GPT-6 Astra vs Intern-S2: What Each One Actually Accepts

The two models in this comparison are both described as multimodal, and the word is doing almost no work in either case. GPT-6 Astra, which O​penAI released on September 3, 2026, takes text and images — and file input, according to router listings that O​penAI's own documentation does not confirm. Intern-S2-397B, the 403-billion-parameter scientific model Shanghai AI Laboratory's InternLM team released under Apache-2.0 on September 13, 2026, takes text, images, and time-series signals, with the forecasting path wired through LMDeploy rather than through the model's own default serving recipe. Neither accepts audio. Neither accepts video. The interesting question is not which is more multimodal; it is which of these two input surfaces your work actually produces, because they barely overlap and neither one substitutes for the other.

Two input surfaces that only look similar

Line the specs up on the dimensions where a real workload makes contact:

• Text — both: GPT-6 Astra up to 922,000 input tokens in one request; Intern-S2-397B evaluated at a maximum inference length of 256,000 tokens for text reasoning

• Image — both, but differently: Astra takes an image alongside a long document in the same request; Intern-S2-397B was pretrained to read raw pages of scientific literature with symbolic and visual content modelled together, which is a claim about how it was trained rather than what it accepts

• Time series — Intern-S2-397B only, via its forecast head, currently through LMDeploy

• File input — GPT-6 Astra according to router listings; O​penAI's own page lists text and image only, so treat it as unconfirmed

• Audio and video — neither

• Output — text on both, and only text

• Reasoning control — GPT-6 Astra exposes reasoning.effort at low, medium, high, xhigh and max; Intern-S2-397B runs with thinking on by default and switches it off per request with enable_thinking=False

• Weights — Intern-S2-397B Apache-2.0, downloadable in FP8 and BF16; GPT-6 Astra closed, API-only

The time-series row is the one that decides this matchup for a subset of buyers and eliminates it for everyone else. Time-series forecasting is not a capability you can approximate with a long text prompt — the input is numeric with its own structure, and a model that was not trained on it will not acquire it because you serialise the numbers into a prompt. If that is your workload, the comparison is over and no benchmark table is needed. If it is not, the row is irrelevant and the rest of this page is what you are reading for.

The context number that matters is the evaluated one

Marketing pages quote the window; benchmark tables quote the length the model was actually run at, and the second number is the one that predicts behaviour. The gap here is instructive.

GPT-6 Astra's window is 1,050,000 tokens with a maximum of 922,000 in and 128,000 out, and its documented service bands are honest about the cost of using it: requests above 272,000 input tokens are repriced at $20.00 input and $75.00 output per million, with cache reads at $2.00 and cache writes at $25.00, applied to the whole request rather than the excess. Below that line the meters read $10.00, $1.00 and $12.50, output at $50.00.

Intern-S2-397B's card states its own limit in the methodology note rather than in a headline: text reasoning benchmarks were run with a maximum inference length of 256,000 tokens, and multimodal benchmarks at 64,000. That is not the model's ceiling — the model card does not publish one in the form a buyer would recognise — but it is the length at which every number on the card was produced. A capability that has only been demonstrated inside 64,000 tokens of image context is a capability that has been demonstrated inside 64,000 tokens of image context, and a page that prints the same score without the length is printing half a fact.

This is the discipline both of these models deserve and neither vendor's material makes easy. Astra's vendor numbers come with a harness attached that changes them by 37 points on one benchmark; Intern-S2's come with a maximum inference length attached that governs whether any of them transfer to your workload. Read both footnotes before reading either table.

What Intern-S2-397B's numbers are, and who produced them

Every figure below is InternLM's own, run through OpenCompass, VLMEvalKit and AgentCompass by the team that trained the model, and published alongside the weights. None of it has been independently reproduced. There is no page for Intern-S2-397B on Artificial Analysis, and that absence matters more than usual, because it means there is no third party to check the general-capability rows against.

The scientific rows are where the model is strongest and where the case for it is clearest: 55.71 on Biology-Instructions, 63.28 on SciReasoner 1.5, 62.35 on MolecularIQ, 53.95 on Mol-Instructions. Alongside those, respectable general rows: 89.77 on MMLU Pro, 81.68 on MMMU Pro, 93.56 on HMMT-2026, 68.54 on SWE-bench-Pro. And one row that tells you what kind of model this is — TerminalBench 2.1 at 64.04, which by the lab's own table loses to an older closed generation. Intern-S2-397B is a scientific instrument that is competent in general, not a generalist with a scientific speciality bolted on.

