Site navigation

ModelsPreview snapshot

Bring the agents you already use. We pick the right model for each part.

HiveBase works with Claude, ChatGPT and Codex, and Grok. It sits above them, not instead of them: one Mission can run lanes of different agents, and an independent check grades the result.

Below: which routes are open today, and the evidence behind routing. A class publishes a pick only when the evidence we are allowed to show backs it. There is no “#1.”

THE LINEUP

Open today, and what is still opening.

Runs on the Claude and ChatGPT plans you already pay for, on your Mac. HiveBase Cloud is rolling out. Status below is read from the same route registry the product runs on.

Claude

Claude Code

Available · on your Mac

Builds lanes of a Mission with the Claude plan you already pay for, signed in by you on your own Mac.

Codex

ChatGPT and Codex

Preview · on your Mac

Builds lanes with the ChatGPT plan you already pay for. Claude and Codex lanes can run in one Mission.

Grok

Grok

Coming

Wired into the same route registry and still being certified, so it is not open to everyone yet. Grok Bot can already connect to HiveBase over MCP to read company context.

Connect Grok Bot over MCP →

  • Lanes of different agents. Claude and Codex lanes can run in one Mission.
  • An independent check. A role HiveBase plays, wearing its own mark, grades the work against the plan.
  • They can ask HiveBase. Codex and Claude Code can ask HiveBase for a company fact over MCP.
LAST REFRESHED
Aug 15, 2026, 12:00 AM UTC

Preview snapshot — not live production data. The live feed was unreachable for this render. The marketing site never fails a build on an empty or unreachable feed.

Classes
11
Verdicts published
0
Auto-reverts recorded
—
Feed
preview
WORK CLASSES

Eleven jobs. Placeholders, not verdicts.

Jump a class, or read them in order. Coding rows always name the harness. n is the first-party sample; intervals are shown when present. Classes without an evidence-backed verdict show methodology only — no model, n, or interval is claimed.

code.agentic

Agentic coding

Multi-file repository work: implement, recover, verify. Ranked as (harness, model, role).

n unpublished
Unpublished default · harness × model
No published default

No published HiveBase default in this preview snapshot.

Layout and methodology preview only. This snapshot does not name a preferred model. Live defaults appear when the public `current` publication exists.

Applicable: Multi-file repository work: implement, recover, verify. Ranked as (harness, model, role).

  • Research-based defaults stay useful before HiveBase outcome data accumulates.
  • Outcome statistics are observational HiveBase-owned dogfood, not a causal ranking.
  • The same model in a different harness is a different row.
  • Artificial Analysis numbers are tracked internally and never rehosted.
Research basis

Preview methodology only. Not a live research-based default.

  • Terminal-Bench

    Attributed agentic CLI source. Numbers stay on the source.

    Source
  • SWE-Rebench

    The public Hugging Face dataset is evaluation tasks, not a model-score feed.

    Source
  • AA Coding Index

    Tracked internally — see artificialanalysis.ai

    artificialanalysis.ai
code.frontend

Front-end / UI

Generated websites, components, and full-stack UI. Taste plus deployable quality.

n unpublished
Unpublished default · harness × model
No published default

No published HiveBase default in this preview snapshot.

Layout and methodology preview only. This snapshot does not name a preferred model. Live defaults appear when the public `current` publication exists.

Applicable: Generated websites, components, and full-stack UI. Taste plus deployable quality.

  • Research-based defaults stay useful before HiveBase outcome data accumulates.
  • Outcome statistics are observational HiveBase-owned dogfood, not a causal ranking.
  • The same model in a different harness is a different row.
  • Artificial Analysis numbers are tracked internally and never rehosted.
Research basis

Preview methodology only. Not a live research-based default.

  • Design Arena

    Blind human Elo on generated UI — attributed, not mirrored. Official API is optional at launch.

    Source
  • LMArena WebDev

    Human preference on web artifacts. See the dataset card.

    Source
code.review

Code review

Finding real defects without comment fatigue. Cross-family review is required.

n unpublished
Unpublished default · harness × model
No published default

No published HiveBase default in this preview snapshot.

Layout and methodology preview only. This snapshot does not name a preferred model. Live defaults appear when the public `current` publication exists.

Applicable: Finding real defects without comment fatigue. Cross-family review is required.

  • Research-based defaults stay useful before HiveBase outcome data accumulates.
  • Outcome statistics are observational HiveBase-owned dogfood, not a causal ranking.
  • The same model in a different harness is a different row.
  • Artificial Analysis numbers are tracked internally and never rehosted.
Research basis

Preview methodology only. Not a live research-based default.

  • Endor Agent Security League

    Attributed security-correctness source for harness × model interaction.

