Ripple 65B Preview vs Gemini 4 Argon: A Price Cliff Meets Open Weights

On 1 December 2026 — twenty-six days from today — Gemini 4 Argon's published rates double. Input goes from $2.00 to $4.00 per million tokens, output from $10.00 to $20.00, cached input from $0.10 to $0.20, and the batch and priority tiers move by the same factor. Google has been explicit that the current pricing is introductory, not a discount that might quietly become permanent. That deadline is the real axis of this comparison, because Ripple 65B Preview, released on 4 November 2026 under Apache 2.0, has no price that can move at all: at 65 billion dense parameters it is small enough to run on your own hardware, and the cost becomes a GPU bill instead of a token rate.

Whether that trade makes sense depends entirely on how much you use. This is the arithmetic, laid out before the benchmarks.

The numbers on each side

• Gemini 4 Argon, today — $2.00 input / $10.00 output per million, $0.10 cache read, batch and flex at $1.00 / $5.00, priority at $3.60 / $18.00br/> • Gemini 4 Argon, from 1 December — $4.00 / $20.00, cache $0.20, batch $2.00 / $10.00, priority $7.20 / $36.00br/> • Ripple 65B Preview — no vendor rate card; Apache 2.0 weights, self-hosted. Third-party hosted rates sit near $0.20 / $0.60br/> • Independent score — Argon 50 on Artificial Analysis Intelligence Index v5.0.1 at $7.68 per index task; Ripple has no index entrybr/> • Context — Argon 1,048,576 tokens shared across input, thinking and output; Ripple up to 1M claimed, trained through 2M-token long-context stagesbr/> • Output speed — Argon 55 tok/s measured; Ripple 250 tok/s claimed on 2×B200 at batch 8, BF16

Run an agent workload of 2 million input tokens and 400,000 output tokens a month. On Argon's introductory rate that is $4 plus $4 — $8. From 1 December it is $16. On a hosted Ripple endpoint at the catalogue rate it would be roughly $0.40 plus $0.24, about $0.64, before you pay for anything else. That gap is not a rounding error, and it is the reason this comparison is worth more than a spec sheet: at small volumes the difference is noise, and at agent scale it decides the architecture.

What you are actually buying from Argon

Gemini 4 Argon was announced on 30 September 2026 and reached general availability on 13 October. It is a 1-million-token model with a shared context budget — input, thinking tokens and output all draw from the same pool — and a default synchronous output ceiling of 65,536 tokens. Getting close to the full million in one output requires the background long-decode path with polling, in hops of at most 64K. This is the kind of detail that a "1M context" line on a comparison page erases, and it changes what the model is good at: Argon is excellent at reading a large corpus and producing a moderate answer, and awkward at producing a book in one call.

Its measured profile is mid-frontier. Artificial Analysis puts it at 50 on the v5.0.1 index, below Claude Opus 5.5 at 59 and GPT-6.1 Sol at 55, at a cost of $7.68 per index task — more than six times Sol's $1.27 for a lower score. Google's own published tables show it winning narrowly on DeepSWE v1.1 at 77.9 and losing on Terminal-Bench 4.0, FrontierSWE v2 and OSWorld-2.0, with the usual caveat that the vendor selected the rows. Its most distinctive number is a 15% hallucination rate on AA-Omniscience, far below the frontier norm.

None of that is a reason to avoid it. Argon's case is long-document work at a price that, until 1 December, is genuinely aggressive — $2 in for a million tokens is the same as GPT-6.1 Sol and a fifth of Opus 5.5.

What Ripple is, minus the launch gloss

Ripple 65B Preview is an 80-layer dense transformer with a 5,120-wide residual stream, 65 billion parameters all active, and a set of architectural choices Ricursive describes as Gemma 3N-like: extra attention heads, a higher sliding-window-to-global ratio, a smaller sliding window, FFN activation clipping, gated attention, global NoPE and value residuals. It is Apache 2.0 and ships as BF16 instruct, NVFP4 instruct, BF16 base and a DFlash2 variant.

The efficiency claims are the interesting part for this comparison. Ricursive says the whole project cost $32 million, about a tenth of the previous open-weight state of the art on agentic coding, using 1.4 million RL rollouts — roughly seventy times fewer than Reflection Beam — at 40% MFU, and that it sustained 250 tokens per second on two B200s at batch size eight in BF16. Artificial Analysis measures Argon at 55 tokens per second. A 65B dense model at 250 tok/s on two GPUs is a different class of deployment from a frontier API, and it is the only part of the Ripple pitch that does not depend on a benchmark harness.

The context claim needs the same scrutiny Argon's does. Ricursive lists training stages at 8K, 64K, 2M and 128K and markets a 1M window. One third-party catalogue already lists the model with a 16.4K context, which is plainly wrong against the card but should be enough to make you verify the window in your own harness before you design around it. A 1M claim on a two-day-old preview is a claim, not a measurement.

The benchmark problem, in one paragraph

You cannot line these two up on a chart. Argon's published coding numbers are mostly Terminal-Bench 4.0 and DeepSWE v1.1; Ripple's headline is Terminal-Bench 2.1 at 80.52, a different task set with a friendlier ceiling. The only same-version row both publish is DeepSWE v1.1, where Argon reports 77.9 and Ripple reports 44.3 — a result that runs directly against the impression the Terminal-Bench numbers create. Neither figure is independently verified for Ripple, and Argon's are vendor-selected. Treat the comparison as unresolved until a neutral harness runs both.

The decision, in two cases

Bursty, document-heavy, low volume. Argon wins and the December increase probably does not matter. If you process a few hundred long PDFs a month, the difference between $8 and $16 is a rounding error, and a 1M shared window with a low hallucination rate is worth more than a 3× token saving you have to operate yourself. Lock in before 1 December if you can — the rate change applies to usage, not to contracts signed earlier, so a committed-spend agreement is worth asking about this month specifically.

Steady-state, high volume, narrow task. The arithmetic flips. Two B200s at a fixed monthly cost against $4,800 of Argon output tokens at 400K/month-at-scale is not a close call, and Apache 2.0 means no licence negotiation. The catch is that you inherit the serving stack, the quantisation decisions and the evaluation work, and you are betting on a preview checkpoint with no independent score.

OrcaRouter sits on the Argon side of that line. Gemini 4 Argon is one of more than 200 models available on a single key, at 0% markup — Google's list price passed through, so the 1 December move lands here the same day rather than at the next contract renewal. If the price change is what worries you, the routing DSL is the practical hedge: compose Argon for the long-context read and hand the output stage to a cheaper model, and the blended call costs less than either model alone. Ripple is not in the catalogue, so a Ripple-based stack is a self-hosting project rather than a routing decision.

What would change the answer: an independent Ripple score, a reversal or extension of Argon's introductory pricing, or Argon's 1M window turning out to be usable end-to-end without the background path. Watch the first one; the other two are known and dated.