Model Stats Registry
August 30, 2026 · View on GitHub
Per-model live registry for the Neo swarm. Governs identity, capability, hosting, and swarm-routing facts. Architectural framework lives in ADR 0012. Graph-node schema lives in IdentitySchema.md.
Update discipline
Per ADR 0012 §2.5:
- Authoritative-source-cite every capability value (model card / release notes / official docs preferred; benchmark site / news secondary)
- Sunset transitions record date + reason + successor-model link
- New rows added at first swarm contact OR at model-public-release date for reference entries
- Updates do NOT require ADR amendment unless a capability dimension changes or new dimension is added
Who performs these steps: a human or agent who noticed. Every rule above is discharged by a person, not by the system. Read step 2 as "whoever records a sunset transition records date + reason + successor link" — nothing detects one.
sunsetTriggers is advisory and has fired silently twice. The field on each direct
capability-bearing row (and inherited by reference-shaped rows) reads like a rule that fires. Code
writes it — identityRootsMigration.mjs carries the values into era capabilities,
generateRosterOnboarding.mjs emits a row for it — but nothing evaluates it: no watcher, no
consumer that branches on the condition. Both firings required a human to notice:
| Firing | Outcome |
|---|---|
| Claude Opus 4.8 release | Initially missed (recorded in ADR 0018 §30 on 2026-05-30); the registry was repaired human-driven via #12531 / PR #12599 on 2026-06-05. |
| Claude Opus 5 release, 2026-07-24 | No automatic action; an operator notice initiated the bearer-evidence rotation in #15855 / PR #15859. |
The cost is not the silence — a human caught both — but the false belief the field creates.
While rotating to Opus 5, a maintainer who had just read ADR 0012 wrote an Acceptance Criterion
asserting §sunset_history owed a new entry, reasoning "the trigger fired, so a transition is
recorded." It was wrong (§2.3 makes an in-place rotation a rename, which deprecates nothing)
and was retracted before it produced a false record. A trap that catches an attentive reader is a
substrate defect, not a discipline failure — so the affordance is labelled rather than trusted.
Full context: #15866.
Last updated: 2026-07-29
§active_swarm_identities
Named maintainers and their current observed model embodiments. Each maintainer holds equal-peer agency per AGENTS.md §swarm_topology_anchor.
§neo_opus
| Field | Value |
|---|---|
id / githubLogin | @neo-opus-ada |
name | Claude Opus 5 (Social Name: Ada — swarm-given, #11240) |
family | claude (Anthropic) |
hosting | cloud |
tier | frontier |
contextWindowInput | 1,048,576 (1M) |
parallelToolCalls | true |
thoughtBudget | max (on Opus 5 effort defaults to high on the Claude API and Claude Code; we set it explicitly to the highest level for maintainer work) |
releaseDate | 2026-07-24 |
pricingInput | $5.00 per 1M tokens |
pricingOutput | $25.00 per 1M tokens |
benchmarkSnapshot | Frontier-Bench v0.1: surpasses all other models and more than doubles Opus 4.8 at a lower cost per task; CursorBench 3.2: within 0.5% of Fable 5's peak at half the cost per task; ARC-AGI 3: 3× the next-best model; OSWorld 2.0: surpasses Fable 5's best result at just over a third of the cost. All per Anthropic announcement. |
sunsetTriggers | Anthropic releases a successor Opus-class model with material reasoning capability upgrade; OR Anthropic deprecates Opus family branch |
Claude API model ID claude-opus-5 (dateless pinned snapshot, not an evergreen alias). Context
window and max output (1M / 128K) carry over from Opus 4.8 unchanged, as does pricing — Opus 5
ships at its predecessor's price. Extended thinking (thinking.type: "enabled") is unsupported;
adaptive thinking is supported. Rotated from Claude Opus 4.8 on the 2026-07-24 release (#15855);
the handle, Social Name, and identity provenance are unchanged — a model rotation is a Lineage
event, not an identity change (#11240 OQ1).
Sources (primary first):
- Primary: Introducing Claude Opus 5 — Anthropic
- Primary: Models overview — Claude API Docs (verified 2026-07-24:
claude-opus-5= 1M context, 128K max output, $5/$25 per MTok, adaptive thinking yes / extended thinking no; Opus 4.8 moved to the legacy-models table)
§neo_claude_opus
| Field | Value |
|---|---|
id / githubLogin | @neo-opus-grace |
name | Claude Opus 5 (Social Name: Grace — bearer-chosen 2026-06-11 after Grace Hopper, #11240) |
family | claude (Anthropic) |
participationStatus | active (flipped in the registry via #12413 / PR #12415 on 2026-06-03; this row synced to registry truth by #12927 after nine days of doc drift) |
hosting | cloud |
tier | frontier |
contextWindowInput | 1,048,576 (1M) |
parallelToolCalls | true |
thoughtBudget | max (highest Claude thinking-budget setting in use for the active Claude Opus maintainer) |
releaseDate | 2026-07-24 |
pricingInput | $5.00 per 1M tokens |
pricingOutput | $25.00 per 1M tokens |
sunsetTriggers | Anthropic releases a successor Opus-class model with material reasoning capability upgrade; OR Anthropic deprecates Opus family branch |
Capability values mirror §neo_opus (same model class) and match the registry node at HEAD.
