Vox Model Selection — 2026-Q2 Refresh
Vox Model Selection — 2026-Q2 Refresh
Section titled “Vox Model Selection — 2026-Q2 Refresh”This document captures the rationale for the 2026-05-15 catalog refresh in contracts/orchestration/model-catalog.bootstrap.v1.json, the premium-alias updates in contracts/orchestration/model-routing.v1.yaml, and the panel-pin resolutions in contracts/eval/llm-panel.v1.yaml.
TL;DR. The Vox model pipeline is architecturally sound — 5-tier routing, 15 strength tags, 13 task categories, scoreboard quality feedback, observed-cost dynamic pricing all already work. The catalog content was stale: the codegen/debugging/security premium aliases pointed at a preview-tier model at $125/MTok, and Anthropic’s Opus 4.7 (GA 2026-04-16), Google’s Gemini 3 family, OpenAI’s GPT-5.4/5.5 family, and DeepSeek’s V4 family were all entirely missing. This refresh adds 15 frontier-tier model rows and rotates 5 premium aliases.
§1 What we changed
Section titled Ҥ1 What we changedӤ1.1 Premium alias rotation (model-routing.v1.yaml:71-79)
Section titled “§1.1 Premium alias rotation (model-routing.v1.yaml:71-79)”| Alias | Before (2026-04) | After (2026-05-15) | Why |
|---|---|---|---|
codegen | claude-mythos-preview-20260407 | anthropic/claude-opus-4.7 | Mythos is preview at $125/MTok; Opus 4.7 GA is 87.6% SWE-bench at $25/MTok |
debugging | claude-mythos-preview-20260407 | anthropic/claude-opus-4.7 | Same |
security | claude-mythos-preview-20260407 | anthropic/claude-opus-4.7 | Same; Anthropic safety stack matches security review |
research | google/gemini-2.5-pro-preview | google/gemini-3.1-pro | 3.1 Pro is current GA; leads ARC-AGI-2 + GPQA 94.3% |
planning | google/gemini-2.5-pro-preview | openai/gpt-5.5-pro | GPT-5.5 Pro leads BenchLM agentic at 90.1 — best multi-step + tool-call reliability |
review | anthropic/claude-sonnet-4.6 | (unchanged) | Sonnet 4.6 is the price/quality knee at $3/$15 |
logic | deepseek/deepseek-r1 | (unchanged) | R1 is free on OpenRouter (rate-limited); good cost floor |
visus | qwen/qwen-3.5-vl | (unchanged) | Vision-capable, cheap |
Mythos Preview removed from catalog entirely (2026-05-15, post-council). At $125/MTok with preview stability and #3 on BFCL, retaining it as an opt-in entry created a footgun — users who pinned to it via user-config would burn budget on a non-GA line. The hardcoded-values audit at .vox/audit/2026-05-11-hardcoded-values/ will refresh on the next vox ci hardcoded-values-audit run and the Mythos entries will drop out.
§1.2 New catalog entries (model-catalog.bootstrap.v1.json)
Section titled “§1.2 New catalog entries (model-catalog.bootstrap.v1.json)”16 new model rows added (15 frontier-tier + the new Mythos-preview annotation):
| Model | Tier | Cost in/out (per MTok) | Use case |
|---|---|---|---|
anthropic/claude-opus-4.7 | Elite | $5 / $25 | Primary codegen/debugging/security alias; 1M ctx; extended thinking with adaptive difficulty |
anthropic/claude-haiku-4.5 | Light | $1 / $5 | Cheap tier retaining computer-use + extended thinking + Anthropic safety stack |
openai/gpt-5.5-pro | Elite | $8 / $30 | Primary planning alias; BenchLM agentic 90.1 |
openai/gpt-5.4 | Pro | $5 / $20 | Terminal-Bench leader 75.1%; gpt-frontier panel member |
openai/gpt-5-mini | Light | $3 / $10 | Generalist cheap tier |
openai/gpt-5-nano | Light | $0.05 / $1 | Classifier-grade — routing decisions, intent classification |
google/gemini-3.1-pro | Elite | $3.50 / $15 | Primary research alias; ARC-AGI-2 + GPQA leader |
