Skip to main content
initializing...
    Ivy Biosciences logo
    Ivy BiosciencesEvergreen Health
    docspricing

    the honest side-by-side

    choose the route that fits the question.

    not every question needs the same system. compare what each option runs, what it records, and how you reach it — and where something else is the better call.

    one structural difference up front: you are not locked to one vendor's stack — anthropic direct or through openrouter, across the model families your plan includes — or bring your own key once you sign in.

    start a runjump to the matrix

    11 options, same rows

    what each option runs, and what it can show you afterwards.

    .bloom* against schrödinger, sandboxaq, openeye / cadence, benchsci (emet), benchling ai, amazon bio discovery, tamarind bio, claude science (anthropic), phylo / biomni lab and tooluniverse (harvard). cells reflect each vendor's own public materials as read in october 2026 (benchsci: august 2026); “not advertised” means their public materials don't list it — not proof of absence. contact us if your team would like a deeper side-by-side.
    .bloom* compared with ten alternatives
    platformevidence — what it runswhat feeds whatrecordaccessself-host
    .bloom*500+ biomedical tools and 50+ databases, composed by ivy biosciencesruns the models — folding and docking on GPUs, admet and retrosynthesis on CPU — hands parts of a question to sub-agents, and results name the tool and job that produced themIvy, inside one question — each job's output feeds the next as a job referencesigned-in runs from the web app, the terminal, curl or MCP leave a tamper-evident record you can check, and so does a guest chat, from the browser tab it is open in, while the guest session lasts. edit a row and the chain breaks at the verifierfree tier + paid plans + enterprise self-hostenterprise, until the planned source release — needs a GPU compute provider, Clerk and S3
    schrödingercomputational platform for molecular discovery and designruns physics simulation you drive from its GUI (maestro) + python apiworkflows you build in maestro or its python apinot advertisedvia sales — no public price listinstalls on your own linux, mac or windows machines and clusters; also deploys to cloud
    sandboxaqlarge quantitative models for physics-grounded drug discovery (aqbiosim)runs its quantitative models on its own compute; integration with chat assistants is listed as coming soonnot advertisednot advertisedenterprise licensing, mcp integration or co-development — no public price listadvertises keeping your models and data inside your walls
    openeye / cadenceai- and physics-powered molecular designruns docking, shape search and simulation in orion or its toolkits, driven from its UI or pythonfloes you build in orion, from its ui or pythonnot advertisedcommercial licence; free academic research licencetoolkits install locally with a licence file; orion runs in its cloud
    benchsci (emet)agentic evidence workbench over a curated knowledge graphagentic skills that reason over evidence already in its graph; no folding, docking or simulationno compute to chain — its skills reason over evidence already in its graphclaims trace to literature, not to a runenterprise contractno — saas only
    benchling aiai built into the benchling notebook and registryruns hosted models on GPUs; agents register molecules, write notebook entries and process results; models chain as steps in larger workflowsmodels run as steps in larger workflows, beside your eln and registry dataadvertises audit trails for ai actions and gxp / 21 cfr part 11 compliancebenchling subscription; free access for academic scientistsnot advertised — benchling's cloud
    amazon bio discoveryagentic app for lab-in-the-loop antibody designagents pick and run hosted antibody models in a no-code app, then send top designs to partner labs for synthesis and assaysits agents chain hosted antibody models, then hand designs to partner labsadvertises experiment history and lineagefree academic tier + paid monthly plans + enterprise; no existing aws account needednot advertised — runs in aws, no infrastructure to deploy
    tamarind biomanaged compute for 300+ ai molecular design toolsruns published models on its GPUs at batch scale; pipelines feed one tool's output into the next, from the web app, api, cli, mcp or its agentspipelines you build feed one tool's output into the next; its design agent runs binder campaignsnot advertisedfree tier + paid plans + enterprise; agents can provision a trial key without signing upnot advertised — hosted on its own cloud
    claude science (anthropic)research app (beta) with life-sciences connectors and skillsclaude writes and runs analysis code — including folding and docking models such as boltz-2 and diffdock — on your machines, your cluster or your own modal accountcode claude writes and runs on your own machines, cluster or modal accountadvertises provenance per result: the code, environment and conversation behind itpro, max, team and enterprise plans; compute billed by the provider you connectthe app and compute run on machines you choose; model calls go to anthropic
    phylo / biomni labagentic biology workspace (biomni lab)agents run analyses across its integrated tools and dataagents run analyses across its integrated tools and datakeeps the code, sources and execution details behind each analysisfree tier + paid plans + enterpriseenterprise: customer-hosted VPC; the public biomni agent code self-hosts
    tooluniverse (harvard)open toolkit for building ai scientist agents from any llmyour own agent calls its tools over mcp, a cli or python; most tools query a data provider, and local ml tools run on your machineyour own agent calls one tool after another over mcp, a cli or pythonnot advertisedfree; you install and run ityes — runs locally
    • .bloom*
      500+ biomedical tools and 50+ databases, composed by ivy biosciences
      evidence — what it runs
