A generated title card reading 'LingSante-3.1-flash vs Gemini 3.8 Flash' with the subtitle 'One reads the note. The other reads the scan.' Three cards read 'LingSante: text input only', 'Gemini: image, video, audio, PDF' and 'Gemini $0.75 / $3.75 per 1M', with a footer 'Ant figures from its own card; Gemini figures labelled third-party.' The OrcaRouter logo is composited along the bottom edge.
Guides & Insights

LingSante-3.1-flash vs Gemini 3.8 Flash: The Scan It Cannot Read

Auteur

Elias Hawthorne

Publicatiedatum

Nieuwste modellen · 20Bekijk alle modellen →
Benchmarks: Artificial Analysis · dagelijks bijgewerkt
Terug naar alle berichten

A discharge summary arrives as a photograph of a printed page. A lab result arrives as a PDF with a table rendered as an image. A referral arrives as a scan of a fax. In each case, a model that cannot see the input has to be preceded by an OCR pipeline that someone has to build, monitor and pay for. That is the first fact of the comparison between LingSante-3.1-flash and Gemini 3.8 Flash: the medical specialist is text-only, and the general-purpose flash model reads images, video, audio and PDFs natively.

The second fact is that this matchup does not appear anywhere in Ant's evaluation. The LingSante card compares against Gemini 4 Argon, the flagship, not Gemini 3.8 Flash. Google's launch material does not mention LingSante at all. So everything below is assembled from two separate vendor records and one independent leaderboard, and the seams are labelled.

A comparison Ant did not run

The closest row in Ant's table is Gemini 4 Argon, a larger and more expensive model. On HealthBench Professional, LingSante-3.1-flash scores 65.72 and Gemini 4 Argon scores 55.91. That is not evidence about Gemini 3.8 Flash, but it is a useful ceiling: if the specialist beats the bigger Gemini on the medical rubric, it is likely to beat the smaller one too. Likely is the operative word.

What the independent record adds complicates that inference. Third-party HealthBench Professional boards place Gemini 3.8 Flash at roughly 51 to 52 — below Gemini 4 Argon, as expected, and well below LingSante's vendor figure. Artificial Analysis, meanwhile, gives Gemini 3.8 Flash an Intelligence Index of 30.8 at medium reasoning on the v5.0.1 scale, while LingSante-3.1-flash has no Index at all. So the closed model has a general-capability score and a third-party medical score; the open specialist has a vendor medical score and one 23.3% result on a harder independent medical test. Neither side has the full picture.

Vision changes the input, not the answer

It is tempting to treat multimodality as a feature checkbox. In a clinical pipeline it changes the architecture of the product.

• With Gemini 3.8 Flash, a scanned form, a photograph of a rash, a PDF discharge summary and an audio clip of a consultation all become prompt content. The model reads them directly at $0.75 per million input tokens.

• With LingSante-3.1-flash, each of those has to become text first. That means an OCR or speech-to-text stage, its own error rate, its own latency, and its own per-page cost. The specialist never sees the original artefact, only a transcription of it.

For typed clinical notes, order sets, structured records and text-based evidence retrieval, that difference is small. For the messy front door of a hospital — faxes, scans, photographs, voicemail — it is the whole problem.

There is a counter-argument, and it is legitimate: OCR errors are visible and auditable, whereas a multimodal model's misreading of an image is harder to catch. A pipeline that transcribes first can log the transcription. A pipeline that lets the model look can only log the answer. Neither design is obviously safer; they fail differently.

Screenshot of the OrcaRouter model page for google/gemini-3.8-flash captured 23 November 2026 in English: header tags Vision, Audio, Tools, JSON and Reasoning with a 2026-09-02 date; a specification panel showing 1M-token context, 65K max output, text+image+video+file+audio input, text output, p50 TTFT 8.98 s and Zero Data Retention; price tiles reading $0.75 input and $3.75 output per 1M tokens; and a traffic figure of 134.6M tokens over seven days.

Two different kinds of unverified

The evidence quality on each side deserves separate treatment, because "vendor-reported" and "independently benchmarked" do not mean the same thing here.

LingSante-3.1-flash: 41 rows, all run by Ant Group, with footnotes naming the harness for each family — Claude Code v2.1.205 for SWE-Bench Pro, Mini-SWE-Agent v2.4.6 for Terminal-Bench 4.0, a defined temperature and top-p. That is more methodological disclosure than most launch cards provide. It is still one organisation grading its own model. The only outside data point is Artificial Analysis's MLCR-AA score of 23.3%, and that page shows no Intelligence Index.

Gemini 3.8 Flash: Google's launch numbers are also vendor-run, but the model has been in the market since 2 September 2026, which means third parties have had time to run it. Artificial Analysis has an Index of 30.8. HealthBench leaderboards carry it in the low fifties. Terminal-Bench 2.1 is reported around 89.4, DeepSWE around 73.7, HLE-Verified around 54.9. These are still mostly launch-day figures, but they come from more than one organisation and they agree with each other.