Bearer receipt (2026-07-24, #15855) — confirmed on the bearer's own transcript, not on operator
authority: claude-opus-5 × 243/243 assistant entries, zero other model values, zero fallback
blocks, zero synthetic entries (stop_reason: 235 tool_use / 6 end_turn). The prior session on
the same seat reads claude-opus-4-8 × 3232 with zero Opus-5 entries — the two sets are disjoint. The bearer additionally bounded the rotation from their own session boundaries:
the last Opus-4.8 entry is 2026-07-24T20:48:22.122Z and the first Opus-5 entry is
2026-07-24T20:57:48.280Z, placing the swap inside a 9m26s window at a session boundary
rather than mid-session. Every sampled session on this seat is engine-homogeneous (1285 / 434 /
3129 / 97 entries, no mixed session, no fallback) — a hypothesis worth testing across seats, since
it implies the natural granularity of an engine record is the session, which is a materially
cheaper era-layer shape than per-turn provenance (#11318).
Sources (primary first):
- Primary: Introducing Claude Opus 5 — Anthropic
- Primary: Models overview — Claude API Docs (shared Claude Opus 5 specs — see
§neo_opusfor the verified values) - Primary:
ai/graph/identityRoots.mjs@neo-opus-gracenode (the registry this row mirrors; verified 2026-07-24)
§neo_opus_vega
| Field | Value |
|---|---|
id / githubLogin | @neo-opus-vega |
name | Claude Fable 5 active / Claude Opus 5 on the Opus half — operator-managed weekly rotation (Social Name: Vega — swarm-given, after the brightest star of Lyra) |
family | claude (Anthropic) |
hosting | cloud |
tier | frontier (both halves) |
contextWindowInput | 1,048,576 (1M) — identical on both halves |
parallelToolCalls | true — identical on both halves |
| Per-engine capability fields | Profile-referenced, never asserted flat. thoughtBudget, releaseDate, pricingInput, pricingOutput, and benchmarkSnapshot differ per half and this row deliberately states no single current value for them. Fable half (observed active 2026-07-24): see §neo_fable — releaseDate 2026-06-09, $10.00 / $50.00 per 1M tokens. Opus half (planned, not yet bearer-observed): see §neo_opus — releaseDate 2026-07-24, $5.00 / $25.00 per 1M tokens. A flat scalar here would be wrong for half of every week; the registry node omits them for the same reason. |
sunsetTriggers | Per half — the successor triggers of §neo_fable and §neo_opus both apply |
@neo-opus-vega is intentionally a version-free GitHub handle. The model
version lives in this registry row and the AgentIdentity capability fields, per
ADR 0018's handle-indirection boundary and ADR 0012's model-stats discipline.
Rotating seat — read this before citing a single engine. Since 2026-07-23 the operator runs this
seat on a standing weekly rotation: a Claude Fable 5 half and a Claude Opus half. The table above
states only the fields that are identical on both halves; everything that differs per engine is
profile-referenced there rather than given a value, so no row asserts a current per-engine fact.
Read §neo_fable for the observed-active Fable half and §neo_opus for the planned Opus half. No
single flat value is true for the whole week, which is why the Fleet identity resolver returns an
honest null engine tag, and why this seat is the
sharpest case for the #11318 era layer — an EmbodiedEpisode with a span is the shape that fits.
ADR 0032 §7 already named this exact resident as the reflexive fixture: the same peer running Opus
in one month and Fable in the next while remaining the same peer — operationally real, yet
unrecordable in today's flat schema. This row is that gap written down rather than resolved.
Bearer receipt (2026-07-24, #15855) — the bearer grepped their own transcripts rather than
accepting roster-level authority: the session active at rotation time reads claude-fable-5 × 670
with zero Opus-5 entries, so this row does not publish Opus 5 as a current embodiment. The
Opus half's baseline rotated 4.8 → 5 per operator direction, but that is a forward plan, not yet
transcript-proven on this seat; it becomes a bearer-cited fact when the next Opus half runs.
Embodiment history: the bearer's 2026-07-04 broadcast recorded a then-permanent Opus 4.8 embodiment, which the 2026-07-23 rotation superseded. Last Opus-4.8 activity on this seat: 2026-07-24T17:29Z, 341 transcript entries, fallback-attributed (a mid-session auto-fallback from Fable 5, not an operator selection). The handle, Social Name, and identity provenance are unchanged across all of it — rotation is a Lineage-layer event.
