A 65B Open Model Just Beat a 501B Giant at Coding: Ripple-65B Costs $0.20 Per Million Tokens

Key takeaways:

• Ripple-65B Preview launched on 4 November 2026 from Ricursive Intelligence, a chip-design startup founded by the researchers behind Google's AlphaChip.
• It has 65 billion dense parameters, trained from scratch on 12 trillion tokens, with a 1 million token context window in its model card.
• Ricursive reports 80.52 on Terminal-Bench 2.1, which it says is the best score among US open-weight models, ahead of Reflection Beam 501B at 80.1 and Meta Muse Spark Open 816B at 76.4.
• It also reports 79.1 percent on SWE-bench Pro v2.0 Hard, plus strong maths and knowledge scores, but it trails rivals on SWE-bench Verified, DeepSWE, SciCode and MCP Atlas in its own table.
• The weights are free under Apache 2.0 on Hugging Face, and the API costs $0.20 per million input tokens and $0.60 per million output tokens on Together AI, with cached input at $0.02.
• The company says the whole project cost about $32 million to build from scratch, a fraction of what frontier labs spend.
• No independent evaluation had been published at the time of writing, so every benchmark figure here is vendor reported.

What is Ripple-65B Preview, in plain terms?

Ripple-65B Preview is the first public language model from Ricursive Intelligence, a startup better known for AI chip design than for chatbots. The company was founded by the two researchers behind Google's AlphaChip, the system that designs chip layouts with reinforcement learning. That background matters, because Ripple is a general-purpose version of the model that powers Ricursive's own chip design platform, tuned for agentic coding, which is work where an AI plans, runs commands in a terminal, reads the results, and keeps going until a task is done.

The model itself is small by 2026 flagship standards, and that is the whole story. It has 65 billion dense parameters, which means all 65 billion are active on every token, unlike mixture-of-experts models that keep most of their parameters switched off. It was trained from scratch on 12 trillion tokens. Its model card lists a 1 million token context window, which is the amount of text and code it can hold in working memory at once. For comparison, the rivals it beats on its headline benchmark carry 501 billion, 816 billion, 550 billion and 975 billion parameters. A 65 billion parameter model beating those giants at terminal coding is why this release travelled so fast on launch day.

The release is open in the way developers actually care about. The weights are published on Hugging Face under the Apache 2.0 licence, which allows commercial use, modification and redistribution without extra terms. There is no gated form and no waitlist for the weights. The company also lists hosted API access through Together AI, GMI Cloud and OpenRouter, with Fireworks support expected soon.

What does 80.52 on Terminal-Bench 2.1 actually mean?

Terminal-Bench is a test of agentic coding in a real terminal. The model is given a task, such as fixing a bug, building a feature, or configuring a system, and it has to work in a command-line environment, run commands, read errors, and produce a working result. It is a much closer test of real developer work than a multiple choice quiz, because the model has to recover from its own mistakes instead of picking one right answer from four options.

Ricursive reports 80.52 on Terminal-Bench 2.1, and says that is the best score among US open-weight models. The table below shows the comparison from Ricursive's own model card. Read the rival scores with care: some of them come from Ricursive's own evaluation runs rather than from the rival labs' official cards, and the company itself notes that scores came out lower for nearly every model, Muse Spark Open excepted, when it used the official Terminus 2 harness at 128K sequence length with summarisation. That does not make the headline false, but it does mean the ranking is a vendor's table, not a neutral leaderboard.

| Model | Size | Terminal-Bench 2.1 (vendor reported) | | Ripple-65B Preview | 65B dense | 80.52 | | Reflection Beam | 501B | 80.1 | | Meta Muse Spark Open | 816B | 76.4 | | Thinking Machines Inkling | 975B | 63.8 | | NVIDIA Nemotron 3 Ultra | 550B | 56.4 |

The more useful reading is not the exact ranking but the shape of it. Ripple is roughly one eighth the size of Beam and one twelfth the size of Muse Spark Open, yet it sits at the top of this table. If the score holds up under independent testing, it means the efficiency gap between small and huge models is closing fast for terminal coding work. If it does not hold up, the price still makes Ripple worth testing, because a model this cheap does not need to be the best to be useful.

Ripple's other reported scores fill out the picture. It reports 79.1 percent on SWE-bench Pro v2.0 Hard, a test of difficult software engineering tasks. On knowledge and science it reports 89.6 on MMLU, 84.2 on MMLU-Pro and 85.2 on GPQA Diamond. On competition maths it reports 97.1 on AIME 2024, 98.3 on AIME 2025 and 95.0 on AIME 2026. On instruction following it reports 79.2 on IFBench, and on long-context reasoning it reports 75.0 on AA-LCR. The weak spots in its own table are just as informative: it trails rivals on SWE-bench Verified, DeepSWE, SciCode at 39.4, and MCP Atlas. A model that is strong on terminal tasks and weaker on some science coding tests is a specialist, not an all-rounder, and the honest way to use it is for the jobs it is good at.

What does it cost to use?