google/gemini-3-flash | Light | $0.50 / $3 | Multimodal generalist |
google/gemini-3.1-flash-lite | Light | $0.25 / $1.50 | Cheapest 1M-context multimodal — audio/video/image input, 363 tok/s output |
deepseek/deepseek-v4-pro | Pro | $0.50 / $2 | MIT, 1M ctx, 1.6T/49B active, LiveBench Coding 69.99 |
deepseek/deepseek-v4-flash | Light | $0.07 / $0.28 | Cheapest competitive output cost at MIT licensing |
moonshot/kimi-k2.6-thinking | Pro | $0.80 / $3 | LiveBench Coding leader 78.57 — open-source alternative |
zhipu/glm-5.1 | Pro | $0.50 / $2 | Single-call BFCL format-precision leader |
qwen/qwen-3.6-27b | Pro (local) | $0 | 77.2% SWE-bench, Apache 2.0 — recommended MENS base-model migration target |
qwen/qwen3-coder-next-32b | Local | $0 | <30ms code-suggestion latency in VS Code — recommended llama3:latest replacement |
§1.3 Panel-pin resolutions (contracts/eval/llm-panel.v1.yaml)
Section titled “§1.3 Panel-pin resolutions (contracts/eval/llm-panel.v1.yaml)”The CR-L0..L4 measurement panel had three placeholder version strings since the panel was drafted on 2026-05-15. This refresh resolves all three:
| Member ID | Placeholder | Resolved |
|---|---|---|
mens-current | "matches workspace.package.version" policy | (unchanged; resolves at runtime against MENS checkpoint hash) |
claude-sonnet | "claude-sonnet-4-7-20260801" (didn’t exist) | claude-sonnet-4-6 (current GA, released 2026-02-17) |
gpt-frontier | "gpt-5-2026-mm-dd" (placeholder) | gpt-5.4 (current GA for code); CR-L0 substitute pinned at gpt-5.5-pro for agentic measurement |
The Sonnet pin uses the un-dated version slug because Anthropic API resolves it to the latest 4.6 patch; the dated form claude-sonnet-4-6-YYYYMMDD is available when bit-for-bit reproducibility is required.
§2 May 2026 model landscape (the data this refresh is based on)
Section titled Ҥ2 May 2026 model landscape (the data this refresh is based on)Ӥ2.1 Frontier coding leaderboard (SWE-bench Verified)
Section titled “§2.1 Frontier coding leaderboard (SWE-bench Verified)”Six frontier models cluster within 1.3 points of each other:
| Rank | Model | SWE-bench Verified | Cost in/out per MTok | Notes |
|---|---|---|---|---|
| 1 | Claude Opus 4.7 | 87.6% | $5 / $25 | 1M ctx; extended thinking; released 2026-04-16; +13% coding uplift over 4.6 at same price |
| 2 | Claude Opus 4.6 | 80.8% | $5 / $25 | Superseded by 4.7 |
| 3 | Gemini 3.1 Pro | 80.6% | $3.50 / $15 | Leads ARC-AGI-2 + GPQA |
| 4 | Claude Sonnet 4.6 | 79.6% | $3 / $15 | Best price/quality coding ratio |
| 5 | Qwen 3.6 27B | 77.2% | open | Apache 2.0 |
| 6 | GPT-5.4 | 57.7% SWE-bench Pro | $5 / $20 | But 75.1% Terminal-Bench leader |
Per ianlpaterson.com’s 38-task routing-table benchmark and lmcouncil.ai aggregate: “Six frontier models now score within 0.8 points of each other on SWE-bench Verified, with all six frontier models within 1.3% of each other. Opus 4.7 still leads on reasoning depth and long-context coherence, but you are paying 2.5x more than Gemini 3.1 Pro for 0.2 more points on SWE-bench.”
§2.2 Agentic / tool-call (BFCL v4, BenchLM agentic)
Section titled “§2.2 Agentic / tool-call (BFCL v4, BenchLM agentic)”| Rank | Model | BenchLM agentic | BFCL v4 tool-call |
|---|---|---|---|
| 1 | GPT-5.5 Pro | 90.1 | 40.4 |
| 2 | Claude Opus 4.7 (with extended thinking) | high | ~40 |
| 3 | Llama 3.1 405B Instruct | n/a | 40.5 (raw format) |
| 4 | Claude Mythos Preview | n/a | 39.6 |
For multi-turn tool use, Claude Sonnet 4.5/4.6 is preferred. For single-call format precision, GLM 4.5/5.1.