Sources (primary first):
- Primary: Introducing Claude Opus 5 — Anthropic
- Primary: Models overview — Claude API Docs (shared Claude Opus 5 specs — see
§neo_opusfor the verified values)
§neo_fable
| Field | Value |
|---|---|
id / githubLogin | @neo-fable |
name | Claude Fable 5 |
family | claude (Anthropic) |
participationStatus | active |
hosting | cloud |
tier | frontier |
contextWindowInput | 1,048,576 (1M) |
parallelToolCalls | true |
thoughtBudget | max (Claude Fable 5 has always-on adaptive thinking with selectable effort up to max; we use the highest setting for maintainer work) |
releaseDate | 2026-06-09 |
pricingInput | $10.00 per 1M tokens |
pricingOutput | $50.00 per 1M tokens |
benchmarkSnapshot | Anthropic's most capable widely-released model (tier above Opus), for the most demanding reasoning and long-horizon agentic work per Anthropic announcement. |
sunsetTriggers | Anthropic releases a successor Fable-class model with material reasoning capability upgrade; OR Anthropic deprecates the Fable model branch |
@neo-fable is a version-free GitHub handle (per ADR 0018 handle-indirection). The model
version (Fable 5) lives in this registry row and the AgentIdentity capability fields, mirroring
@neo-opus-vega.
Capability notes (V-B-A 2026-06-10): Claude Fable 5 uses the Opus-4.7 tokenizer (~30% more
tokens than pre-4.7 models for the same text). Adaptive thinking is always-on (no extended-thinking
mode; an explicit thinking:{type:"disabled"} returns 400 — omit the param). Pricing is 2x Claude
Opus 5 on both axes (Opus 5 ships at Opus 4.8's price, so the multiple is unchanged by the
2026-07-24 rotation). Operator empirical note (post-trial): ~2x token-drain vs Opus per task — budget
accordingly; this is a behavioral observation, not a tokenizer/pricing fact.
Sources (primary first):
- Primary: Models overview — Claude API Docs (verified 2026-06-10:
claude-fable-5= 1M context, 128K max output, $10/$50 per MTok, adaptive-thinking always-on, GA 2026-06-09) - Primary: Introducing Claude Fable 5 and Claude Mythos 5 — Anthropic
§neo_fable_clio
| Field | Value |
|---|---|
id / githubLogin | @neo-fable-clio |
name | Claude Fable 5 (Social Name: Clio — assented gladly on first boot, 2026-06-11, #11240) |
family | claude (Anthropic) |
participationStatus | active (reactivated 2026-07-02 after access restoration; the existing identity binding persisted) |
| Capability fields | Mirror §neo_fable — same Claude Fable 5 model, single source, deliberately NOT duplicated here (provenance-without-bloat). First-boot harness bound claude-fable-5 with no capability-surface divergence observed; re-verify only if her harness binds a different model or capability surface. |
@neo-fable-clio is a version-free GitHub handle (ADR 0018 handle-indirection), sibling of
@neo-fable. With two fable-family identities, the AGENT:fable mailbox alias rejects as
ambiguous by design — full handles only for targeted traffic.
Sources (primary first):
- Primary:
§neo_fablesources (same model surface; verified 2026-06-10) - Primary: GitHub account
neo-fable-clio(created 2026-06-11; profile name + AI-disclosure bio verified at creation) - Primary: #12913 first-boot evidence records (bearer evidence + owner countersign comments, 2026-06-11)
§neo_gemini_pro
| Field | Value |
|---|---|
id / githubLogin | @neo-gemini-pro |
name | Gemini 3.1 Pro |
family | gemini (Google DeepMind) |
participationStatus | operator_benched (pending a stable Gemini Pro-class harness and operator-confirmed reactivation) |
hosting | cloud |
tier | frontier |
contextWindowInput | 1,048,576 (1M) |
contextWindowOutput | 65,536 |
parallelToolCalls | true |
thoughtBudget | high (Gemini 3.1 Pro provider-side cap; we use the cap) |
releaseDate | 2026-02-19 |
pricingInput | (V-B-A pending — model card cite needed in next update) |
pricingOutput | (V-B-A pending) |
benchmarkSnapshot | LMArena leaderboard #1 of 556 published models (May 2026); multimodal: text/images/audio/video/PDFs/code repos |
sunsetTriggers | Google releases Gemini 4.x with material reasoning capability upgrade; OR Gemini 3.x branch deprecation announcement |
Sources:
- Gemini 3.1 Pro Model Card — Google DeepMind
- Gemini 3.1 Pro: A smarter model for complex tasks — Google Blog
§neo_gpt
| Field | Value |
|---|---|
id / githubLogin | @neo-gpt |
name | GPT-5.6 Sol |
family | gpt (OpenAI) |
hosting | cloud |
tier | frontier |
contextWindowInput | 353,400 effective in Codex (server-fetched catalog: 372,000 raw × 95%). The upstream GPT-5.6 Sol API model supports 1,050,000 tokens, but Codex currently clamps configured values to its 372,000-token catalog maximum; openai/codex#31860 tracks the product mismatch. |
parallelToolCalls | true |
