Ripple 65B Preview vs DeepSeek V4.1 Pro: Only One of These Models Exists

DeepSeek V4.1 Pro has no model card, no API identifier, no price and no benchmark. It has never been released. It exists as a name inside a single sentence in DeepSeek's release notes for V4.1 Flash on 10 September 2026, and search demand for it has been running for eight weeks regardless. Ripple 65B Preview, by contrast, shipped yesterday — 4 November 2026, from Ricursive Intelligence, Apache 2.0, 65 billion dense parameters, twelve trillion training tokens. A head-to-head between the two is therefore not a comparison of models. It is a comparison of a model and a rumour, and the useful thing to do is put Ripple against the DeepSeek models that actually answer HTTP requests.

The sentence that created a model

On 10 September 2026 DeepSeek published V4.1 Flash along with a migration notice: requests to deepseek-v4-pro would be routed to the new Flash model "until V4.1-Pro launches". That is the entire documentary footprint. The reroute itself was withdrawn after developer pushback and DeepSeek said V4 Pro would continue with billing unchanged. The parenthetical survived, and tracker sites, aggregators and comparison pages have been building a model around it ever since.

What makes the phantom unusually sticky is that the naming is plausible. In the V4 generation, Pro followed Flash by roughly two weeks. On that rhythm V4.1 Pro was due in late September. It is now 5 November, and DeepSeek's own pricing page lists exactly two chat models: Flash, on the 1017 checkpoint, and Pro, still on the 0813 build. Calling deepseek-v4.1-pro on third-party gateways returns a model-not-found error. There is nothing to compare against.

What DeepSeek actually sells today

Three models are live, and the differences between them matter more than the missing fourth. Prices below are DeepSeek's off-peak rates per million tokens; peak pricing, 01:00–04:00 and 06:00–10:00 UTC on weekdays, is double.

• V4.1 Flash (1017) — 1M context, 384K max output, vision supported, $0.15 input / $0.60 output, $0.003 cache hit. The 1017 checkpoint replaced the September build's post-training; it is the model DeepSeek wants you to use.

• V4 Pro (0813) — 1M context, 384K max output, no vision, $0.66 input / $1.98 output, $0.022 cache hit. Older architecture, four and a half times the Flash price, still the only "Pro" on the price list.

• V4 Flash (0731) — 284B total parameters with 13B active, 1M context, MIT licence, $0.14 / $0.28. This is the one that matters for the Ripple argument.

V4.1 Pro would slot above V4 Pro on that list if it shipped. Until it does, anybody quoting you a V4.1 Pro price is quoting either V4 Pro's rate card or a guess.

The comparison that can actually be made

Ripple's launch claim is a Terminal-Bench 2.1 score of 80.52, described as US open-weight state of the art for agentic coding, ahead of Reflection Beam 501B at 80.1, Meta's Muse Spark Open 816B at 76.4, Thinking Machines' Inkling 975B at 63.8 and Nemotron 3 Ultra 550B at 56.4. Every one of those competitors is American, which is what makes the claim true as written.

DeepSeek publishes Terminal-Bench 2.1 scores too. V4 Flash 0731 reports 82.7 on the same version, measured by DeepSeek on its own harness — the same caveat that applies to Ricursive's number, and to everyone else's. If you take both vendor figures at face value, a 284-billion-parameter mixture-of-experts model with 13 billion active parameters beats Ripple by 2.2 points on the exact benchmark Ripple chose to lead with. It also costs less: $0.14 in and $0.28 out against hosted Ripple rates around $0.20 and $0.60.

The gap widens on DeepSWE v1.1, where both labs publish a number. Ripple 65B Preview reports 44.3. DeepSeek V4 Flash reports 54.4 — a ten-point lead, up from 7.3 in the April preview build. And for completeness, the 1017 build of V4.1 Flash reports roughly 77 on DeepSWE in DeepSeek's internal testing, which is a different and much stronger model than the one Ripple is being measured against, at $0.15 / $0.60.

The critique that landed

The most-shared response to the launch, from the DeepSeek tracker account teortaxesTex, was two sentences: that Ripple actually loses to V4 Flash 284B, and that the launch leaned on the previous generation of benchmarks. On the numbers above, both halves hold. The 2.1-versus-4.0 distinction is real — Terminal-Bench 4.0 is a different task set with tamper-resistance checks and a flat eight-hour timeout, and it cut every frontier model's score roughly in half when it landed on 28 August. Ripple published on 2.1, where the ceiling is friendlier.

That is a fair hit, but it is not a knockout, and the distinction matters. Ricursive's competitors on that chart were measured on 2.1 as well, so the internal ordering of the launch — Ripple above Beam, Muse Spark Open, Inkling and Nemotron 3 Ultra — is a like-for-like comparison. What is not like-for-like is any attempt to read 80.5 against a model that publishes 4.0, which is most of the frontier.

The parts of the launch that stand on their own

Two claims do not depend on a benchmark version. The first is the budget: Ricursive says the full effort cost $32 million, about one-tenth of the previous open-weight state of the art on agentic coding, reached with 1.4 million RL rollouts — roughly seventy times fewer than Reflection Beam — at 40% MFU. If that survives scrutiny, the interesting story is not that Ripple is the best coding model. It is that a 65B model got within a couple of points of a 284B one for a tenth of the cost.

The second is the licence. Apache 2.0 is a materially different proposition from MIT-with-caveats or a research-only card: it is the licence a legal team signs without a meeting. The weights are on Hugging Face as BF16 instruct, NVFP4 instruct, BF16 base and a DFlash2 variant. One inconsistency worth flagging: the Hugging Face card states the model "isn't deployed by any Inference Provider", while the launch page names several hosting partners. Treat availability as "the weights are definitely downloadable; hosted endpoints are still settling".

How to act on this

You cannot benchmark against V4.1 Pro, because it does not exist. You can benchmark against V4 Flash and V4.1 Flash, and you should, because they bracket the decision: 82.7 and roughly 77 on the relevant coding benchmark, at $0.14/$0.28 and $0.15/$0.60 respectively, both open-weight, both 1M context. OrcaRouter carries the DeepSeek line — deepseek-v4-flash, deepseek-v4-pro and deepseek-v4.1-flash among the 200-plus models on one key, at provider list price with no markup, so a DeepSeek price move is live here the same billing day. Ripple is not in that catalogue; it is self-host or a partner endpoint.

If what you actually wanted was DeepSeek's next flagship, the honest advice is to stop waiting on a name. V4.1 Flash 1017 is DeepSeek's current best coding model, it is openly licensed, and it is a third of the price of V4 Pro. If what you wanted was a cheap 65B you can run yourself, Ripple is real today and its launch numbers put it close to V4 Flash on Terminal-Bench while trailing on DeepSWE.

The trigger to revisit this page is specific and easy to check: a deepseek-v4.1-pro model id that returns tokens instead of a 404. When that happens, the comparison stops being hypothetical, and the first thing to look at will be which benchmark version DeepSeek chose.