      runs the models — folding and docking on GPUs, admet and retrosynthesis on CPU — hands parts of a question to sub-agents, and results name the tool and job that produced them
      what feeds what
      Ivy, inside one question — each job's output feeds the next as a job reference
      record
      signed-in runs from the web app, the terminal, curl or MCP leave a tamper-evident record you can check, and so does a guest chat, from the browser tab it is open in, while the guest session lasts. edit a row and the chain breaks at the verifier
      access
      free tier + paid plans + enterprise self-host
      self-host
      enterprise, until the planned source release — needs a GPU compute provider, Clerk and S3
    • schrödinger
      computational platform for molecular discovery and design
      evidence — what it runs
      runs physics simulation you drive from its GUI (maestro) + python api
      what feeds what
      workflows you build in maestro or its python api
      record
      not advertised
      access
      via sales — no public price list
      self-host
      installs on your own linux, mac or windows machines and clusters; also deploys to cloud
    • sandboxaq
      large quantitative models for physics-grounded drug discovery (aqbiosim)
      evidence — what it runs
      runs its quantitative models on its own compute; integration with chat assistants is listed as coming soon
      what feeds what
      not advertised
      record
      not advertised
      access
      enterprise licensing, mcp integration or co-development — no public price list
      self-host
      advertises keeping your models and data inside your walls
    • openeye / cadence
      ai- and physics-powered molecular design
      evidence — what it runs
      runs docking, shape search and simulation in orion or its toolkits, driven from its UI or python
      what feeds what
      floes you build in orion, from its ui or python
      record
      not advertised
      access
      commercial licence; free academic research licence
      self-host
      toolkits install locally with a licence file; orion runs in its cloud
    • benchsci (emet)
      agentic evidence workbench over a curated knowledge graph
      evidence — what it runs
      agentic skills that reason over evidence already in its graph; no folding, docking or simulation
      what feeds what
      no compute to chain — its skills reason over evidence already in its graph
      record
      claims trace to literature, not to a run
      access
      enterprise contract
      self-host
      no — saas only
    • benchling ai
      ai built into the benchling notebook and registry
      evidence — what it runs
      runs hosted models on GPUs; agents register molecules, write notebook entries and process results; models chain as steps in larger workflows
      what feeds what
      models run as steps in larger workflows, beside your eln and registry data
      record
      advertises audit trails for ai actions and gxp / 21 cfr part 11 compliance
      access
      benchling subscription; free access for academic scientists
      self-host
      not advertised — benchling's cloud
    • amazon bio discovery
      agentic app for lab-in-the-loop antibody design
      evidence — what it runs
      agents pick and run hosted antibody models in a no-code app, then send top designs to partner labs for synthesis and assays
      what feeds what
      its agents chain hosted antibody models, then hand designs to partner labs
      record
      advertises experiment history and lineage
      access
      free academic tier + paid monthly plans + enterprise; no existing aws account needed
      self-host
      not advertised — runs in aws, no infrastructure to deploy
    • tamarind bio
      managed compute for 300+ ai molecular design tools
      evidence — what it runs
      runs published models on its GPUs at batch scale; pipelines feed one tool's output into the next, from the web app, api, cli, mcp or its agents
      what feeds what
      pipelines you build feed one tool's output into the next; its design agent runs binder campaigns
      record
      not advertised
      access
      free tier + paid plans + enterprise; agents can provision a trial key without signing up
      self-host
      not advertised — hosted on its own cloud
    • claude science (anthropic)
      research app (beta) with life-sciences connectors and skills
      evidence — what it runs
      claude writes and runs analysis code — including folding and docking models such as boltz-2 and diffdock — on your machines, your cluster or your own modal account
      what feeds what
      code claude writes and runs on your own machines, cluster or modal account
      record
      advertises provenance per result: the code, environment and conversation behind it
      access
      pro, max, team and enterprise plans; compute billed by the provider you connect
      self-host
      the app and compute run on machines you choose; model calls go to anthropic
    • phylo / biomni lab
      agentic biology workspace (biomni lab)
      evidence — what it runs
      agents run analyses across its integrated tools and data
      what feeds what
      agents run analyses across its integrated tools and data
      record
      keeps the code, sources and execution details behind each analysis
      access
      free tier + paid plans + enterprise
      self-host
      enterprise: customer-hosted VPC; the public biomni agent code self-hosts
    • tooluniverse (harvard)
      open toolkit for building ai scientist agents from any llm
      evidence — what it runs
      your own agent calls its tools over mcp, a cli or python; most tools query a data provider, and local ml tools run on your machine
      what feeds what
      your own agent calls one tool after another over mcp, a cli or python
      record
      not advertised
      access
      free; you install and run it
      self-host
      yes — runs locally