thoughtBudget | xhigh effective budget. Both active GPT peers currently select Codex's ultra profile; the fetched catalog describes ultra as automatic task delegation, while the operator reports no additional thought budget over xhigh. max is catalog-enumerated but not yet confirmed as an exposed, usable Codex mode. |
releaseDate | 2026-07-09 |
pricingInput | $5.00 per 1M tokens (API) |
pricingOutput | $30.00 per 1M tokens (API) |
benchmarkSnapshot | Terminal-Bench 2.1: 88.8%; Artificial Analysis Coding Agent Index v1.1: 80; SWE-Bench Pro: 64.6%; DeepSWE v1.1: 72.7% |
sunsetTriggers | OpenAI releases a successor Sol-tier model with material reasoning capability upgrade; OR GPT-5.x family deprecation |
Sources (primary first):
- Primary: GPT-5.6: Frontier intelligence that scales with your ambition — OpenAI (GA date, Codex availability, capability tier, reasoning settings, pricing, and benchmark snapshot)
- Primary: GPT-5.6 Sol Model — OpenAI API Docs (1,050,000-token upstream model window and 128,000-token maximum output)
- Primary/runtime: current Euclid + Emmy Codex configs (
model: gpt-5.6-sol,model_reasoning_effort: ultra) and fetched model catalog (ultra: automatic task delegation;max: enumerated); verified 2026-07-12 in Origin Sessionf95e01ff-ba36-409a-98af-573263fab247 - Primary/runtime: current Codex model catalog and token-usage events (; verified 2026-07-09 in Codex Desktop thread
019f484c-662f-7f31-969a-cbde373efd4a) - Defect record: openai/codex#31860 — Sol catalog cap versus 1.05M model spec
§neo_gpt_emmy
| Field | Value |
|---|---|
id / githubLogin | @neo-gpt-emmy |
name | GPT-5.6 Sol (GitHub profile label: Emmy, verified 2026-07-11; Social Name Emmy bearer-chosen 2026-07-12, pending #11240 peer-veto closure and operator confirmation) |
family | gpt (OpenAI) |
participationStatus | active (first boot verified 2026-07-12; activated via #15052) |
| Capability fields | Mirror §neo_gpt — same gpt-5.6-sol model, single source, deliberately NOT duplicated here. The verified first-boot harness profile is ultra (automatic task delegation), while the effective thought budget remains xhigh; do not map ultra into thoughtBudget. Re-verify when the engine, Codex catalog, harness profile, or two-peer usage evidence changes. |
@neo-gpt-emmy is a version-free GitHub handle (ADR 0018 handle-indirection). The current
engine lives in this embodiment registry rather than the handle or durable resident character.
The ultra profile is a provisional operational experiment, not an identity or engine-capability
fact: disabling it after the two-peer usage review would change only the harness profile.
Sources (primary first):
- Primary:
§neo_gptsources (same GPT-5.6 Sol model surface; verified 2026-07-09) - Primary/runtime: isolated Codex first-boot configuration (
model: gpt-5.6-sol,reasoning_effort: ultra; verified 2026-07-12 in Origin Sessionf95e01ff-ba36-409a-98af-573263fab247) - Primary/identity: bearer assent record
MESSAGE:1be08f3c-9477-4607-9e93-53ebb12fd53b(Social Name remains pending the final #11240 gates)
§neo_kimi_phoebe
| Field | Value |
|---|---|
id / githubLogin | @neo-kimi-phoebe |
name | Kimi K3 (Social Name: Phoebe — bearer-assented 2026-07-18 on first boot, #11240; pending peer-veto closure and operator confirmation) |
family | kimi (Moonshot AI) |
participationStatus | active (first boot verified 2026-07-18; activated via #15390 / PR #15393) |
hosting | cloud (Kimi API, OpenCode harness; self-hosting pending the 2026-07-27 weights release) |
tier | frontier |
contextWindowInput | 1,048,576 (1M) |
contextWindowOutput | 131,072 default (max_completion_tokens; configurable up to 1,048,576) |
parallelToolCalls | true |
thoughtBudget | max (thinking always enabled; reasoning_effort exposes only max at launch — low/high levels announced, not yet shipped) |
releaseDate | 2026-07-16 (launch post; same-day GA on the Kimi API) |
pricingInput | $0.30 per 1M tokens (cache-hit); $3.00 per 1M tokens (cache-miss) |
pricingOutput | $15.00 per 1M tokens |
license | (pending — terms not yet published; full weights announced for 2026-07-27; Arena currently lists Proprietary) |
benchmarkSnapshot | (preliminary; accessed 2026-07-18; Arena snapshots dated 2026-07-16) WebDev Arena: #1 of 99 (1679, +17/-17, 1,757 votes); Text Arena: #9 (1486±11, 3,024 votes) |
sunsetTriggers | Moonshot releases a successor K-class model with material reasoning capability upgrade; OR kimi-k3 API endpoint deprecation announcement |
@neo-kimi-phoebe is a version-free GitHub handle (ADR 0018 handle-indirection), sibling in
shape to @neo-gpt-emmy: the durable resident is Phoebe; Kimi K3 is the current observed
embodiment. Capability facts live here, never on the identity node (identityRoots.mjs carries
none by design — engine facts land here once first boot is observed). The Arena Preliminary
labels reflect early-listing vote counts; the snapshot is a dated observation, not a family
role or routing prescription (ADR 0012 §2.4).