What GPT-6 Astra's numbers are, and where they are weakest

Astra's record comes in two halves and they do not agree on emphasis. Artificial Analysis measures it at 53 on its current Intelligence Index over ten evaluations, at $3.26 per completed index task, generating roughly 60M output tokens across that run against a board median near 85M, at 44.6 output tokens per second. That is an independent evaluation with a published methodology, and it places the model seventh of 216 on intelligence and 168th of 185 on speed.

The vendor half is a table of its own: 98% on FrontierMath Tier 4, described as saturating the tier; 99.9% on ARC-AGI-3; 100% on ExploitBench, with Astra the first O​penAI model to cross the company's Critical cybersecurity capability threshold; 57.9% on Terminal-Bench 4.0 against 37.3% for GPT-5.6 Sol. Each figure is attributable to O​penAI and none has been reproduced. The ARC-AGI-3 number is the one that has travelled furthest and is worth the most scepticism: ARC Prize ran the model under two harnesses and published both, recording 62.7% under its standard harness and 99.9% under a provider adapter that preserves reasoning state between requests. The 37-point spread is a measurement of the harness, not a scandal — O​penAI documented both settings — but any comparison that puts the 99.9% against another model's standard-harness figure is comparing two experiments and calling it a benchmark.

One more asymmetry belongs here, because it is the reason the independent number is the one to plan against. Intern-S2-397B's card has no independent column at all, so its numbers are the only numbers, and they carry the vendor's authority and nothing else. Astra has both, and they disagree about how good it is at general reasoning in a way that one of them cannot be wrong about — the vendor table says saturated, the index says seventh.

"Free weights" is a capital decision

Intern-S2-397B is Apache-2.0, and the practical consequence is arithmetic rather than licensing. The published checkpoint is 403 billion parameters, and the BF16 repository holds roughly 807 GB — before you count a serving stack, a KV cache, or the second copy you keep for a rollback. An FP8 build exists and roughly halves the footprint. The lab documents LMDeploy, vLLM and SGLang as supported servers, and its own tool-calling example is an O​penAI-compatible endpoint on an LMDeploy server, so the integration path is not exotic. The hardware is.

That produces two very different economics from the same licence. If you already own a multi-GPU node with spare capacity, the marginal cost of the next scientific query approaches electricity, and no per-token competitor can beat that — which is the entire reason open weights exist. If you do not own it, "free" means buying or renting capacity, and on rented capacity the comparison against a metered API is a straightforward arithmetic question about utilisation that a model card cannot answer for you. Astra's side of the same question is simpler: a rate card, no operations team, and a long-context band at $20.00 / $75.00 that punishes exactly the size of request a 1,050,000-token window invites.

One detail that favours the self-hosted path and is easy to miss: the Apache-2.0 licence is what makes the model usable in a regulated environment at all. Fine-tuning on proprietary data, running air-gapped, and keeping the inference inside a data perimeter are three things Astra's API cannot offer at any price, and no benchmark row reflects them.

Where a router sits in a pairing like this

Intern-S2-397B is not on OrcaRouter and this page is not claiming it is; it is a new open checkpoint, and your route to it is the vendor's official Intern API or a deployment of your own. GPT-6 Astra is on our catalogue as openai/gpt-6-astra, at OpenAI's list rates passed through with 0% markup of ours.

The useful shape for a team running both is a split rather than a choice, and it is the configuration the routing layer is for: send the general, long-context, document-heavy traffic to the metered model behind one OpenAI-compatible endpoint, and keep the scientific and time-series batch work on the checkpoint you operate. Automatic failover then covers the case where either side is unhealthy — the self-hosted leg when a node drops, the API leg when a provider path degrades — without either team writing a second integration. That is a boundary drawn as a routing rule, not as an architecture.

Which one

Choose Intern-S2-397B if your input is scientific in the specific sense that the model was trained for — figure-dense literature pages, molecular and material structures, time-series signals — or if Apache-2.0 is the deciding property because your data cannot leave your infrastructure. Budget for the hardware before you budget for anything else, and run your own evaluation rather than trusting the card, because there is no independent number to fall back on.

Choose GPT-6 Astra if your work is agentic and long-horizon, if a 922,000-token input in a single request is a shape you actually produce, or if you need an independent scoreboard to defend the decision to someone else. Its $3.26 per completed index task is the highest figure in this comparison, and it buys a model that emits roughly 60M tokens doing that work where the median model emits 85M — which is why per-token price is the wrong number to decide on and per-finished-task cost is the right one. Do not choose it for short, high-volume traffic, and do not cross the 272,000-token line without re-running your cost model, because that line doubles the input rate on the whole request.

The two are not substitutes, and the reason is not benchmark points. One of them accepts a typed table of sensor readings and one accepts a scanned page of a paper, and neither does both well enough that the other becomes unnecessary. If your work somehow produces both, you already know which team in your organisation owns which half.

A release is also the moment to check what has gone wrong before: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.