For a procurement decision, that asymmetry is worth more than the 13-point gap in the vendor medical score. A model with an independent Index and cross-checked third-party rows is easier to put in front of a clinical governance committee than a model with one vendor table and one 23.3% outside result.

Cost per thousand documents

The token prices look far apart and the total cost is closer than it looks.

• Gemini 3.8 Flash — $0.75 input and $3.75 output per million tokens through 31 December 2026, then $1.50/$7.50 from 1 January 2027. Cached input is $0.075. Output includes thinking tokens.

• LingSante-3.1-flash — no first-party list price. Self-hosting needs eight H200s in FP8 or four H20-3e cards in MXFP4; the vendor's launch partner sells API access with a $20-per-month token plan on the free tier.

At low volume, Gemini is cheaper because there is no capital cost. At high volume, the specialist's fixed cost amortises. The break-even is not a token-price calculation; it is a utilisation calculation. A hospital processing a few thousand documents a month will not amortise four H20-3e cards. One processing millions will.

The routing answer is more practical than either extreme. Gemini 3.8 Flash is on OrcaRouter at $0.75/$3.75 with no markup, so a team can send image-bearing requests there and text-only medical requests to a self-hosted specialist, using one endpoint and an automatic failover rule. That is a routing problem, not a vendor-selection problem, and it is the design most clinical pipelines will end up with.

A generated two-column scoreboard titled 'LingSante-3.1-flash vs Gemini 3.8 Flash — the scoreboard'. Left column values read 'Input: text only', 'HealthBench Pro: 65.72 vendor', 'AA Index: not listed', 'Context: 262K shipped', 'Price: self-host only', 'Medical safety: MedSCE 83.11'. Right column values read 'Input: text, image, video, audio, PDF', 'HealthBench Pro: about 52 third-party', 'AA Index: 30.8', 'Context: 1M tokens', 'Price: $0.75 / $3.75', 'Medical safety: none published'. A footer reads 'Gemini Index per Artificial Analysis; HealthBench rows labelled by source.'

What the benchmarks cannot settle

Three questions decide this matchup and none of them is answered by either card.

The first is transcription loss. If your OCR stage loses a drug dose or a negation, the specialist's better clinical judgement never gets a chance to matter. Nobody has published an end-to-end comparison of LingSante-3.1-flash plus OCR against Gemini 3.8 Flash reading the same scans. Until someone does, the multimodal model has an unmeasured advantage on exactly the inputs where medicine is hardest.

The second is output length. Gemini 3.8 Flash is a reasoning model with thinking levels; Artificial Analysis measured roughly 48,000 output tokens per task and about $0.58 per task at high effort. LingSante's card does not publish a tokens-per-task figure, and its launch post claims a 20% reduction in search steps versus accuracy-only multi-agent reinforcement learning, which suggests its designers were thinking about exactly this cost. A per-task cost comparison would be more useful than a per-token one, and neither vendor provides it.

The third is safety behaviour in deployment. LingSante scores 83.11 on Ant's MedSCE safety benchmark; Gemini 3.8 Flash has no comparable public medical-safety row. A model that is better at medical reasoning but untested on refusal behaviour is not obviously the safer choice for a patient-facing product.

The verdict depends on your input

If the input is text and the task is clinical judgement — triage, evidence synthesis, structured record review, patient communication — LingSante-3.1-flash is the more specialised tool, and its vendor numbers on HealthBench Professional, EHR-Complex and PatientAgentBench are strong enough to justify a pilot. Accept that you are piloting an unverified model and budget for your own evaluation.

If the input is anything a camera or a PDF produced, Gemini 3.8 Flash is the only one of the two that can do the job without a separate pipeline. It is more expensive per token, it has no medical tuning, and its medical scores are lower — but it can read the document, and a medical model that cannot read the document is not in the same competition.

The honest framing is not "which model is better." It is "which input do you actually receive." For most health systems, the answer is both, which is why the interesting engineering is the router between them rather than the choice between them.

Screenshot of the Hugging Face model card for inclusionAI/LingSante-3.1-flash captured 23 November 2026 in English: 27 likes and a blank downloads-last-month figure; tags Text Generation, Safetensors, bailing_hybrid, conversational, custom_code and fp8 with License mit; sidebar showing 572B parameters in BF16, F32 and F8_E4M3, the line 'This model isn't deployed by any Inference Provider', and a six-item LingSante collection; introduction stating 560 billion total parameters, 25 billion per token and a 1-million-token context window; the research-and-evaluation-only notice; and a grid of benchmark chart thumbnails whose legend names LingSante-3.1-flash, DeepSeek-V4.1-Flash-1017, GLM-5.3, Kimi-K3.1-Pro, Claude Opus 5.5, GPT-6.2 Astra and Gemini 4 Argon.