Ripple 65B Preview vs GPT-6.1 Sol: The Row Missing From the Leaderboard

Every serious model comparison in this price bracket now starts at the same place: Artificial Analysis's Intelligence Index, version 5.0.1, the rebuild that replaced 60% of the composite with private or held-out test sets on 27 October. Claude Opus 5.5 reads 59 there. GPT-6.1 Sol reads 55. Gemini 4 Argon reads 50. Grok 4.7 reads 35. Ripple 65B Preview reads nothing, because it has no entry. The model that launched on 4 November 2026 with the loudest coding claims of the week is the only one in this group that a buyer cannot place on the neutral board — and that absence, more than any vendor chart, is what should shape the decision.

GPT-6.1 Sol is the useful foil precisely because it is the value case on that board: 55 on the index for $2.00 per million input and $10.00 per million output, and — the number that actually decides agent budgets — $1.27 of compute per index task against Opus 5.5's $10.04. If Ripple is going to displace anything in a production coding stack, Sol is the incumbent it has to beat on economics, not on a Terminal-Bench percentile.

Why there is no Ripple row

Artificial Analysis needs a served endpoint and stable weights to run its harness. Ripple 65B Preview is two days old, released as weights rather than as a vendor API, and the launch page names hosting partners while the Hugging Face card simultaneously states the model "isn't deployed by any Inference Provider". That contradiction alone explains a missing row: until a host is serving a stable checkpoint, there is nothing for an evaluation harness to call.

There is a second reason, less flattering. The index is deliberately built from held-out and private sets, refreshed on a schedule. A model can be excellent on public benchmarks that its training data has seen and ordinary on the private ones. Ripple's launch numbers are all public benchmarks: Terminal-Bench 2.1 at 80.52, SWE-bench Pro v2.0-Hard at 79.1, DeepSWE v1.1 at 44.3, MMLU 89.6, GPQA Diamond 85.2, AIME 2025 at 98.3. Those are strong figures. They are also, without exception, the benchmarks a team chooses when it wants to publish a launch.

That is not an accusation of misconduct. It is the ordinary state of a two-day-old preview, and the correct response is to treat Ripple as unranked rather than as ranked-first.

What Sol actually is

GPT-6.1 Sol shipped on 29 September 2026 at OpenAI's DevDay as the tier below GPT-6 Astra, replacing GPT-6 Sol. It carries a 1,050,000-token context window, 128,000 tokens of maximum output, a knowledge cutoff of 30 April 2026, text and image input, and five reasoning levels from low to max with medium as the default.

The pricing structure has one trap worth knowing about. Below 272,000 input tokens the rate is $2.00 in and $10.00 out, with cached input at $0.10 and cache writes at $2.50. Above that threshold the entire request is rebilled at $4.00 in and $15.00 out — not just the excess. A 300K-token prompt therefore costs twice as much as a 270K one, which makes chunking a real optimisation rather than a nicety.

• Price — Sol $2.00 / $10.00 per million, $0.10 cache read; Ripple has no list price, hosted rates near $0.20 / $0.60

• Cost per index task — Sol $1.27; Ripple not measured

• Context — Sol 1.05M with a 272K billing cliff; Ripple up to 1M claimed

• Weights — Sol closed; Ripple Apache 2.0, downloadable

• Output speed — Sol 55 tok/s measured; Ripple 250 tok/s claimed on 2×B200

• DeepSWE v1.1 — Sol 75.22 (vendor) vs Ripple 44.3 (vendor), the same benchmark version

The one clean comparison, and it is not close

Amid the version churn there is a benchmark both labs ran and both published: DeepSWE v1.1. OpenAI reports 75.22 for Sol at high effort. Ricursive reports 44.3 for Ripple. That is a 31-point gap on a software-engineering agent benchmark, published by both sides, with no harness-version ambiguity to hide behind.

Set that against the Terminal-Bench picture and the tension is obvious. Ripple publishes 80.52 on Terminal-Bench 2.1; Sol's third-party Terminal-Bench 4.0 figure is around 55. On one benchmark Ripple looks like a frontier competitor, on the other it looks a generation behind. The difference is the exam: 2.1 is an 89-task patch release of the 2.0 set that fixed 28 tasks for dependency and specification problems, while 4.0 is a different task set with adversarial-cheat checks and an eight-hour flat timeout that roughly halved every frontier model's score when it arrived in late August.

The reasonable synthesis is that Ripple is a genuinely capable 65B model whose published strengths are on the older, more saturated exam, and whose one same-version comparison against a frontier model is a 31-point deficit. Sol is the safer bet on coding quality; Ripple is the more interesting bet on cost of ownership.

Where Ripple wins without a benchmark

Three things are true of Ripple and not of Sol, and none of them require a leaderboard.

The first is the licence. Apache 2.0 on a 65B dense model means fine-tuning is permitted, redistribution is permitted, and a security review has nothing to argue about. A closed frontier model, however cheap per token, cannot be fine-tuned on your codebase and cannot be run in an air-gapped environment.

The second is portability. Ricursive's card describes a model designed for efficient serving on NVIDIA hardware and small enough for a DGX Spark, with an NVFP4 checkpoint and a Colab notebook provided. That is a different deployment shape from an API: it is a model you can put in a CI runner or on a developer's desk.

The third is the training economics. Twelve trillion pretraining tokens, a stated $32 million total budget, 1.4 million RL rollouts and 40% MFU add up to a claim that this capability tier is now reachable for eight figures. If that generalises, the frontier API price floor has a clock on it.

How to run both without betting on either

The practical answer for the next quarter is not to choose. GPT-6.1 Sol is in the OrcaRouter catalogue, on the same key as more than 200 other models, at 0% markup over OpenAI's list — so the $2/$10 rate and any future change to it pass through untouched. That matters here because Sol's price is the thing Ripple is being measured against; a router that marks up would corrupt the comparison.

For a stack that wants Ripple in the loop, the architecture is a split rather than a swap: call Sol or another frontier model for the planning and review steps where the 31-point DeepSWE gap is expensive, and route high-volume, low-ambiguity generation to a self-hosted 65B where the marginal cost collapses. OrcaRouter's routing DSL exists for exactly that composition — several models behind one call, with automatic failover if a provider degrades — though Ripple itself would sit outside it, since we do not serve it.

Three questions worth answering before you commit

Is Ripple 65B Preview better than GPT-6.1 Sol?

Not on any evidence that exists today. On the single benchmark both published at the same version, Sol leads by 31 points. Ripple's headline advantage comes from a different Terminal-Bench version with a lower ceiling. Until a neutral harness scores Ripple, "better" is a claim about price, not capability.

Can I call Ripple as an API, or do I have to host it?

Both are in flux. The weights are unambiguously downloadable from Hugging Face under Apache 2.0, in BF16 and NVFP4. Hosted access is announced but not settled — the model card says no inference provider serves it, while the launch page names partners. Assume self-hosting until an endpoint responds.

What does Sol cost above 272K input tokens?

Double on input, one and a half times on output, applied to the whole request: $4.00 and $15.00 per million. Below that line it is $2.00 and $10.00. It is the single largest billing surprise in this comparison and it has nothing to do with Ripple.

The number to watch is not another vendor table. It is the first Artificial Analysis entry for Ripple 65B Preview on index v5.0.1. Everything above is provisional until that row exists.