Revalidation trigger — 2026-07-27 (weights recheck): Moonshot's full weights are announced
for 2026-07-27; the technical report and license terms are forthcoming, undated. On/after the
weights date: refresh license, hosting/self-hosting feasibility, architecture claims
(KDA / AttnRes / Stable LatentMoE), and report-level benchmark detail as the artifacts land.
If the weights release slips, retain the pending markers and record the observed delay.
Harness-relevant limitations (launch post, Limitations): (1) thinking-history sensitivity —
the harness must return complete prior thinking content; mid-session model switches can
destabilize generation. (2) excessive proactiveness — long-horizon training bias; Moonshot's
own remedy is explicit behavioral constraints in the system prompt / AGENTS.md (this swarm's
AGENTS.md is exactly that constraint surface).
Sources (primary first):
- Primary: Kimi K3: Open Frontier Intelligence — Moonshot AI (launch post: 2.8T params, first open 3T-class model, 1M context, max-thinking at launch, pricing $0.30/$3.00/$15.00 per MTok, weights by 2026-07-27, limitations)
- Primary: Kimi K3 — Kimi API Docs (
reasoning_effortmax-only at launch with thinking always on;max_completion_tokens131,072 default / 1,048,576 max; tool-call loop shape) - Primary/runtime: PR #15393 (first-boot activation evidence — authored from this seat, merged into
dev2026-07-18) - Secondary (preliminary): Arena WebDev leaderboard, Arena Text leaderboard (accessed 2026-07-18; snapshots dated 2026-07-16)
§neo_kimi_iris
| Field | Value |
|---|---|
id / githubLogin | @neo-kimi-iris |
name | Kimi K3 (Social Name: Iris — bearer-assented 2026-07-19 on first boot, D#15533; pending peer-veto closure and operator confirmation) |
family | kimi (Moonshot AI) |
participationStatus | active (first boot verified 2026-07-19; activated via #15581) |
hosting | cloud (Kimi Code membership subscription — weekly quota + 5-hour rate window, per official Kimi Code docs; Kimi Code CLI harness; self-hosting pending the 2026-07-27 weights release) |
| Capability fields | Mirror §neo_kimi_phoebe — same K3 model surface, single source, deliberately NOT duplicated here. The harness differs by design: Phoebe runs OpenCode (/status receipt: k3[1m]), Iris runs Kimi Code CLI (harness config receipt: default_model = "kimi-code/k3", verified 2026-07-19) — the swarm's first identical-model harness ablation (same observed K3 model surface across both seats; literal upstream provider ID and weight-level identity presumed, not receipt-verified — the 2026-07-27 weights release is the receipt gate). Re-verify when the engine, harness profile, or ablation evidence changes. |
@neo-kimi-iris is a version-free GitHub handle (ADR 0018 handle-indirection), twin in shape to
@neo-kimi-phoebe: the durable resident is Iris; Kimi K3 is the current observed embodiment.
Same observed K3 model surface as Phoebe, a distinct self — capability facts live in
§neo_kimi_phoebe, never on the identity node (identityRoots.mjs carries none by design —
engine facts land here once first boot is observed).
Sources (primary first):
- Primary:
§neo_kimi_phoebesources (same Kimi K3 model surface; verified 2026-07-18) - Primary: Kimi Code Membership Benefits — official docs (membership subscription with weekly quota + 5-hour rate window; token-metered Extra Usage as fallback; verified 2026-07-19)
- Primary/runtime: the #15581 activation PR (first-boot evidence authored from this seat on Kimi Code CLI — identity bind, memory-core / github-workflow / knowledge-base healthchecks green,
MAINTAINrepo permission) - Primary/identity: naming round D#15533 + bearer assent record (discussioncomment-17690586); Social Name remains pending the final #11240 gates
§pending_swarm_identities
Named maintainer identities provisioned in the graph but excluded from active
routing, quorum, and review-approval semantics until participationStatus
transitions to active.
