GPT-6.1 Sol vs Intern S2: One Has a Price, the Other Has Weights

The strangest thing about comparing GPT-6.1 Sol with Intern S2 is that one side of the comparison has no price you can look up. Shanghai AI Laboratory's Intern-S2-397B, published 2026-09-13, ships Apache 2.0 weights on Hugging Face, a technical report, and a vendor-hosted API endpoint — and no published per-token rate anywhere. Open​AI's GPT-6.1 Sol, released 2026-09-29, costs $2.00 per million input tokens and $10.00 per million output, with a documented tier above 272,000 prompt tokens, and every one of those numbers appears on a public pricing page. So the honest way to frame this pairing is not "which is better" but "which of the two can you actually forecast a bill for" — and the answer differs for reasons that have nothing to do with capability.

Everything below splits cleanly into two evidence bases, and keeping them apart is most of the work. GPT-6.1 Sol's figures are independent, from Artificial Analysis, except where a sentence says otherwise. Intern-S2's figures are InternLM's own, unreproduced by any third party, and the sentences carrying them say so.

Two systems under one name

Intern S2 is not a single model, and the confusion is worth clearing before anything else. The family contains at least two published artifacts:

• Intern-S2-Mobius — a 35-billion-parameter open-weight model under Apache 2.0, continually pre-trained from Qwen3.5-35B, using a Mobius-v0 architecture with a shared Memory of 2,560 experts, four Reasoner blocks and 40 layers. Deployment guidance describes roughly 3 billion active parameters per token with a 3:1 hybrid attention pattern, a 256,000-token context and a vision tower. Weights landed 2026-07-29; the GitHub repository and technical report followed on 2026-08-05; vLLM support merged 2026-08-06.

• Intern-S2-397B — the flagship, published 2026-09-13. Roughly 397 billion total parameters in a Qwen3.5-MoE architecture, 512 experts with 10 active, Apache 2.0 weights on Hugging Face alongside a model card, an official Intern API entry, and vLLM Day-0 support announced the following day, with a recipe published for FP8 inference on eight H100s or eight H200s. Its published limits are 256,000 tokens of text context and 64,000 of multimodal context.

GPT-6.1 Sol is the simpler object: one snapshot identifier, gpt-6.1-sol, no dated variant, 1,050,000 tokens of context, 128,000 maximum output, text, image and file input, five reasoning effort levels with medium as the default, and a stated knowledge cutoff of April 30, 2026.

What each side's numbers rest on

This is the axis on which the two are furthest apart, and it is not close.

GPT-6.1 Sol's independent record, all from Artificial Analysis:

• AA Intelligence Index — 51.8, rank 10 of 147 models, 93.2nd percentile.

• Humanity's Last Exam — 52.9%.

• Long-Context Recall — 83%.

• SciCode — 54.2%.

• Terminal-Bench 4.0 — 56.06%.

Intern-S2's record, all InternLM's own, none reproduced by a third party:

• MMLU Pro — 89.77% for the 397B flagship.

• HMMT-2026 — 93.56%.

• MMMU Pro — 81.68%.

• SWE-bench-Pro — 68.54%.

• TerminalBench 2.1 — 64.04%.

• Scientific domain rows — Biology-Instructions 55.71%, SciReasoner 1.5 63.28%, Mol-Instructions 53.95%, MolecularIQ 62.35%.

Those two lists are not comparable, and it is worth being explicit about why. The first set is a party with no stake in the outcome asking the same questions of many models under one protocol. The second set is the lab that trained the model reporting how its own model did. That does not make the second set false — labs with a published technical report and downloadable weights have a reputation stake in the numbers — but it does mean the numbers are a claim rather than a measurement, and there is currently no independent evaluator, and no independent inference provider, publishing Intern-S2 scores anywhere.

The rows that carry real information

Two things in InternLM's own materials are more informative than the headline scores, precisely because they are not flattering.

The first is the Mobius benchmark table for the 35B model, which compares it against the Qwen3.5-35B base it was continually pre-trained from. That is a genuinely useful comparison — same lineage, same scale, so the delta isolates what the architecture change bought. Mobius wins seven of nine general benchmarks and loses Humanity's Last Exam and UGD hard. Reading that honestly, the Mobius architecture buys moderate general gains and gives up something on the hardest reasoning set. The science-domain rows in the same table are larger, but they are the signature of domain continued pre-training rather than of the architecture — the training data is where the biology and chemistry numbers come from, not the expert layout.