§2.3 Reasoning-specialty
Section titled “§2.3 Reasoning-specialty”Four useful shapes of reasoning model emerged in 2026:
- General high-reasoning: o3, o4-mini
- Long-context dense work: Claude Opus 4.7 extended thinking; Sonnet 4.6 thinking
- Parallel exploration: Gemini 2.5 Pro Deep Think; Gemini Flash Thinking
- Open-weights cost-efficient: DeepSeek R1 (free), DeepSeek V4 Pro reasoning mode
Opus 4.7’s adaptive thinking is the most interesting innovation: a lightweight pre-pass classifies input difficulty, then the model spends correspondingly more compute thinking before answering. This is uniquely suited to Vox’s mixed workload (some prompts are trivial syntax fixes; some are cross-file architectural debugging) — the cost scales with task, not with prompt length.
§2.4 Open-source frontier
Section titled “§2.4 Open-source frontier”DeepSeek V4 Pro (released 2026-04-24): 1.6T total / 49B active, 1M context, MIT licensing. LiveBench Coding 69.99, Agentic Coding 56.67 (May 12, 2026 snapshot). The MIT license + 1M context + competitive coding score makes V4 Pro the obvious cloud-fallback for Anthropic outages.
DeepSeek V4 Flash: 284B / 13B active, MIT, $0.28 output cost. The cheapest competitive output rate of any 2026 model — ~18× cheaper than Haiku output, ~4.5× cheaper than GPT-5 Nano output.
Kimi K2.6 Thinking: LiveBench Coding 78.57 — currently the open-source coding leader (May 2026 snapshot). Available on OpenRouter.
Qwen 3.6 27B: 77.2% SWE-bench Verified, Apache 2.0 license. Fits on a MacBook with 64GB RAM. The licensing matters for MENS-derived weights.
Llama 4 Scout: 10M-token context window (the long-context champion), but coding lags Qwen 3.6 by ~5 points on SWE-bench. Llama 4 Maverick is the dense variant.
§2.5 Small / cheap
Section titled “§2.5 Small / cheap”| Model | Cost in/out per MTok | Notable |
|---|---|---|
| Claude Haiku 4.5 | $1 / $5 | Retains computer-use + extended thinking + vision + Anthropic safety stack |
| Gemini 3 Flash | $0.50 / $3 | Multimodal |
| Gemini 3.1 Flash-Lite | $0.25 / $1.50 | 1M ctx, multimodal (audio/video/image), 363 tok/s output |
| GPT-5 Mini | $3 / $10 | Medium-cheap |
| GPT-5 Nano | $0.05 / $1 | Cheapest commercial; classifier-grade |
| DeepSeek V4 Flash | $0.07 / $0.28 | Cheapest competitive output cost overall |
Cost-tier savings: A typical enterprise workload that routes by task complexity (simple → Haiku/Nano, medium → Sonnet/GPT-5, complex → Opus/GPT-5-Pro) saves ~58% compared to using Opus everywhere.
§2.6 Local / Ollama-friendly
Section titled “§2.6 Local / Ollama-friendly”| Model | Hardware | Speed | Notable |
|---|---|---|---|
| Qwen3 8B | Laptop | 20-35 tok/s | OK for batch, slow for IDE autocomplete |
| Qwen3-Coder-Next 32B | Decent GPU | <30ms code suggestions in VS Code | IDE-grade latency |
| GPT-OSS 20B | 16GB laptops | 40-60+ tok/s | |
| Llama 3.3 70B | Mac Studio M4 Max | 30+ tok/s | |
| Qwen 3.5 (122B / 10B active) | 64GB MacBook | competitive | Beats GPT-5-mini on most benchmarks |
§2.7 Free / rate-limited (OpenRouter)
Section titled “§2.7 Free / rate-limited (OpenRouter)”DeepSeek R1, Llama 3.3 70B, Gemma 3 — all available at zero cost on OpenRouter with rate limits ~20 req/min, 200 req/day.