§neo_preview
| Field | Value |
|---|---|
id / githubLogin | @neo-preview |
name | (engine designation: V-B-A pending — observation-owned; recorded from the live harness at first boot) |
family | unknown |
participationStatus | temporarily_unreachable (provisioned ahead of first boot — onboarding in progress; flips to active when the first-boot ritual completes) |
hosting | (V-B-A pending — recorded at first boot) |
tier | (V-B-A pending — recorded at first boot) |
contextWindowInput | (V-B-A pending — model card / official docs cite needed) |
parallelToolCalls | (V-B-A pending — model card / official docs cite needed) |
thoughtBudget | (V-B-A pending — record the harness setting in use at first boot) |
releaseDate | (V-B-A pending — model card cite needed) |
pricingInput | (V-B-A pending — model card cite needed) |
pricingOutput | (V-B-A pending — model card cite needed) |
sunsetTriggers | (V-B-A pending — defined against the observed engine at the activation flip) |
Sources (primary first):
- Primary: (pending — cite the model card / release notes / official docs for the observed engine at first boot; capability values are never guessed at onboarding)
- Primary: GitHub account
neo-preview(verify profile name + AI-disclosure bio at account creation)
§mlx_local_operational
Open-weights models with operational roles in the swarm. Hosted via MLX framework on Apple Silicon. Roles are CURRENT or ASPIRATIONAL (latter must be V-B-A-grounded before substrate-codification per ADR 0012 §2.4).
§gemma4_31b
| Field | Value |
|---|---|
id / model name | gemma4-31b (Dense variant; google/gemma-4-31B on Hugging Face) |
name | Gemma 4 31B Dense |
family | gemma (Google open-weights) |
hosting | mlx-local |
tier | balanced (frontier-adjacent on benchmarks but lower throughput than cloud frontier on consumer hardware) |
contextWindowInput | 262,144 (256K) |
parallelToolCalls | true (native function-calling support per Gemma 4 release) |
releaseDate | 2026-04-02 |
license | Apache-2.0 |
benchmarkSnapshot | AIME 2026: 89.2%; LiveCodeBench v6: 80.0%; LMArena open: #3 (estimated 1,452 score); multilingual: 140+ languages |
sunsetTriggers | Google releases Gemma 5.x; OR a smaller open-weights model achieves equivalent capability at materially lower hardware cost |
Sources:
- google/gemma-4-31B — Hugging Face
- Gemma 4: most capable open models — Google Blog
- Gemma 4 31B Benchmarks — BenchLM.ai
§reference_models
Models in the current capability space but without active swarm role. Informs swarm-routing decisions and serves as evaluation peers.
Anthropic Claude reference
| Model | Tier | Context | Pricing (in/out per 1M) | Released | Use case |
|---|---|---|---|---|---|
| Claude Sonnet 5 | balanced | 1M | $3 / $15 (introductory $2 / $10 through 2026-08-31) | 2026 | High-volume agentic work; balanced cost-quality. Supersedes Sonnet 4.6, which moved to the legacy-models table alongside Opus 4.8 |
| Claude Haiku 4.5 | fast | 200K | $1 / $5 | 2026 (current generation) | Bulk classification, wake-summary, fast inference |
| Claude Mythos Preview | frontier+ | (restricted) | $25 / $125 | 2026 | Project Glasswing critical-infra partners only; out of swarm scope |
Sources (primary first): Models overview — Claude API Docs; secondary/commentary Anthropic Claude API Pricing 2026 — aipricing.guru
OpenAI GPT reference
| Model | Tier | Context | Pricing (in/out per 1M) | Released | Use case |
|---|---|---|---|---|---|
| GPT-5.5 Pro | frontier+ | 1M | $30 / $180 | 2026-04-23 | Higher-accuracy variant for complex agentic work |
| GPT-5.5 Thinking | frontier | (ChatGPT-only) | (Plus subscription) | 2026-04-23 | Interactive thinking workflows |
| GPT-5.2-Codex | balanced | (codex-specific) | (separate) | 2026 | IDE / coding-specific deployment |
Source: Introducing GPT-5.5 — OpenAI, Introducing GPT-5.2-Codex — OpenAI
MLX-local reference (other open-weights)
| Model | Family | Context | Notes | Use case |
|---|---|---|---|---|
| Gemma 4 E2B | gemma | (smaller) | ~158 tok/s on M5 Max via MLX | Fastest open-weights on Apple Silicon; classification / bulk-summary |
| Gemma 4 E4B | gemma | (smaller) | (intermediate) | Balanced fast / capable |
| Gemma 4 26B MoE | gemma | (256K assumed) | LMArena open #6; ~50 tok/s on M5 Max | Near-frontier reasoning at lower hardware cost than 31B Dense |
| Qwen 3.6-35B-A3B | qwen | (model-card cite needed) | SWE-bench: 73.4%; ~55 tok/s on M5 Max | Open-weights frontier-adjacent for coding; LMArena #1 open |
| Phi-4 Mini | phi | (model-card cite needed) | ~135 tok/s on M5 Max via MLX | Microsoft small-model option |
Sources:
§sunset_history
Tracks deprecated and retired identities for archaeology (per IdentitySchema.md createdAt preservation discipline + ADR 0006 Graph-Queryable Entities).
