muse Spark 1.1 on July 9 through Meta's first-ever paid Model API — $20 in free credits at launch, closed weights, no Hugging Face release — and OpenRouter listed it July 16 under a new meta/ vendor prefix, distinct from the older meta-llama/ namespace. At $1.25/$4.25 with a 1M context it is priced to compete, but it launched US-only with a single Meta-hosted provider. Thinking Machines Lab — Mira Murati's outfit — followed on July 17 with Inkling, an open-weight 975B-total / 41B-active multimodal MoE at $1/$4.05, the lab's first major OpenRouter listing.

The quiet value story is Kwaipilot's KAT-Coder V2.5, released around July 10–11 in two tiers positioned to take a whole coding issue autonomously. Pro lists at $0.74/$2.96 and Air at $0.15/$0.60 — both with 256K context. Air's pricing is the notable one: it undercuts every general-purpose model in the wave except poolside's Laguna line while targeting agentic coding specifically, which makes it the budget candidate for issue-to-PR automation pipelines.

Two releases define July's bottom rungs. Meituan's LongCat 2.0 — a sparse MoE with 1.6T total and roughly 48B active parameters, a ~1M-token window, and a repo-level coding and long-horizon agent focus — reached OpenRouter on July 20 at $0.30/$1.20, three weeks after its actual June 30 launch. And poolside shipped Laguna S 2.1 on July 21: an open-weight 118B-total / 8B-active MoE at $0.10/$0.20, trained start-to-launch in under nine weeks on 4,096 H200 GPUs (vendor-stated), with weights in four quantizations on Hugging Face on day one under the OpenMDW-1.1 license — plus a free OpenRouter variant at $0/$0 with a 256K window.

That line is poolside's own framing — a vendor self-description, not an independent ranking. The disclosed numbers behind it are more modest and more credible than the slogan: 70.2% on Terminal-Bench 2.1 against Fable 5's 88.0% and GPT-5.6 Sol's 88.8%, and 78.5% on SWE-Bench Multilingual, roughly level with much larger models. poolside is the one lab in this wave that publishes its harness and full trajectories, which makes its self-reported figures the most auditable of the month even though they remain vendor-run. The full technical read is in our Laguna S 2.1 launch analysis.

Run the arithmetic on a representative agentic task — 500K input tokens plus 100K output. On Laguna XS 2.1 that costs about $0.042 (500K × $0.06/M + 100K × $0.12/M). On GPT-5.6 Sol the same token volume costs about $5.50 (500K × $5/M + 100K × $30/M) — roughly 131x more. The June wave's equivalent spread was about 25x. The gap between the cheapest capable model and the most expensive frontier model widened by a multiple in one month, which makes routing discipline — not model loyalty — the thing that controls an AI budget.

Google compressed its entire July move into a single July 21 announcement: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a security-focused 3.5 Flash Cyber variant, shipped together. 3.6 Flash replaces 3.5 Flash as the Gemini app and API default at $1.50/$7.50 — with the output rate cut from a prior $9 — and is currently Google's strongest publicly available model, because Gemini 3.5 Pro remains unreleased and in partner testing. Google frames the release as an efficiency gain (roughly 17% fewer output tokens per task, per Artificial Analysis measurement it cites) rather than a frontier jump. Flash-Lite lists at $0.30/$2.50, streams around 350 tokens per second, and is explicitly positioned for subagents in multi-agent systems.

We keep this section short deliberately, because the per-model depth already exists: the full Gemini 3.6 Flash launch analysis, how it stacks up against GPT-5.6 and Kimi K3 on benchmarks, and Flash-Lite's subagent economics. The roundup-level takeaway is positioning: Google now covers the $0.30 and $1.50 rungs of the ladder with same-day OpenRouter listings, and it told the market plainly where the next jump comes from.

That is Google confirming Gemini 4 pre-training is officially underway — in training only, with no release date — alongside a launch it explicitly framed as an efficiency play. Read together, the message is that the 3.x line is now Google's cost-performance ladder while the capability bet rides on 4. For teams routing production traffic, that makes 3.6 Flash a stable default to adopt rather than a stopgap to wait out.

A pricing table is only useful once it becomes a routing decision. The matrix below pairs each major workload class with the July models we would shortlist for it, grounded in the live pricing above and with vendor benchmark claims treated as claims. It follows the same discipline as our AI cost-optimization playbook: match the task to the cheapest model that clears your quality bar, and verify on your own prompts before committing a default.

Multi-model routing is also no longer exotic plumbing — OpenRouter's Fusion ensemble endpoint makes cross-model synthesis a config option rather than an engineering project. For teams that want the eval harness, routing policy, and spend governance built rather than described, that is the core of our AI transformation engagements.

Per our June roundup, that wave was a ten-day burst — five major models from May 27 to June 4, spanning a roughly 25x input-price range. July sustained nearly the same per-day cadence across a full month: ten families in 22 days, an ~83x input spread, and a four-model single day on July 21. The structural reading we made in June — that every price tier now carries a credible long-context option — hardened in July into something stronger: every price tier now has competition within it, from three models clustered at or below $0.30 input to three more contesting the $1–$2 workhorse band.

The absences are as telling as the arrivals. DeepSeek, Mistral, and the Meta-Llama open-weight line shipped nothing new on OpenRouter in this window — Meta's July move was the closed, API-only Muse Spark under a brand-new vendor prefix, a strategic U-turn from the lab that defined open weights. DeepSeek V4's GA status remained secondary-sourced-only as of July 21, with no primary confirmation. Meanwhile the catalog itself churns: a live pull at publication returned 343 models, a point-in-time figure that fluctuates weekly as free and paid variant pairs and deprecations come and go. For the pre-wave baseline, see our April 2026 OpenRouter rankings.

Projecting forward: if the July cadence holds even approximately, the second half of 2026 delivers a new production-relevant model roughly every week, and the shelf life of any "best model" claim compresses toward days. Kimi K3's promised July 27 weights release is the next scheduled event on that calendar. The durable posture is the one both roundups now point to — maintain a routing policy with named fallbacks per workload, and re-run the comparison monthly rather than re-litigating it per launch.