Ripple 65B Preview vs Claude Opus 5.5: Two Agentic Coders, Two Different Exams

Ripple 65B Preview reports 80.5 on Terminal-Bench. Claude Opus 5.5 reports 66.4. Put those two numbers in one column and a 65-billion-parameter open-weight model released on 4 November 2026 by Ricursive Intelligence appears to beat Anthropic's flagship by fourteen points on the benchmark closest to agentic coding work. It does not. Ripple's figure is Terminal-Bench 2.1. Opus 5.5's is Terminal-Bench 4.0. They are not the same exam, they are not scored the same way, and the history of the two versions is the single most useful thing a buyer can know before treating either number as a verdict.

That gap is not a technicality invented to protect the incumbent. It is a documented collapse. When SemiAnalysis ran the same models across both versions on 8 September 2026, Claude Fable 5.1 fell from 91.4 on 2.1 to 55.8 on 4.0, GPT-6 Astra from 88.4 to 57.7, and Gemini 3.8 Flash from 89.4 to 19.1 — a 70-point drop for a model that had not changed. Terminal-Bench 4.0 removed eight tasks, fixed nineteen more, applied a flat eight-hour timeout and added tamper-resistance checks. A 2.1 score and a 4.0 score share a name and nothing else.

What each model actually is

Ripple 65B Preview is a dense model, not a mixture of experts: 65 billion parameters, all of them active on every token, in an 80-layer "thin-and-deep" transformer with a 5,120-wide residual stream. Ricursive's model card lists extra attention heads, a higher sliding-window-to-global attention ratio, a smaller sliding window, FFN activation clipping, gated attention, global NoPE and value residuals — a set of choices it describes as Gemma 3N-like. It is Apache 2.0, trained on 12 trillion tokens, and claims up to 1 million tokens of context.

Claude Opus 5.5 is a closed model released on 22 September 2026, with a 1-million-token window, 128,000 tokens of maximum output, text and image input, and adaptive thinking that is always on — a breaking change from Opus 5, where thinking was opt-in. It exposes five effort levels with medium as the default, which matters: almost every published Opus 5.5 benchmark was run at max or xhigh, not at the setting most API calls will use.

• Parameters — 65B dense, all active vs undisclosed, closed weights

• Licence — Apache 2.0, downloadable vs proprietary, API only

• Context — up to 1M claimed vs 1M, with 128K max output

• Release — 4 November 2026 vs 22 September 2026

• Price — no vendor list price; hosted rates around $0.20 / $0.60 per million tokens vs $4.00 input / $20.00 output, $0.20 cache read

• Independent score — none yet vs 59 on Artificial Analysis Intelligence Index v5.0.1

The one row where the versions match

Terminal-Bench is where the launch wants the argument. SWE-bench Pro is where it gets interesting, and DeepSWE v1.1 is where it gets uncomfortable.

Opus 5.5's SWE-bench Pro figure is 89.9, from Anthropic's system card. Ripple's is 79.1 — but on v2.0-Hard, and its own card separately lists 61.42 on v1.0. Three numbers, two vendors, at least two task sets. Nothing here can be subtracted from anything else.

DeepSWE v1.1 is different. Both labs published a score on it, and neither has re-versioned it in the way Terminal-Bench was re-versioned. Opus 5.5 scores 74.2. Ripple 65B Preview scores 44.3. That is a thirty-point gap on a benchmark both sides chose to publish, and it points the opposite way from the Terminal-Bench chart. DeepSWE is a software-engineering agent benchmark built around real repository work rather than shell tasks, so for the audience Ripple is pitching — coding agents — it is arguably the more relevant of the two.

This is the honest reading of the launch: Ripple's headline is a 2.1 score, its own card shows a 44.3 on the one agentic-coding benchmark where a same-version comparison exists, and the model it is being compared to wins that row by a wide margin. None of which makes Ripple a bad model. It makes the 80.5 a number about a specific exam.

What $32 million buys, and what it does not

Ricursive's most interesting claim is not a benchmark. It is that the whole research and development effort cost $32 million, roughly one-tenth of what the company says the previous open-weight state of the art on agentic coding cost, reached with 1.4 million reinforcement-learning rollouts — about seventy times fewer than Reflection Beam 501B — at 40% model FLOPs utilisation.

If that holds up, it is a bigger deal than 80.5 on any benchmark, because it says the price of entering this capability tier has fallen by an order of magnitude. It also explains the shape of the model: 65B dense at 80 layers is a design you can train and serve cheaply, and Ricursive reports 250 tokens per second on two B200s at batch size eight, before any quantisation. A comparison point: Opus 5.5 measures 112 tokens per second on Artificial Analysis's harness.

What $32 million does not buy is independent verification. There is no published third-party re-run of Ripple's numbers yet, and the model does not appear on the Artificial Analysis Intelligence Index — the standard neutral board — at all. Opus 5.5 sits at 59 on that index, version 5.0.1, at $10.04 of compute per index task. Until Ripple gets an entry, every cross-model claim about it rests on vendor tables.

The price argument, done properly

The tempting sentence is that Ripple is twenty to thirty times cheaper per token. The accurate sentence is that Ripple has no list price, because there is nothing to list a price for: the weights are Apache 2.0 and downloadable. Third-party catalogues quote hosted rates near $0.20 per million input tokens and $0.60 per million output, and those are hosting rates, not a vendor rate card. Self-hosting turns the cost into GPUs and engineering time, which is cheap at steady state and expensive on day one.

Opus 5.5 is the opposite trade. Four dollars in and twenty dollars out per million tokens is a real number you can put in a spreadsheet today, with a $0.20 cache read that makes long agent loops with a stable prefix considerably cheaper than the headline suggests. On OrcaRouter it is one of more than 200 models reachable on a single key at 0% markup — the provider's list price passed through, so an Anthropic price change is live on the same day rather than at the next billing reconciliation.

Ripple is not in the OrcaRouter catalogue, and nothing in this article should be read as us serving it. Its weights are on Hugging Face, and the launch names hosting partners; if you want to call it as an API rather than run it, that is the route.

Which one to pick

If the work is a coding agent on a proprietary codebase with a compliance review, Opus 5.5 is the defensible choice and the benchmark argument is not close: it has an independent index score, a system card, a published SWE-bench Pro number, and a 1M window with 128K of output. You are paying for the audit trail as much as the model.

If the work is a high-volume, narrow agent loop where the prompt is stable, the corpus is yours, and the bill is the constraint, a 65B Apache-2.0 model that fits on two B200s is a genuinely different proposition from a $20-per-million-output API. The question to answer before committing is not "is 80.5 bigger than 66.4" — it is "what does Ripple score on DeepSWE against the model I am replacing", and right now the answer on record is 44.3 against 74.2.

The thing worth watching is not another vendor chart. It is the first independent run of Ripple 65B Preview on Terminal-Bench 4.0. Everything in this comparison waits on that number.