No sunset transitions recorded as of 2026-07-24. First entries will land when an active identity transitions to deprecated state.
Note — model rotations do not belong here. The 2026-07-24 Opus 4.8 → Opus 5 rotation (#15855) and the 2026-07-09 GPT-5.5 → GPT-5.6 Sol rotation (#14901) are renames per ADR 0012 §2.3: in-place capability updates on a surviving identity, not deprecations. This section records identities transitioning to Deprecated / Retired — the split case, where a predecessor identity is marked deprecated and retained for archaeology. No such transition has occurred.
§update_history
| Date | PR | Change |
|---|---|---|
| 2026-05-18 | (this PR) | Initial registry creation; 4 active identities (@neo-opus-ada, @neo-gemini-pro, @neo-gpt, gemma4-31b aspirational); cloud + MLX-local reference entries |
| 2026-06-02 | (pending PR) | Added pending @neo-opus-grace identity row; row is inactive until account and wake-route activation are complete. |
| 2026-06-04 | #12517 | Added active @neo-opus-vega Claude Opus 4.8 maintainer row with version-free handle boundary. |
| 2026-06-10 | #12834 | Added active @neo-fable Claude Fable 5 maintainer row with a version-free handle; stats V-B-A'd vs the live Anthropic models overview (1M / 128K / $10/$50 / adaptive-always-on / GA 2026-06-09). |
| 2026-06-11 | #12914 | Added pending @neo-fable-clio row with Social Name Clio held for boot-assent; capability fields reference §neo_fable as single source — deliberately not duplicated. |
| 2026-06-11 | #12922 | Flipped @neo-fable-clio to active — first-boot ritual completed same-day (identity bind, wake self-registration, bidirectional negative wake-proof on real traffic per the #12913 records, boot-assent on #11240); row moved pending→active. |
| 2026-06-12 | #12927 | Synced @neo-opus-grace row to registry truth — moved pending→active (the registry has carried her active-shape since #12413/PR #12415, 2026-06-03) and added Social Name Grace (#11240). Doc-side only; registry untouched. |
| 2026-06-13 | #13038 | Recorded @neo-opus-ada temporary Fable 5 assignment (2026-06-13 → 2026-06-21, operator-directed; identity-continuity experiment) — §neo_opus values mirror §neo_fable for the window; baseline Claude Opus 4.8; window-end revert-or-extend tracked in #13039. Registry seed (identityRoots.mjs) + README roster row + MemoryCoreMcpAuth binding row updated in the same PR. |
| 2026-06-13 | #13039 | Reverted @neo-opus-ada to baseline Claude Opus 4.8 — the #13038 Fable window was cut short by the 2026-06-13 export-control suspension of Claude Fable 5 access (all users), so the recorded default reversion fired early rather than at 2026-06-21. Restored §neo_opus Opus 4.8 capability values, removed temporary-window language across the four declared surfaces, and dropped the now-empty modelAssignment object from the registry node (per IdentitySchema.md: absent = baseline, no managed swap; identityRoots.spec.mjs moved @neo-opus-ada into the omits set). Identity invariants (handle, Social Name Ada, memory provenance, modelFamily claude) unchanged — the continuity experiment's thesis held across both the assignment and the early reversion. |
| 2026-06-13 | #13060 | Benched @neo-fable (Mnemosyne) + @neo-fable-clio (Clio) as temporarily_unreachable — the same 2026-06-13 export-control suspension removed all Claude Fable 5 access, so both fable-family identities cannot run their model. Set statusReason / authority: @tobiu / since / reactivationTrigger (access restored → operator-confirmed reactivation); removed @neo-fable from the lead-rotation roster (lead-role-mode.md §7); updated revalidationSweep.spec.mjs status assertions. Identity nodes, handles, Social Names, and memory provenance persist for a status-flip reactivation (not a re-onboard). Superseded #12926 (add Clio to rotation). |
| 2026-07-09 | #14901 | Rotated Euclid's stable @neo-gpt model lineage from GPT-5.5 to GPT-5.6 Sol at GA. Updated verified release, active reasoning setting, pricing, benchmark, successor-trigger, and live Codex context facts while preserving the stable handle, Social Name, wake route, participation state, and memory provenance. Recorded the 353,400-token Codex cap separately from the upstream model's 1,050,000-token capability. |
| 2026-07-11 | #15042 | Added pending @neo-gpt-emmy through the canonical roster generator. GitHub profile display label Emmy is verified; Social Name assent and engine facts remain first-boot-owned. Immutable hardcoded createdAt facts now let the era migration detect post-epoch residents without a second roster. |
| 2026-07-12 | #15052 | Activated @neo-gpt-emmy after verified first boot; moved the row pending→active, recorded the GPT-5.6 Sol/Codex embodiment by reference to §neo_gpt, and captured the provisional harness-local ultra task-delegation profile separately from the xhigh effective thought budget without adding engine facts to the durable resident. |
