Nemotron 3.5 Bolt vs GPT-5.6 Sol: One of These Two Models Has a Price

You cannot compare these two models on cost, and that is the most useful thing about the comparison. NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB is a 24GB file: download it, and the marginal price of the millionth token is the electricity. GPT-5.6 Sol is a hosted flagship billed at $4.00 per million input tokens and $20.00 per million output tokens, with a published p50 time to first token of 1.63 seconds and 535.5 million tokens of traffic across the last seven days. One of those two has a rate card. The other has a hardware prerequisite, which is a different kind of bill that arrives once.

Both models are real and both are shipping. NVIDIA's Nemotron 3.5 Bolt 74B-A9B 24GB checkpoint sat on Hugging Face from 7 October 2026 with nothing attached to it but a model card, and the vendor's own announcement of the Bolt line followed on 15 October — eight days later, in a post that names both this model and a 4GB sibling. GPT-5.6 Sol has been on OpenAI's price list since 9 July 2026. If you are choosing between them this week, the decision is not which one is smarter. It is whether the workload can leave your machine.

What the twenty-one against fifty-nine gap measures

Artificial Analysis scores Nemotron 3.5 Bolt 74B-A9B at 21 on its Intelligence Index, ranking it fourth of 67 models in the comparison class it is placed in. GPT-5.6 Sol sits in the high fifties on the same index version. A gap somewhere near forty points is enormous, and stated that way it ends the argument before it starts.

The gap is also two different measurements wearing one number. The 21 is the best score any open-weight model has achieved at this footprint — Artificial Analysis records the median for comparable open-weight models at 8, and notes that Bolt lands "amongst the leading models in intelligence when comparing to other open weight models of similar size." The approximately 59 is what a frontier lab produces when memory is not a constraint. Neither number is wrong. They are answering different questions: how much intelligence fits in 24GB, and how much intelligence money can buy.

What the panel could not measure is more interesting than what it could. The Artificial Analysis page for Bolt lists speed as N/A and cost per Intelligence Index task as N/A, for the plain reason that there is no hosted endpoint to call. Every speed figure that exists for this model comes from NVIDIA's own bench: the research note states a design target of at least 80 tokens per second of decoding and reports 84 tokens per second on the AIPerf SPEED-Bench on RTX Spark, with 83.99 tokens per second in a llama.cpp run on a Surface Laptop Ultra with 128GB. Those may well be accurate. Nobody outside NVIDIA has reproduced them, and you should treat them as a vendor claim until someone does.

Where the hosted model is structurally ahead

Three differences are not about intelligence at all, and each one disqualifies the local file for a class of work.

• Context — GPT-5.6 Sol serves a 1,050,000-token window with a 128,000-token maximum output; Nemotron 3.5 Bolt serves up to 262,144 tokens. A four-times-wider window is not a luxury feature when the job is a repository or a contract archive in one request. Bolt's window also has to share 24GB with the KV cache, the images, the video frames and the conversation history, which means the practical ceiling arrives earlier than the specification suggests.

• Input types — GPT-5.6 Sol accepts text, images and files; Nemotron 3.5 Bolt accepts text, images and video and returns text. That is a genuine edge for Bolt on video and a real gap on documents, because a PDF passed as a file does not have to be rasterised, paginated or reasoned about frame by frame.

• Availability — GPT-5.6 Sol is a URL. Nemotron 3.5 Bolt is a vLLM deployment requiring Ampere, Hopper or Blackwell silicon, or a llama.cpp run on an RTX Spark or DGX Spark device. One of those you can be using in four minutes.

There is a fourth difference that is less about capability and more about metering: every call to the hosted model produces a line item, and every call to the local one does not. For a high-volume classification or extraction loop, that inversion matters more than forty index points.

