Intern-S2-397B vs Grok 4.6: Who Gets to Turn the Knobs

Intern-S2-397B and Grok 4.6 shipped about a month apart and answer the same underlying question in opposite ways: who gets to control the reasoning. Grok 4.6, the x​AI flagship that went live on August 12, 2026, sells you a dial — a configurable reasoning-effort control on a closed API priced at $2.00 per million input tokens and $6.00 per million output tokens, with a 500,000-token window and x​AI's server-side tools billed on top. Intern-S2-397B, the 403-billion-parameter scientific multimodal model that Shanghai AI Laboratory's InternLM team released under Apache-2.0 on September 13, hands you the whole machine — weights, thinking toggle, vision path, time-series head — and expects you to run it yourself. One is rented with knobs; the other is owned outright. The comparison that matters here is not which scores higher on a benchmark table; it is which kind of control your workload actually needs, and what each kind costs.

Two different dials

The cleanest way to see these two apart is to look at where the reasoning controls live. Grok 4.6 exposes a reasoning-effort setting through x​AI's API, and around it sits a suite of server-side tools — function calling, web search, X search, code execution — that x​AI operates and bills separately. You tune effort, you chain the tools it gives you, and you never touch a weight. The model's behavior is x​AI's property and x​AI's problem; your lever is a parameter in the request.

Intern-S2-397B gives you a different set of levers, and they are lower down the stack. Thinking is on by default and you switch it off per request with enable_thinking=False. Because the model is Apache-2.0, you can also take the weights and fine-tune the reasoning behavior itself on your own data, then serve the result on LMDeploy, vLLM or SGLang. Its input surface is wider than a text-and-image API: it accepts time-series signals and produces forecasts, a capability currently wired through LMDeploy only, and it was trained for long-horizon agent work in sandboxed environments. The trade is equally clear — that level of control is only meaningful if you are prepared to operate the hardware and the serving stack that come with it.

• Reasoning control — Grok 4.6: reasoning-effort dial on a closed API vs Intern-S2-397B: enable_thinking toggle plus full weight-level control under Apache-2.0.

• Tools — Grok 4.6: x​AI's server-side web search, X search and code execution, billed separately vs Intern-S2-397B: tool calling you wire up yourself, plus a time-series forecasting path through LMDeploy.

• Inputs — Grok 4.6: text, image, file vs Intern-S2-397B: text, image, time series.

• Context — Grok 4.6: 500,000 tokens vs Intern-S2-397B: 256K text reasoning, 64K multimodal, both figures from the model card.

• Price — Grok 4.6: $2.00 / $6.00 per 1M, tiering to $4.00 / $12.00 past 200K tokens vs Intern-S2-397B: no rate card; self-host or the official Intern API.

What the numbers actually cover

Every Intern-S2-397B figure below is InternLM's own, run through OpenCompass, VLMEvalKit and AgentCompass and published alongside the weights with no independent reproduction. Grok 4.6's numbers are a different kind of thing: they accumulated through the public arena and Artificial Analysis after the model went live, so they carry a third-party label.

On the independently measured side, Grok 4.6 sits at 44 on the current Artificial Analysis Intelligence Index, with a GPQA Diamond of 94.9, a Humanity's Last Exam of 61.0 and a coding score of 76.8, and Artificial Analysis puts its TerminalBench 2.1 figure near 80.3. On the vendor side, Intern-S2-397B's card reports 89.77 on MMLU Pro, 93.56 on HMMT-2026, 81.68 on MMMU Pro and 68.54 on SWE-bench-Pro, alongside scientific rows where it looks like a different species: 55.71 on Biology-Instructions, 63.28 on SciReasoner 1.5, 53.95 on Mol-Instructions, 62.35 on MolecularIQ. The one number in the vendor table that should make an agent builder flinch is TerminalBench 2.1 at 64.04 — a figure the model's own makers report losing to an older closed generation by a wide margin, in exactly the terminal-work lane a rented frontier model like Grok 4.6 is strongest in.

Read that split as a portrait, not a ranking. Intern-S2-397B is a scientific specialist that is a competent general model; Grok 4.6 is a generalist that has a passing interest in science. The scientific rows being lopsided is expected — no generalist flagship was pre-trained on raw pages of scientific literature with figures, layout and symbolic notation together, and that is precisely the paradigm Intern-S2-397B is built around. Nothing in either evidence base supports "Intern-S2-397B is as good as Grok 4.6 at arbitrary reasoning," and nothing in Grok 4.6's evidence suggests it was tuned for multi-omics sequence analysis.

The price of control

Grok 4.6 has a price you can put in a spreadsheet: $2.00 per million input tokens and $6.00 per million output, with the rate climbing to $4.00 / $12.00 once a prompt crosses 200,000 tokens, and an optional priority tier for faster responses. It is available through OrcaRouter at x​AI's list price passed through at 0% markup, so a rate change at x​AI lands here the same day with no middleman delta — the dial you rent stays priced the way the vendor prices it.

Intern-S2-397B has no published per-token rate at all. The model card references an official Intern API, but the economics people actually want to compare are the self-host ones: a 403B-parameter checkpoint, roughly 807 GB of weights across 188 shards in BF16, so a serious multi-GPU node, with the FP8 build roughly halving the footprint. If you already own that capacity, the marginal cost of the next scientific query approaches electricity, and that is the entire point of the Apache-2.0 position. If you do not own it, the "free" model is a capital-expenditure pilot, not a line item.

What a router buys in this specific matchup is the seam rather than the price. Grok 4.6 sits on the key you already have for general production traffic; Intern-S2-397B is not routed on OrcaRouter — it is a brand-new checkpoint, and your path to it is the vendor's own API or a self-hosted deployment. Route the general, latency-sensitive reasoning to the metered model, keep the scientific batch work on the model you run, and let automatic failover cover you when either side's endpoint is unhealthy. The boundary between them becomes a routing rule instead of a second integration.

Which one to pick

Pick Grok 4.6 when the job is general, the reasoning has to come back fast and correct, and you do not want to own the stack that produces it. Its 500K window, its measured GPQA Diamond at 94.9 and its tool suite are production features, and the $2/$6 meter is what you pay for not operating anything.

Pick Intern-S2-397B when the input is scientific — a figure-dense paper, a molecular diagram, a time-series signal — and when the reasoning has to happen on data that cannot leave your environment, or on a behavior you want to fine-tune yourself. Apache-2.0 makes that possible; no rented API does. Budget for the hardware, expect no independent scoreboard for a while, and treat every number on its card as a claim until someone outside InternLM reproduces one.