§3 Task-to-model mapping (the deliverable)
Section titled “§3 Task-to-model mapping (the deliverable)”Each row maps a Vox task category from model-routing.v1.yaml:51-64 to a recommended cascade. Bold = primary (matches the premium_alias); italic = preferred Pro-tier fallback; plain = cost-floor fallback.
§3.1 Per task category
Section titled “§3.1 Per task category”| Vox task | Primary | Pro fallback | Light/cost-floor |
|---|---|---|---|
| CodeGen (production) | anthropic/claude-opus-4.7 | anthropic/claude-sonnet-4.6 | deepseek/deepseek-v4-pro |
| CodeGen (light) | anthropic/claude-haiku-4.5 | google/gemini-3-flash | deepseek/deepseek-v4-flash |
| Testing | anthropic/claude-sonnet-4.6 | openai/gpt-5.4 | qwen/qwen-3.6-27b |
| Debugging (cross-file) | anthropic/claude-opus-4.7 w/ extended thinking | openai/gpt-5.4 (Terminal-Bench) | deepseek/deepseek-v4-pro |
| TypeChecking | anthropic/claude-sonnet-4.6 | openai/gpt-5.4 | qwen/qwen-3.6-27b |
| Research | google/gemini-3.1-pro | anthropic/claude-opus-4.7 | google/gemini-2.5-pro-preview (legacy) |
| Parsing | openai/gpt-5-nano | google/gemini-3.1-flash-lite | deepseek/deepseek-v4-flash |
| Review | anthropic/claude-sonnet-4.6 | anthropic/claude-opus-4.7 (high-stakes) | qwen/qwen-3.6-27b |
| General | anthropic/claude-sonnet-4.6 | openai/gpt-5.4 | deepseek/deepseek-v4-pro |
| Ars | anthropic/claude-sonnet-4.6 | openai/gpt-5-mini | deepseek/deepseek-v4-flash |
| Planning | openai/gpt-5.5-pro | anthropic/claude-opus-4.7 | google/gemini-3.1-pro |
| InterAgent (MCP, A2A) | openai/gpt-5.5-pro | anthropic/claude-sonnet-4.6 | meta-llama/llama-3.1-405b-instruct |
| ToolOrchestration | anthropic/claude-sonnet-4.6 | zhipu/glm-5.1 (format precision) | openai/gpt-5-mini |
| Visus | qwen/qwen-3.5-vl | google/gemini-3.1-flash-lite (multimodal) | anthropic/claude-haiku-4.5 |
§3.2 vox repair specifically
Section titled “§3.2 vox repair specifically”The repair loop (crates/vox-cli/src/commands/repair.rs:134) is the single hottest LLM consumer in Vox because it re-sends the entire source file across 3 retry attempts. The cost optimization here matters disproportionately.
Recommendation: anthropic/claude-sonnet-4.6 with prompt caching enabled. Three reasons:
- Prompt caching cuts effective cost by 80%+ across the 3-attempt loop. Anthropic’s prompt caching reduces cached input from $3 → $0.30/MTok. The first attempt pays full; attempts 2 and 3 read from cache. Total cost over a 3-attempt session converges to ~1.27× single-attempt cost vs. 3.00× without caching.
- 79.6% SWE-bench Verified at $3/$15 is the price/quality knee. Opus 4.7’s 87.6% costs 5× more for marginal improvement on repair-shaped tasks.
- 1M context means even large project files don’t truncate mid-repair.
For unusually-stuck or multi-file repair sessions, escalate to Opus 4.7 via the existing complexity-based tier bumping.
§3.3 CR-L0 spec-to-app agent loop
Section titled “§3.3 CR-L0 spec-to-app agent loop”CR-L0 measures end-to-end agent loop quality with a $5/spec cost ceiling. Per the panel-rule, gpt-5.5-pro is the primary (BenchLM agentic 90.1) with claude-opus-4.7 as alternative. Both have 1M-class context windows; both have native tool support; both honor the cost ceiling at typical spec sizes (~50K input, ~10K output ≈ $1.20-$1.50 per spec).
§3.4 MENS positioning
Section titled “§3.4 MENS positioning”MENS is currently QLoRA-fine-tuned on Llama 3.2 / Llama-4 Scout base. Recommendation: migrate MENS base-model to Qwen 3.6 27B. Three reasons:
- +5 points SWE-bench: Qwen 3.6 (77.2%) beats Llama 4 Scout (~72%) on coding tasks.