    where it loses

    reach for something else when the question is narrower.

    .bloom* is for the loop between these tools: a question answered with computed, cited, recorded evidence in one run.
    • schrödinger · openeye

      physics-grade free-energy work on a clinical candidate.

      you'll be done sooner there. .bloom* is for the loop before it: triage a series with computed, cited evidence, then hand the survivors to the physics.

    • emet (benchsci)

      reasoning over curated literature and reagent evidence, including a closed-access corpus you cannot crawl.

      its claims trace to literature; it can't run a fold or a dock, so it can't make evidence that isn't in the graph yet.

    • claude science

      your data and compute already live on your laptop, cluster or modal account, and you want an agent that writes and runs the analysis code there.

      it writes the code and you run it on your own compute; admet and retrosynthesis aren't advertised. .bloom* chains the jobs on hosted compute and, signed in, keeps a tamper-evident record.

    • benchling ai

      your data already lives in benchling and you need gxp / 21 cfr part 11 records around it.

      its models sit beside your eln; docking and retrosynthesis aren't advertised. .bloom* chains fold → dock → admet → retro in one question, and exports to benchling when you're done.

    • amazon bio discovery

      you're optimising antibodies and want the designs synthesised and assayed by partner labs.

      it's antibody-only and no-code; .bloom* docks, scores admet and plans the synthesis for small molecules, from the browser, the terminal, curl or MCP.

    • phylo / biomni lab

      the question is omics analysis, literature or experiment design, worked through by a team in one shared environment.

      it keeps the code and execution details behind each analysis; docking, admet and retrosynthesis aren't advertised. .bloom* chains them in one question and, signed in, keeps a tamper-evident record.

    • tooluniverse

      you want a free, local catalogue of database and api lookups to hand your own agent.

      your agent does the chaining and the local models run on your machine; .bloom* runs the gpu models for you, chains the jobs and, signed in, records each step tamper-evidently.

    what only .bloom* does in this set

    the run is the product: computed, kept, checkable.

    each claim below is checked in our own source, and against the rows above.
    • tamper-evident

      signed-in runs leave a tamper-evident record you can check.

      edit a row and the chain breaks at the verifier. it detects modification; it does not prevent it, and it is not access control.

      see how the record works
    • chains the jobs

      folding and docking on GPUs, then admet and retrosynthesis — Ivy chains them in one question, each job's output feeding the next, on hosted compute with no cloud account of your own.

    • won't fabricate

      a runtime gate refuses to dock or predict on a SMILES no tool returned in the chat; citations are grounded and deduped.