The second is the efficiency story, which is the most concrete claim in the release: 2.9 times the throughput of the baseline at batch 16, rising to 4.6 times at batch 256. Those are ratios against an internal baseline, measured by the lab, on hardware the lab chose — and they are also the kind of claim that is easy to verify once you have the weights, which you do, under Apache 2.0.

That verifiability is the actual value of an open release, and it points somewhere the headline scores cannot: you do not have to trust the throughput ratio, because eight H100s and a vLLM recipe are now the only things between you and checking it.

Open weights against a priced endpoint

Here is where the two models stop being alternatives and start being different decisions.

GPT-6.1 Sol is a service. You pay per token, you get a 1,050,000-token window, 128,000 output tokens, five effort levels, and no operational surface at all. The bill scales linearly with use and the marginal cost of a spike is a line item, not a capacity problem. The trade is that you cannot fine-tune it — Open​AI publishes no fine-tuning path for this model — you cannot run it on your own hardware, and your unit economics are set by someone else's pricing page.

Intern-S2-397B is a download. Apache 2.0 means you can run it, modify it, fine-tune it, and serve it inside your own network without any data leaving it. Eight H100s is a real capital cost and a real operational burden, but above a certain sustained volume that capital cost is cheaper than a per-token rate, and it is the only option if the data cannot leave your perimeter at all. The trade on the other side is that everything about quality is now your problem: there is no independent benchmark to plan against, no external provider to fail over to, and no vendor to escalate a regression to.

Between those two positions, the deciding variable is almost always utilisation. A monthly burst of a few hundred million tokens does not amortise eight GPUs. A steady multi-billion-token monthly flow inside a regulated perimeter frequently does.

The number that is missing from one column

Intern-S2-397B has no published per-token rate. That is not an oversight to be worked around with an estimate; it means the standard comparison this blog usually runs — cost per million tokens against cost per million tokens — cannot be run at all against the self-hosted path.

The comparison that can be run is against the vendor-hosted endpoint. InternLM operates an official Intern API at chat.intern-ai.org.cn/api/v1 with the model identifier intern-s2-397b, and the rate for that endpoint is what you would need quote-to-quote against GPT-6.1 Sol's $2.00 and $10.00. Until that rate is published, the only honest statement is that the API path's price is unknown, and the self-hosted path's price is a function of your GPU contract and your utilisation rather than of a rate card.

Which means the arithmetic this article can do is narrow, and worth doing anyway: ask what your own cost per million tokens would be at your utilisation, and compare it to $2.00/$10.00 with the above-272K tier applied. For most teams the answer to that arithmetic is that the closed endpoint wins below a few billion tokens a month and the open weights win above it — and the entire question is which side of that line the workload is on.

Running the two of them differently

There is no single answer here, because the two models are not reached the same way, and pretending otherwise would be the same error as quoting the index rows side by side.

GPT-6.1 Sol is on OrcaRouter as openai/gpt-6.1-sol, at Open​AI's list price with 0% markup — $2.00 input, $10.00 output, cached input at $0.10, the above-272K tier applied exactly as Open​AI sets it. Because we pass the provider price through rather than marking it up, an Open​AI rate change is live here the same day rather than at the next billing cycle, and automatic failover sits underneath the route. That last part matters more for an unproven model than a proven one — which is the same argument that applies to the Intern-S2 side, except that the Intern-S2 side cannot use it.

Intern S2 is not on OrcaRouter and we do not host it. It is reachable through InternLM's own API, and it is downloadable from Hugging Face under Apache 2.0 for anyone with the hardware. If you want to try it, the weights are the low-commitment path — you can evaluate the 397B flagship on your own prompts before deciding whether to buy GPUs for it, and the Mobius 35B is small enough to test the family's behaviour on a single node.

The verdict this pairing earns is not a ranking, because the two are not buying the same thing. If you need a model with an independently measured capability level, a published output ceiling, a known per-token price, and no operational surface, GPT-6.1 Sol is the answer and it is not a close call — it is simply the only one of the two that offers those four things. If your constraint is that the data cannot leave your network, or your volume is high enough that GPUs beat tokens, Intern-S2-397B is the better answer and the missing benchmark is a risk you are choosing to carry rather than a reason to walk away. What you should not do is compare the two on their published numbers. InternLM's own eval and Artificial Analysis's leaderboard are not the same instrument, and reading them as one table would produce a confident answer to a question nobody asked.