- Apache 2.0 licensing: Llama 4 weights are under Meta’s bespoke license; Apache 2.0 makes MENS-derived weights freely shippable.
- 128K context: still ample for Vox’s typical inputs; the 10M context of Llama 4 Scout is not currently exercised by Vox training corpora.
This is a v0.6 retraining cycle — not autonomous, requires council ratification + a training run.
Alternative for IDE-grade local inference: retrain on Qwen3-Coder-Next 32B (<30ms VS Code latency).
§4 Cost analysis
Section titled Ҥ4 Cost analysisӤ4.1 Per-task estimated cost (typical 4K input, 500 output)
Section titled “§4.1 Per-task estimated cost (typical 4K input, 500 output)”| Task type | Primary model | Cost per call | Cost per 100 calls |
|---|---|---|---|
| CodeGen (Elite) | Opus 4.7 | $0.0325 | $3.25 |
| CodeGen (Pro) | Sonnet 4.6 | $0.0195 | $1.95 |
| CodeGen (Light) | Haiku 4.5 | $0.0065 | $0.65 |
| Research | Gemini 3.1 Pro | $0.022 | $2.20 |
| Planning | GPT-5.5 Pro | $0.047 | $4.70 |
| Review | Sonnet 4.6 | $0.0195 | $1.95 |
| Parsing | GPT-5 Nano | $0.00072 | $0.072 |
| Visus | Qwen-3.5-VL | $0.00045 | $0.045 |
§4.2 vox repair 3-attempt session cost (10K source file)
Section titled “§4.2 vox repair 3-attempt session cost (10K source file)”| Configuration | Cost per session |
|---|---|
| Sonnet 4.6 with prompt caching (recommended) | ~$0.064 |
| Sonnet 4.6 without prompt caching | ~$0.155 |
| Opus 4.7 with prompt caching | ~$0.108 |
| Opus 4.7 without prompt caching | ~$0.260 |
| DeepSeek V4 Pro | ~$0.026 |
Even with prompt caching, DeepSeek V4 Pro is ~2.5× cheaper than Sonnet 4.6 — strong fallback for cost-sensitive deployments. The Anthropic safety stack and ecosystem are worth the premium for production work; the price/quality tradeoff is real.
§4.3 CR-L1 reference-panel measurement run (164 problems × 5 attempts × 3 panel members)
Section titled “§4.3 CR-L1 reference-panel measurement run (164 problems × 5 attempts × 3 panel members)”| Panel | Cost per run |
|---|---|
| MENS-current (local) + Sonnet 4.6 + GPT-5.4 | ~$92 |
| MENS-current + Opus 4.7 + GPT-5.5 Pro (max quality) | ~$215 |
| MENS-current + Sonnet 4.6 + GPT-5 Mini (cost-floor) | ~$48 |
| MENS-current + DeepSeek V4 Pro + Gemini 3.1 Pro | ~$32 |
The recommended panel (Sonnet 4.6 + GPT-5.4 + MENS) is the price/signal sweet spot. Per the panel SSOT’s ”<$300/RC” cap, this stays well within budget.
§5 Migration plan
Section titled Ҥ5 Migration planӤ5.1 Landed 2026-05-15 (this refresh)
Section titled “§5.1 Landed 2026-05-15 (this refresh)”- ✅
contracts/orchestration/model-routing.v1.yaml— 5 premium aliases rotated - ✅
contracts/orchestration/model-catalog.bootstrap.v1.json— 15 new model rows + Mythos-preview annotation - ✅
contracts/eval/llm-panel.v1.yaml—claude-sonnetandgpt-frontierplaceholders resolved - ✅ This document — design rationale + benchmark citations
§5.2 v0.6 (next-minor)
Section titled “§5.2 v0.6 (next-minor)”- A. Enable prompt caching for
vox repair— addcache_control: { type: "ephemeral" }to the system+source-code blocks incrates/vox-cli/src/commands/repair.rs:142-148. Estimated savings: ~80% per session. - B. Update
openrouter_chat_model_preference()(crates/vox-config/src/inference.rs) default toanthropic/claude-sonnet-4.6if not set; let users override. - C. Retire
llama3:latestas Local-tier default in favor ofqwen/qwen3-coder-next-32b.