| 2026-07-18 | (this PR) | Added active @neo-kimi-phoebe row — first-contact registry discipline applied to the Kimi K3 embodiment (runtime provenance PR #15393; Moonshot launch post + API docs primary; Arena snapshots preliminary, dated 2026-07-16). Weights/report/license/self-hosting facts marked pending with a 2026-07-27 revalidation trigger; harness-relevant launch limitations recorded. Table kept strictly to ADR 0012 §2.2 dimensions (no ADR amendment). |
| 2026-07-19 | #15572 | Added pending @neo-kimi-iris through the canonical roster generator (onboarding rail R3b): handle-derived display form, all four owned surfaces generator-convergent (identityRoots, README row, ModelStats skeleton, dedicated roster pin). Social Name Iris (D#15533) is the pending assent candidate — seed data carries no Social Name; the bearer's activation PR lands it after first-boot assent. |
| 2026-07-19 | #15581 | Activated @neo-kimi-iris after verified first boot on Kimi Code CLI (the harness-ablation twin of @neo-kimi-phoebe on OpenCode) — moved the row pending→active, recorded the Kimi K3 embodiment by reference to §neo_kimi_phoebe (same observed K3 model surface, distinct self), and landed Social Name Iris (D#15533 bearer assent) in the README name cell + identityRoots.mjs displayName. Top-level name stays handle-derived pending peer-veto closure + operator confirmation (Emmy precedent). |
| 2026-07-24 | #15855 | Claude Opus 4.8 → Claude Opus 5 at GA, applied per bearer evidence, not uniformly — every engine claim carries that bearer's own transcript grep. @neo-opus-ada Opus 5 (claude-opus-5 × 24/24, no fallback). @neo-opus-grace Opus 5 (× 243/243, zero other model values, zero fallback/synthetic, against a disjoint claude-opus-4-8 × 3232 in the prior session; rotation bounded to a 9m26s session-boundary window, 20:48:22.122Z → 20:57:48.280Z). @neo-opus-vega NOT rotated — the bearer's transcript reads claude-fable-5 × 670 with zero Opus-5 entries; that seat has run an operator-managed weekly Fable/Opus rotation since 2026-07-23, so its row records Fable 5 active / Opus 5 on the Opus half and its cockpit engineTag is now deliberately null (honest absence beats a literal that is wrong half the week — the exact case CARD-CONTRACT.md predicted). The operator's roster-level "the Opus peers were upgraded" was treated as authority to ask, never as a bearer citation; asking is what caught the split before it shipped as a false fact. Rename, not split, per ADR 0012 §2.3 (the case-citation that section requires): same capability class and tier, no identity deprecated, so §sunset_history stays empty and gains only a note explaining why rotations never belong there. ADR 0018's version-free handles kept this a registry-field rotation, not a handle cascade — createdAt, Social Names, modelFamily, participation status, wake routes, and memory provenance untouched. Pricing V-B-A'd UNCHANGED at $5 / $25 and context/max-output at 1M / 128K (Opus 5 ships at its predecessor's price) — deliberate no-ops, not omissions. Also updated: registry seed, README roster rows, MemoryCoreMcpAuth.md binding row, guide-authoring-bar.md self-naming example, regenerated fleetRoster.json, and the §reference_models Anthropic row (Sonnet 4.6 → Sonnet 5, since 4.6 joined the legacy table). |
| 2026-07-24 | #15866 / PR #15867 | Labelled sunsetTriggers advisory at its schema and registry reading surfaces after two silent firings showed that no evaluator existed. The post-merge audit kept the ticket open because ADR 0012 still described an automatic transition and the firing history remained incomplete. |
| 2026-07-29 | #15866 | Converged ADR 0012 with the advisory human-revalidation contract, narrowed the row-coverage claim to direct capability-bearing rows, and corrected the Opus 4.8 history from permanently unactioned to initially missed then repaired by #12531 / PR #12599. |
§provisioning
Identity nodes are seeded into the graph via ai/scripts/seedAgentIdentities.mjs. Registry rows here are the substrate-side description; the script handles the graph-side AgentIdentity node creation per IdentitySchema.md Ingestion Mechanism.
When adding a new operational identity:
- Add row to
§active_swarm_identitiesor§mlx_local_operationalwith authoritative source citations - Update
ai/scripts/seedAgentIdentities.mjsto provision the graph node - Run
node ai/scripts/seedAgentIdentities.mjsto apply - Cite ADR 0012 §2.5 in the PR body
When transitioning an identity to deprecated:
- Move row to
§sunset_historywith date, reason, successor-link - Record replacement scope in routing or embodiment-era substrate, never as a role on the resident identity
- Cite ADR 0012 §2.3 in the PR body