Ricursive Intelligence, founded in late 2025 by Anna Goldie and Azalia Mirhoseini — the pair behind Google's reinforcement-learning chip-design software AlphaChip — released Ripple-65B Preview on November 4, 2026, under an Apache 2.0 license that permits commercial use, modification and redistribution without a model-license royalty.

The company describes it as a general-purpose variant of the proprietary model behind its AI chip-design platform. Weights are on Hugging Face in BF16 and NVFP4, with hosted access named for Together AI, GMI Cloud and OpenRouter, and Fireworks support expected soon.

The model is a dense 65-billion-parameter Transformer with 80 layers, pretrained from scratch on 12 trillion tokens, with a context window Ricursive says stretches up to 1 million tokens.

The upset is the size math. Ricursive reports Ripple scoring 80.5 on Terminal-Bench 2.1, just ahead of Reflection Beam at 80.1 and Meta's Muse Spark Open at 76.4 — a margin of roughly 0.4 percentage points over a rival with 501 billion total parameters, per Crypto Briefing.

On SWE-bench Pro v2.0 Hard, Ricursive's launch graphic places Ripple first among the compared group at 79.1. The company says the results come with roughly 15 times fewer total parameters than Thinking Machines' 975B Inkling.

Crypto Briefing notes the model is not a clean sweep, however: Ricursive's own figures show it trailing larger models on other knowledge and reasoning assessments.

VentureBeat reports discrepancies inside Ricursive's own materials: the model card lists SWE-bench Pro v2.0 Hard scores of 52.3 for Meta and 2.0 for Inkling versus 72.4 and 39.6 in the launch graphic, and acknowledges difficulty reproducing agent benchmarks, noting changing the evaluation framework affected scores.

On DeepSWE 1.1, Meta scores 60.0 against Ripple's 44.3. Reflection's own published comparison lists several Chinese-developed models above Ripple's Terminal-Bench result.

OrcaRouter highlights the cleanest frontier comparison: OpenAI reports 75.22 for GPT-6.1 Sol on DeepSWE v1.1 at high effort, a 31-point gap over Ripple's 44.3 on the same benchmark version — while Ripple's Terminal-Bench lead comes on version 2.1, not the tougher Terminal-Bench 4.0 set.

Ricursive puts total development cost — including preliminary experiments, dataset preparation, evaluations and final runs — at about $32 million, per VentureBeat, which notes this estimate is not a directly comparable accounting disclosure from both companies.

The efficiency case rests on reaching its Terminal-Bench result after only 1.4 million reinforcement-learning rollouts, which Ricursive says is roughly 70 times fewer than Reflection's reported 100-million-plus rollout run.

OrcaRouter points out Ripple has no entry on Artificial Analysis's Intelligence Index v5.0.1, where GPT-6.1 Sol reads 55, and that the Hugging Face card states the model 'isn't deployed by any Inference Provider' even as the launch page names hosting partners.

Pricing is clearest on OpenRouter's GMI Cloud endpoint at $0.20 per million input and $0.60 output tokens, cached input at $0.05, with only 16,384 tokens of context currently served despite the advertised million-token window.

For builders, Ripple matters in three ways that do not require a leaderboard, as OrcaRouter argues: an Apache 2.0 license permitting fine-tuning on your codebase and air-gapped deployment, an NVFP4 checkpoint and dense architecture suited to a DGX Spark or CI runner, and training economics suggesting this capability tier is now reachable for eight figures.

If that generalizes, the frontier API price floor has a clock on it. The number to watch is the first Artificial Analysis entry for Ripple 65B Preview on index v5.0.1 — everything above is provisional until that row exists.

Until then, the sensible stack may be a split: frontier models for planning and review steps where the DeepSWE gap is expensive, and a self-hosted 65B for high-volume, low-ambiguity generation.