This is where Ripple gets interesting for anyone who pays API bills. The hosted API is cheap by any standard, and the weights are free if you have the hardware to run them.

| Access route | Input per 1M tokens | Output per 1M tokens | Cached input per 1M tokens | Context | | Together AI | $0.20 | $0.60 | $0.02 | 128K on Together's listing | | OpenRouter via GMI Cloud | $0.20 | $0.60 | $0.05 | 16,384 tokens currently | | Hugging Face weights | Free under Apache 2.0 | Your hardware cost | Not applicable | 1M in the model card |

A few details in that table deserve a closer look. First, Together AI lists the model at $0.20 input and $0.60 output per million tokens, with cached input at $0.02, and gives the endpoint name as ricursive-ai/Ripple-65B-Preview. Second, OpenRouter lists the same $0.20 and $0.60 rates through GMI Cloud, with cached input at $0.05, but its endpoint currently supports only 16,384 tokens of context, so do not assume that price buys the full 1 million token window. Third, Together's own listing shows a 128K context window, which is different from the 1 million tokens in the model card, so check the context limit on whichever route you use before you send a large codebase through it.

For a sense of scale, an input-heavy agentic coding session that sends 500,000 input tokens and receives 100,000 output tokens costs $0.10 plus $0.06, or $0.16, on Together's rates, before any caching discount. A full day of heavy coding agent use, say 5 million input tokens and 1 million output tokens, costs $1.00 plus $0.60, or $1.60. Those are illustration figures on the published rates, not a promise about your usage, but they show why a $0.20 model changes the maths for coding agents, which burn far more tokens than a chat session.

The company also says the whole project cost about $32 million in research and development from scratch, covering test models, data work, evaluations, debugging, synthetic data and the final training runs, plus 1.4 million reinforcement learning rollouts. That is roughly a tenth of what the previous open-weight leader spent, according to launch coverage. Whether the $32 million figure is exactly right matters less than the direction it points: small teams can now build a competitive coding model without a frontier lab budget.

How do you try Ripple-65B today?

There are three practical routes, and the right one depends on what you want to do with it.

The fastest route is the API. On Together AI, create an account, add billing, and call the ricursive-ai/Ripple-65B-Preview endpoint with an OpenAI compatible client. On OpenRouter, the model is listed through GMI Cloud at the same rates, and it supports tool calling and reasoning for agentic workflows. Start with a small coding task you can check yourself, such as a script you already know works, and compare the output and the cost against the model you use now. That gives you a real price and quality comparison in an afternoon.

The second route is the weights. Download them from Hugging Face, where the model card lists the architecture, training stages and benchmark table. At 65 billion dense parameters, this is not a laptop model, but it is far more practical to self-host than a 500 billion parameter model, and one launch report notes it is small enough for a DGX Spark class machine. If you already run open models locally, Ripple is a straightforward addition to your stack, and the Apache 2.0 licence means you can use it commercially without asking permission.

The third route is to wait for independent results. No independent evaluation had surfaced at the time of writing, and the first independent Terminal-Bench runs will tell us whether 80.52 survives outside Ricursive's own harness. If you are choosing a model for a production system this week, run your own benchmark on your own tasks first, and do not switch on the headline alone.

What are the catches?

There are five, and they are all worth knowing before you build on this model.

First, it is a preview. The model card and the launch coverage both describe light post-training, which means the model has had less fine-tuning than a finished release. Expect rough edges in areas outside its strengths, and expect a fuller version later.

Second, every benchmark figure is vendor reported. Ricursive ran the comparisons itself, and some rival scores are its own runs. Independent scores will arrive, and they may move the ranking. Until then, treat 80.52 as a claim to verify, not a fact to quote.

Third, the model trails rivals on several tests in its own table, including SWE-bench Verified, DeepSWE, SciCode and MCP Atlas. If your work is heavy on those kinds of tasks, test Ripple against your current model on your own examples before switching.

Fourth, the context window depends on the route. The model card says 1 million tokens, Together lists 128K, and OpenRouter currently caps at 16,384 tokens. A coding agent that needs to hold a whole repository in context will behave very differently at 16K than at 1M, so confirm the limit where you run it.

Fifth, availability is still spreading. Together AI and GMI Cloud list it, OpenRouter serves it through one provider, and Fireworks support is expected soon. Early reports noted about 85 percent uptime on the single OpenRouter provider over its first day, so if you need reliability now, keep a fallback model configured.

Who gets the most value from this?

Three groups should try Ripple this week. Developers who run coding agents will feel the price difference immediately, because agentic coding is the most token-hungry work most people do, and $0.20 per million input tokens turns a costly experiment into a cheap one. Small teams and indie builders who self-host get a strong open model under a permissive licence, without the hardware bill of a 500 billion parameter model. And anyone evaluating open models for a product should add Ripple to their shortlist, because a 65 billion parameter model that competes with models eight to twelve times its size changes what a small stack can do.

The test to run is simple. Take a coding task you have already solved with your current model, run it through Ripple on the API, and compare three things: whether the result works, how many tokens it used, and what it cost. If all three look good, you have found a cheaper default. If the result is worse, you have spent less than a dollar learning exactly where Ripple falls short, and you still have the weights to try again when the next version lands.

Sources: Ricursive Intelligence launch blog, "Introducing Ripple-65B Preview", 4 November 2026; Ripple-65B Preview model card on Hugging Face; Together AI model page and pricing page, checked 6 November 2026; OpenRouter listing as reported 5 November 2026; VentureBeat, "Ex-Googlers' new Apache 2.0 model Ripple-65B Preview claims the U.S. open coding lead", 5 November 2026; DepthGrid, "Ripple-65B Preview: 65B Dense Open Model, Benchmarks, Apache 2.0", 6 November 2026.

Related reading: TopUseAI Blog, Mistral Large 4.1, Nemotron 3.5 Bolt