    Source
reasoning.deep

Deep reasoning

Strategic, second-order, long-horizon analysis.

n unpublished
Unpublished default · model
No published default

No published HiveBase default in this preview snapshot.

Layout and methodology preview only. This snapshot does not name a preferred model. Live defaults appear when the public `current` publication exists.

Applicable: Strategic, second-order, long-horizon analysis.

  • Research-based defaults stay useful before HiveBase outcome data accumulates.
  • Outcome statistics are observational HiveBase-owned dogfood, not a causal ranking.
  • Artificial Analysis numbers are tracked internally and never rehosted.
Research basis

Preview methodology only. Not a live research-based default.

  • Epoch Capabilities Index

    Versioned contextual evidence — attributed, not mirrored.

    Source
  • AA Intelligence Index

    Tracked internally — see artificialanalysis.ai

    artificialanalysis.ai
agentic.tools

Agentic tool use

Multi-step tool calling with schema fidelity and recovery.

n unpublished
Unpublished default · model
No published default

No published HiveBase default in this preview snapshot.

Layout and methodology preview only. This snapshot does not name a preferred model. Live defaults appear when the public `current` publication exists.

Applicable: Multi-step tool calling with schema fidelity and recovery.

  • Research-based defaults stay useful before HiveBase outcome data accumulates.
  • Outcome statistics are observational HiveBase-owned dogfood, not a causal ranking.
  • Artificial Analysis numbers are tracked internally and never rehosted.
Research basis

Preview methodology only. Not a live research-based default.

  • Berkeley Function Calling Leaderboard

    Attributed function-calling source.

    Source
  • LMArena Agent

    Official agent IPS schema, separate from text/vision/webdev ratings.

    Source
writing.narrative

Long-form writing

User-visible prose, brand voice, and durable narratives.

n unpublished
Unpublished default · model
No published default

No published HiveBase default in this preview snapshot.

Layout and methodology preview only. This snapshot does not name a preferred model. Live defaults appear when the public `current` publication exists.

Applicable: User-visible prose, brand voice, and durable narratives.

  • Research-based defaults stay useful before HiveBase outcome data accumulates.
  • Outcome statistics are observational HiveBase-owned dogfood, not a causal ranking.
  • Artificial Analysis numbers are tracked internally and never rehosted.
Research basis

Preview methodology only. Not a live research-based default.

  • LMArena Text

    Human preference on text. See the dataset card.

    Source
extraction.structured

Structured extraction

Schema-valid JSON / Output.object without silent field loss.

n unpublished
Unpublished default · model
No published default

No published HiveBase default in this preview snapshot.

Layout and methodology preview only. This snapshot does not name a preferred model. Live defaults appear when the public `current` publication exists.

Applicable: Schema-valid JSON / Output.object without silent field loss.

  • Research-based defaults stay useful before HiveBase outcome data accumulates.
  • Outcome statistics are observational HiveBase-owned dogfood, not a causal ranking.
  • Artificial Analysis numbers are tracked internally and never rehosted.
Research basis

Preview methodology only. Not a live research-based default.

longctx

Long context

Documents, transcripts, and large codebases. Effective window, not marketed size.

n unpublished
Unpublished default · model
No published default

No published HiveBase default in this preview snapshot.

Layout and methodology preview only. This snapshot does not name a preferred model. Live defaults appear when the public `current` publication exists.

Applicable: Documents, transcripts, and large codebases. Effective window, not marketed size.

  • Research-based defaults stay useful before HiveBase outcome data accumulates.
  • Outcome statistics are observational HiveBase-owned dogfood, not a causal ranking.
  • Artificial Analysis numbers are tracked internally and never rehosted.
Research basis

Preview methodology only. Not a live research-based default.

vision.doc

Document / vision

PDF, OCR, diagram, and image understanding.

n unpublished
Unpublished default · model
No published default

No published HiveBase default in this preview snapshot.

Layout and methodology preview only. This snapshot does not name a preferred model. Live defaults appear when the public `current` publication exists.

Applicable: PDF, OCR, diagram, and image understanding.

  • Research-based defaults stay useful before HiveBase outcome data accumulates.
  • Outcome statistics are observational HiveBase-owned dogfood, not a causal ranking.
  • Artificial Analysis numbers are tracked internally and never rehosted.
Research basis

Preview methodology only. Not a live research-based default.

  • LMArena Vision

    Human preference on vision tasks. See the dataset card.

    Source
speed.classify

Classification / gates

Binary and cheap structured gates. Recoverable if wrong.

n unpublished
Unpublished default · model
No published default

No published HiveBase default in this preview snapshot.

Layout and methodology preview only. This snapshot does not name a preferred model. Live defaults appear when the public `current` publication exists.

Applicable: Binary and cheap structured gates. Recoverable if wrong.