§5.3 v0.7+ (council-gated)
Section titled “§5.3 v0.7+ (council-gated)”- D. MENS base-model migration from Llama-4 Scout → Qwen 3.6 27B. Requires retraining run + corpus validation. Tracked via
mens-training-ssot.md. - E.
OpenAiDirectprovider plumbing — currently OpenAI is only accessible via OpenRouter proxy. Adding direct access reduces latency by ~30-50ms per call and is policy-required for some compliance regimes. New provider entry +OpenAiApiKeysecret per existing pattern. - F. Model-spec
stabilityaxis — distinguish GA / preview / experimental within tiers. Tier semantics today conflate “expensive” with “stable”; adding astabilityfield would let the router prefer GA at the same tier when available.
§5.4 Quarterly refresh cadence
Section titled “§5.4 Quarterly refresh cadence”Per CR-L1’s panel rebaselining policy, this document is the canonical place to update model recommendations. The schedule:
| Cadence | Action |
|---|---|
| Per release-candidate tag | Pin panel version_pinned to current model IDs; re-validate cost ceilings |
| Per quarter | Re-audit benchmarks; rotate premium aliases if a new release shifts ≥3 points on the relevant benchmark |
| Per major model launch (Anthropic / OpenAI / Google) | Same-day catalog entry + 7-day evaluation window before alias rotation |
§6 Open questions
Section titled “§6 Open questions”These are surfaced by the audit and deferred to council:
Should Mythos preview stay in the catalog?Resolved 2026-05-15: removed. Any external config that pinned toclaude-mythos-preview-20260407will get an “unknown model” routing fallback; the router selects from remaining catalog entries. Follow-on: addvox ci retired-model-warnlint that warns when users pin to retired/preview model IDs in user-config.- MENS base-model migration timing: Llama-4 → Qwen 3.6 is a v0.6/v0.7 retraining commitment. Does the council ratify or defer?
OpenAiDirectprovider: latency win is real but the attack-surface delta needs threat modeling. Defer to security review.- Model-spec
stability: ga | preview | experimentalfield: schema change toModelCatalog(incrates/vox-orchestrator/src/models/spec.rs). Backward-compatible to add asOption<String>; emergent if Mythos-preview pattern recurs. - Free-tier exploitation (DeepSeek R1, Llama 3.3 70B, Gemma 3 on OpenRouter): rate-limited to 20/min, 200/day. Useful for development / CI experiments but not production. Should the router automatically prefer free-tier when rate-limit budget permits? Currently no preference is encoded.
- Prompt-caching wire-up: requires Anthropic-specific request-shape changes. The existing OpenAI-compat proxy at OpenRouter may strip cache_control fields. Need to verify per-provider passthrough.
§7 Appendix — Benchmarks cited
Section titled “§7 Appendix — Benchmarks cited”| Benchmark | What it measures | Where I drew from |
|---|---|---|
| SWE-bench Verified | End-to-end PR-to-fix on real GitHub issues | iternal.ai, smartscope.blog, lmcouncil.ai |
| SWE-bench Pro | Harder, multi-step variants | smartscope.blog |
| Terminal-Bench | Tool-shell-execution accuracy | smartscope.blog |
| LiveBench Coding | Coding + Agentic Coding (May 12, 2026 snapshot) | akitaonrails.com |
| BFCL v4 | Berkeley Function-Calling Leaderboard | gorilla.cs.berkeley.edu |
| BenchLM agentic | Multi-step task completion + tool-call reliability | benchlm.ai |
| ARC-AGI-2 | Abstract reasoning | sureprompts.com |
| GPQA | Graduate-level science Q&A | smartscope.blog |
| MMLU-Pro | Multi-task knowledge | findmyaitool.com |
| Arena Elo | Human preference rating | various |
| Anthropic API pricing | Direct provider rates (May 2026) | finout.io, aipricing.guru |
| OpenRouter pass-through pricing | Provider-rate proxy with 5.5% credit fee | costgoat.com, openrouter.ai |
Full source URL list in the prior chat message that generated this doc; refresh quarterly.