What owning the file actually buys

Nemotron 3.5 Bolt 74B-A9B is a sparse mixture-of-experts model with 74 billion parameters and 9 billion active per token, built on a Mamba-2 and Transformer hybrid backbone NVIDIA calls LatentMoE, with a C-RADIO v4-H vision encoder bolted on for the multimodal work. It is a compression artifact in the most literal sense: NVIDIA took the 120B-A12B Nemotron 3.5 Super VL model, applied a router-guided search it calls Star Elastic to shrink it to 74B-A9B, then ran quantization-aware distillation down to 2.0625 bits per expert weight — an IQ2_XXS format — to land inside 24GB without retraining from scratch. The licence is OpenMDW-1.1, which permits commercial and non-commercial use.

Two consequences follow. The first is that the model runs beside other applications on a workstation rather than instead of them, which is what makes a local agent viable at all. The second is that it ships with speculation built in: a LiLiCorr drafter driven through DFlash proposes seven tokens ahead, which is where a meaningful part of that 84 tokens per second comes from.

The vendor's own benchmark table, which nobody outside NVIDIA has reproduced, puts the 24GB checkpoint ahead of every baseline in the same table — a Qwen3.6 35B, a Gemma 4 26B and NVIDIA's own Nemotron 3.5 Lightning 30B-A3B — on GDPVal-AA-V2 (28.47), Terminal-Bench 2.1 (50.98), GPQA Diamond (86.43), SciCode (41.02) and BrowseComp (66.75). Read those as claims about how a compressed 74B compares to uncompressed models a third its size, which is precisely what they are designed to show, and not as a comparison against a hosted flagship.

Pricing the two halves honestly

The hosted half has an unambiguous price. GPT-5.6 Sol is on OrcaRouter's catalogue as openai/gpt-5.6-sol at OpenAI's provider list price passed through with 0% markup added — $4.00 per million input tokens and $20.00 per million output tokens, the same numbers on our page as on the vendor's, which means a vendor-side change lands on our side the same day rather than at the next contract renewal. A 20,000-token input with 3,000 tokens of output costs $0.08 in and $0.60 out, or about $0.68 a call.

The local half has no price we are willing to invent. There is no hosted endpoint for Nemotron 3.5 Bolt, no rate card, and no third-party platform metering it that we can point you at. Pricing it means pricing a GPU, a power draw and an amortisation period — arithmetic that belongs in your spreadsheet and not in ours, because the answer depends entirely on how much you run it.

That asymmetry is where a routing layer does something concrete rather than decorative. Nemotron 3.5 Bolt is not on our catalogue and we do not serve it; nothing here is an offer to host NVIDIA's checkpoint. What a single endpoint does buy is the ability to keep the local and hosted halves of a workflow under one key, so a request can be sent to GPT-5.6 Sol when it needs a million-token window or a file input and to your own Bolt deployment when it does not — with the rule that decides which is which written once instead of hard-coded into every service. If the hosted provider degrades, automatic failover moves the request rather than the request failing.

Which one to pick

• Pick Nemotron 3.5 Bolt 74B-A9B 24GB when the data cannot leave the building, when the workload is video-bearing, when the volume is high enough that per-token billing is the dominant cost, or when you want an agent running on hardware you already own. Accept a 262K window and an index score that is roughly a third of the hosted model's. Verify the throughput claims on your own device before you plan capacity around them.

• Pick GPT-5.6 Sol when the task needs the widest context, file input, the strongest reasoning available, and a support contract behind it. Pay $4 and $20 per million tokens for the privilege, and note that the 128K maximum output is a longer single response than a 24GB memory envelope will comfortably allow.

• Run both when the workflow splits cleanly, which for most agent systems it does. The local file handles the high-frequency, low-stakes, privacy-sensitive half; the hosted model handles the requests that need reach.

The open question is the same one the last compressed Nemotron raised and nobody answered: a first independent throughput and latency measurement of the 24GB weights, run by someone who does not work at NVIDIA. Until that exists, the intelligence floor of this model is externally verified and its speed ceiling is not — and for a local deployment the speed is the half you feel.