    • one agent, every surface

      the web app, curl to /api/v1/orchestrate, the ivybloom terminal app and any MCP client all reach Ivy; a single tool job can also be dispatched directly with curl to /api/v1/jobs, ivybloom run, or a studio project.

      see every way in
    • shareable by link

      token links with no-auth guest viewers for chats, files, jobs, and documents — hand a reviewer the result, no account.

    • real-time

      ~20Hz streaming and live job + card updates.

    • FSL-licensed

      FSL-1.1-Apache-2.0 licensed; source release planned — the whole platform, to read and run yourself once it ships. phylo's public biomni agent code and tooluniverse (apache-2.0) are readable today.

    compute and the bench

    computation narrows the set. the bench still decides.

    computation doesn't replace wet-lab experiments — it makes them smarter: narrow the compound set before it reaches the bench, prioritize high-confidence candidates, focus wet-lab resources on validation, and aim to shorten design-make-test cycles.
    computational methods compared with their wet-lab analogs
    applicationcomputational methodwet-lab analogwhere accuracy holds
    compound screeningcomputational virtual screening — hours to dayshigh-throughput screening (HTS) — weeks to monthsgood correlation with experimental binding
    ADMET predictionML property predictors — seconds to minutesin vivo PK/PD studies — weeks per studystrong performance on well-represented chemotypes
    structure determinationAI structure prediction — seconds to minutesX-ray crystallography / cryo-EM — months to yearsnear-experimental backbones for many well-studied targets; weaker on novel folds
    lead optimizationgenerative molecular design — hours to days for designstraditional medicinal chemistry — weeks per design-make-test cycledependent on scoring function quality
    binding affinityML-based scoring — milliseconds per predictionSPR / ITC / biochemical assays — days per compoundmoderate correlation with experimental values
    toxicity assessmentAI toxicity prediction — seconds per predictionin vivo toxicology studies — weeks to months per studygood sensitivity for known mechanisms
    crystal structure analysisautomated XRD analysis — minutes to hoursmanual XRD analysis — days to weeksstrong performance for known phases
    • compound screening
      computational method
      computational virtual screening — hours to days
      wet-lab analog
      high-throughput screening (HTS) — weeks to months
      where accuracy holds
      good correlation with experimental binding
    • ADMET prediction
      computational method
      ML property predictors — seconds to minutes
      wet-lab analog
      in vivo PK/PD studies — weeks per study
      where accuracy holds
      strong performance on well-represented chemotypes
    • structure determination
      computational method
      AI structure prediction — seconds to minutes
      wet-lab analog
      X-ray crystallography / cryo-EM — months to years
      where accuracy holds
      near-experimental backbones for many well-studied targets; weaker on novel folds
    • lead optimization
      computational method
      generative molecular design — hours to days for designs
      wet-lab analog
      traditional medicinal chemistry — weeks per design-make-test cycle
      where accuracy holds
      dependent on scoring function quality
    • binding affinity
      computational method
      ML-based scoring — milliseconds per prediction
      wet-lab analog
      SPR / ITC / biochemical assays — days per compound
      where accuracy holds
      moderate correlation with experimental values
    • toxicity assessment
      computational method
      AI toxicity prediction — seconds per prediction
      wet-lab analog
      in vivo toxicology studies — weeks to months per study
      where accuracy holds
      good sensitivity for known mechanisms
    • crystal structure analysis
      computational method
      automated XRD analysis — minutes to hours
      wet-lab analog
      manual XRD analysis — days to weeks
      where accuracy holds
      strong performance for known phases

    wet-lab validation remains the gold standard for regulatory approval and clinical translation. predictions should guide experimental design, not replace it.

    benchmarking methodology

    how the claims on this page are made.

    a DiffDock validator runs against a PDBbind subset, and each tool names the published model it runs. the figures elsewhere on this page are qualitative or illustrative, not measured results.

    questions about methodology: [email protected]

    compare plans by domainsee what it computes

    see exactly which capabilities unlock at each tier, organized by research domain.

    Ivy Biosciences logo
    Ivy BiosciencesEvergreen Health
    𝕏in
    © 2026 ivy biosciences•privacy•terms•security•support
    Loading...