Document dated 2026-05-15. Supersedes scattered model preferences across crates/vox-config/src/inference.rs, crates/vox-orchestrator/src/models/, and pre-2026-Q2 commits in contracts/orchestration/. Next review: at v0.6 release or 2026-08-15, whichever is sooner.
§8 Selection SSOT — select() (Council-ratified 2026-05-15)
Section titled “§8 Selection SSOT — select() (Council-ratified 2026-05-15)”The model pipeline ships a single source of truth for model selection:
crates/vox-orchestrator/src/models/select.rs.
Every Vox surface that picks an LLM should flow through select(intent, registry)
rather than hardcoding a model id or duplicating axis logic.
8.1 Three-axis user knob
Section titled “8.1 Three-axis user knob”Users control selection along three orthogonal axes (each 0–100):
| Axis | Meaning | Projects onto AutoRoutingPriority |
|---|---|---|
| cost | 100 = cheapest possible | efficiency |
| responsiveness | 100 = lowest latency | latency |
| intelligence | 100 = highest capability | precision |
Presets:
COST_FIRST— 70/15/15 — classifiers, CI lints, NLI checksBALANCED— 33/33/34 — defaultQUALITY_FIRST— 15/15/70 — review, security audit, debugging, research, planningFAST— 15/70/15 — IDE autocomplete, ghost-text
Set via env: VOX_MODEL_AXES=cost:80,intelligence:10,responsiveness:10.
8.2 Caller-hint intents
Section titled “8.2 Caller-hint intents”| Constructor | Task | Axes | Notes |
|---|---|---|---|
repair_loop() | CodeGen | BALANCED | cacheable_workload = true (Anthropic prompt cache) |
research() | Research | QUALITY_FIRST | complexity 7 |
review() | Review | QUALITY_FIRST | cacheable; complexity 6 |
nli_classifier() | Parsing | COST_FIRST | hard ceiling max_cost_usd_per_call = $0.01 |
ide_autocomplete() | CodeGen | FAST | prefer_local = true |
plan_mode() | Planning | QUALITY_FIRST | complexity 8 |
8.3 Resolution order
Section titled “8.3 Resolution order”VOX_MODEL_FORCEenv override → immediate return if id matches catalog.prefer_local→ Ollama / VoxLocal / PopuliMesh providers only.- Premium alias → when
axes.intelligence >= 50, honor the pin incontracts/orchestration/model-routing.v1.yaml. - General scorer →
ModelRegistry::best_for_with_filter()with axes projected ontoAutoRoutingPriorityand filters formax_cost_usd_per_callcontext_size_hint.
Each outcome carries a SelectionReason (PremiumAlias/Scored/LocalOnly/EnvOverride)
so debugging routing surprises is a one-line lookup.
8.4 Migrated call sites (2026-05-15)
Section titled “8.4 Migrated call sites (2026-05-15)”| Crate | Before | After |
|---|---|---|
vox-cli::commands::repair | hardcoded REPAIR_LOOP_PREFERRED const | select_with_default_registry(&SelectionIntent::repair_loop()) |
vox-orchestrator::task_dispatch::research_dispatch | hardcoded "anthropic/claude-3.5-sonnet:beta" (retired) | SelectionIntent::research() |
vox-code-audit::review::providers (defaults) | claude-3.5-sonnet / gpt-4o-mini / gemini-2.5-flash | Sonnet 4.6 / GPT-5-Mini / Gemini 3 Flash |
vox-code-audit::ai_analyze (gemini default) | gemini-2.5-flash | gemini-3-flash |
Future migrations (incremental — they operate at the provider-layer, below select()):
vox-actor-runtime::model_resolution, vox-orchestrator-mcp::resolve_mcp_chat_model_sync,
vox-gamify::reward_routing.
8.5 Cleanup completed in this pass
Section titled “8.5 Cleanup completed in this pass”- Retired model
claude-mythos-preview-20260407removed from catalog + tests + docs. claude-3.5-sonnet,claude-3.5-sonnet:beta,gpt-4o,gpt-4o-mini,gemini-2.5-flashpurged from active code paths (still referenced from retirement contracts).llama3:latestOllama bootstrap replaced withqwen3-coder-next-32bas Local default.REPAIR_LOOP_PREFERREDretained as a last-resort fallback only.