  • Research-based defaults stay useful before HiveBase outcome data accumulates.
  • Outcome statistics are observational HiveBase-owned dogfood, not a causal ranking.
  • Artificial Analysis numbers are tracked internally and never rehosted.
Research basis

Preview methodology only. Not a live research-based default.

No attributed evidence published for this class yet.

intel.x

X / community intel

Native X search and community recency. Discovery, not deciding.

n unpublished
Unpublished default · model
No published default

No published HiveBase default in this preview snapshot.

Layout and methodology preview only. This snapshot does not name a preferred model. Live defaults appear when the public `current` publication exists.

Applicable: Native X search and community recency. Discovery, not deciding.

  • Research-based defaults stay useful before HiveBase outcome data accumulates.
  • Outcome statistics are observational HiveBase-owned dogfood, not a causal ranking.
  • Artificial Analysis numbers are tracked internally and never rehosted.
Research basis

Preview methodology only. Not a live research-based default.

No attributed evidence published for this class yet.

METHODOLOGY

How a verdict is allowed to exist.

Work classes are the spine. External benchmarks, community consensus, and HiveBase first-party outcomes all hang off the same eleven jobs so they stay commensurable. Fusion never promotes a default from community recency alone — that channel is capped and decays.

Coding is harness-keyed. The ranked unit is (harness, model, role). An executor on Claude Code is not comparable to the same weights inside Cursor or Codex.

Outcomes stay separate. Research-based defaults can exist before HiveBase outcome data accumulates. Dogfood statistics appear in their own block with population, dates, denominator, and uncertainty. n < 20 is below the reporting threshold, not a causal override. Durable selection guidance lives in /docs/guides/model-intelligence. The same snapshot is at GET /api/models/ledger.

Artificial Analysis is not rehosted. AA indexes inform internal routing. This page will only say they are tracked and point at artificialanalysis.ai. Attributed sources (Terminal-Bench, SWE-Rebench, Design Arena, LMArena, BFCL, Epoch) appear as name + outbound link.

Saturated suites (SWE-bench Verified, HumanEval, retired open LLM leaderboards) are weight-zero. A canary that fails the quality floor auto-reverts; those events are public when the live feed includes them.

CHANGELOG

What moved — including auto-reverts.

Promotions that did not hold stay on the record. Auto-reverts are not hidden behind a successful later default. This preview's rows are layout examples, not recorded production events.

  1. Aug 15, 2026, 12:00 AM UTChivebase index run

    Preview snapshot for hivebase.ai/models layout. Not a live production index and not a preferred-model list.

  2. Aug 12, 2026, 09:10 AM UTCauto revert

    Example auto-revert event type the live changelog can show. This preview row is not a recorded production revert.

  3. Aug 10, 2026, 12:00 PM UTCsource stale

    SWE-bench Verified remains weight-zero (saturated / contaminated) and is not used for fusion.

SOURCE RECEIPT

Last refresh by source.

Ingest health for this snapshot. A skipped or empty refresh is shown, not implied successful. Artificial Analysis appears here as a tracked source, never as a copied score.

Last refresh time and status for each Model Intel source
SourceFinishedStatus
first_party—skipped
artificial_analysisTracked internally — artificialanalysis.ai—skipped
openrouter—skipped

Frequently asked questions

Why not one ranking?

A model that is right for cheap classification is the wrong unit for multi-file coding. HiveBase evaluates a verdict per work class and publishes it only when attributed evidence backs it — before that, a class shows methodology, not a pick. For code.agentic, code.frontend, and code.review the unit is (harness, model, role) — the same model on Claude Code versus Cursor is not the same system.

Why are Artificial Analysis scores missing?

Artificial Analysis is tracked internally for routing. Their terms do not allow us to rehost numeric scores. When AA is a source, the page says so and links to artificialanalysis.ai.

What does an auto-revert mean?

If a canary drops below HiveBase's quality floor, routing returns to the previous default. Public auto-reverts are listed in the changelog so a promotion that did not hold is visible, not buried.

Is this live production data?

When the public snapshot is reachable and carries an evidence-backed verdict, the receipt reads live. A reachable publication with no published verdicts renders classes as unpublished — methodology only, not a measured dataset. If environment, table, or query fail, this page renders a labeled preview snapshot so the marketing site never depends on production being up.

Are first-party numbers a ranking?

No. HiveBase-owned dogfood outcomes appear separately with population, dates, denominator, and uncertainty. They are observational. Research-based defaults remain useful before that evidence accumulates. Existing eval consent does not authorize publishing customer traces.

Routing you can inspect.

HiveBase routes work along these classes; a verdict joins them only as evidence earns it. The same plan, check, and receipt as everywhere else.