Stage

Recommended primary tool

Backup or specialist

Cost / API Constraints

Notes

1. Ideation / hypothesis generation

Gemini AI Co-Scientist, GPT-5 Pro

FutureHouse Owl (prior-art check)

Frontier consumer subscription ≈ USD 20–25 per user per month, premium reasoning tiers roughly an order of magnitude higher. AI Co-Scientist is not generally available, so its effective cost to an unaffiliated institution is access-limited rather than priced.

Co-Scientist debate agents particularly strong for novel hypotheses

2. Literature review

OpenScholar, FutureHouse Crow/Falcon

Elicit (structured tables), Semantic Scholar

OpenScholar open weights, free, self-hosting the 8B model needs ≈1×24 GB accelerator. FutureHouse free at web and API tier. Elicit freemium, paid ≈ USD 10–12 per user per month.

Citation accuracy matters more than speed

3. Citation management

Zotero + ResearchRabbit two-way sync

Scite.ai for citation context

Zotero and ResearchRabbit free, Zotero cloud storage metered above the free quota. Scite ≈ USD 12–20 per user per month or institutional licence. Crossref and PubMed APIs free.

Always verify against Crossref/DOI

4. Primer / amplicon design

Primer-BLAST + GPT-5/Gemini for criteria reasoning

Evo 2 for sequence-level checks

Primer-BLAST and Primer3Plus free (NCBI / public). LLM scaffolding cost as row 1. Evo 2 metered by GPU-hour in cloud, or capital cost on-premises.

LLM should augment, not replace, thermodynamic tools

5. CRISPR guide design

CRISPOR / CHOPCHOP + LLM scaffolding

Evo 2 / Nucleotide Transformer V2 (off-target)

CRISPOR and CHOPCHOP free web services. Nucleotide Transformer V2 2.5B runs on ≈1×40 GB accelerator, Evo 2 as row 4.

Off-target prediction remains the limiting step, treat model output as a shortlist for wet-lab validation

6. Bioinformatics analysis

GPT-5 / Gemini 2.5 Pro for code

DeepSeek V3.1 / Llama 4 on-premise for sensitive data

Frontier API metered per million tokens. Budget caveat: self-hosting flagship DeepSeek V3.1 is not a low-cost option it requires a multi-accelerator node even quantised. Llama 4 Scout and 8–32B-class models are the realistic sovereign choices at foundation scale.

Verify outputs against GATK, nf-core

7. Single-cell / multi-omics

scGPT / Geneformer (fine-tuned)

scVI/Harmony baselines

All open-source and free. The real cost is fine-tuning compute (≈1×24-48 GB accelerator) plus expert curation time for labelled data.

Avoid zero-shot foundation-model embeddings for novel discovery 26

8. Protein structure / interaction

AlphaFold 3 (Server, free non-commercial)

Boltz-1, Chai-1 (open-source clones), ESM-3

AlphaFold Server free for non-commercial use. Boltz-1 and Chai-1 free under permissive licences, ≈1×24–40 GB accelerator. ESM-3 largest variants metered via Forge API.

Choose the permissively licensed equivalents wherever commercialisation is foreseeable

9. Data interpretation/visualization

GPT-5 / Gemini 2.5 Pro

Claude (best plotting code style)

Covered by the row 1 subscription, no incremental cost.

Verify all numerical claims

10. Manuscript drafting

Claude Opus 4.5 (writing quality)

GPT-5, Gemini 2.5 Pro

Frontier consumer subscription ≈ USD 20–25 per user per month.

Disclose per ICMJE, never list AI as author

11. Statistical analysis

GPT-5 with Python tool

DeepSeek R1 for cost-sensitive

Covered by the row 1 subscription where code execution is included. DeepSeek R1 open weights free, but subject to the same multi-accelerator hosting constraint as row 6.

Rerun in R/Python with auditable scripts

12. Figure generation

BioRender, matplotlib via GPT-5

DALL-E 3 / Imagen 3 for schematic

matplotlib free. BioRender is the significant recurring cost: individual academic plans run to the low hundreds of USD per year and institutional licences higher and publication rights require a paid plan, as the free tier does not license figures for journal publication.

Disclose any AI-generated images

13. Peer-review preparation

Claude Opus 4.5 for rebuttal tone

GPT-5 for technical depth

Covered by the row 10 subscription.

Do not upload confidential comments to public models (ICMJE 2024)

14. Journal selection

Elsevier Journal Finder, Springer Nature Journal Suggester, SciSpace match

Manual scope review

Publisher finders and JANE free, SciSpace freemium.

Never rely on a single publisher’s suggestion as a scope assessment

15. Plagiarism / AI-detection

iThenticate (canonical)

Originality.ai (AI-text screening)

iThenticate is normally reached through an institutional Similarity Check membership, standalone per-manuscript pricing is material for an institution submitting at volume. Originality.ai is credit-based.

Disclose AI use upfront — best defence