Model Guide

Model Guide

Select between the current frontier models and prompt them correctly.

Models covered: Claude Opus 5 (primary Claude), Claude Sonnet 5 (cost-tier Claude), GPT-5.6 Sol / Terra / Luna (GA 2026-07-09; GPT-5.5 removed), Kimi K3 (Moonshot open-weight, 2026-07-16), and Grok 4.5 through the Cursor subscription pool. Claude Fable 5 remains a metered opt-in (off subscription 2026-07-07; see below). Last updated: 2026-07-24 (Claude Opus 5 replaces Opus 5 as primary Claude default). Active stance: This skill no longer maintains a broad model zoo. Older GPT, Gemini, Grok-4.20-and-earlier, and Sonnet-4.6-and-earlier routes were removed from active guidance. Sonnet 5 is reinstated as a named, cost-tier Claude option (2026-06-30). Grok 4.5 is an opt-in read-only repo critique lane through exact Cursor slugs; the xAI API path remains separate and blocked/unverified locally. Use this guide for high-value frontier decisions; use repo-specific batch tooling or search tools for cheap bulk work.

OPEN QUESTION (2026-06-30, not yet resolved — operator call): the "Architecture / design / high-reasoning critique → NEVER Sonnet" verdict below was reached against Sonnet 4.6 on 2026-06-20. Sonnet 5's system card shows large agentic/coding gains and prompt-injection robustness tying or beating Opus 4.8 in several places, but also the worst prefill/system-prompt-susceptibility numbers of the compared models and measurably more turns/tokens per task (system-card digest: references/sonnet-5-system-card.md). Whether this changes the "NEVER Sonnet" verdict for architecture/critique work is a live question, not re-litigated here — the verdict stands until the operator revisits it.

Claude Fable 5 — status (2026-07-12). Off the claude.ai Pro/Max/Team subscription since 2026-07-07: continued access is priced at metered usage credits, $10/$50 per MTok (2× Opus 5) — press/pricing-page sourced (techtimes.com, bleepingcomputer.com, claude.com/pricing); reconciliation against observed usage is open, see Verified Transport below. Fable is reachable via llmx chat -m claude-fable-5 (claude-cli transport, confirmed working) and headless claude -p --model claude-fable-5 (confirmed 2026-07-04) — not reliably via the Agent tool, where fable-high/fable-low-style dispatches currently serve claude-sonnet-5 regardless of the pin (measured 2026-07-12, see Verified Transport — this is a mechanism bug, not a re-dormancy). Route gated/briefed/review dispatch to opus-low ($0 subscription); reach for Fable (via llmx, not the Agent tool) only with a named Fable-specific capability-edge justification over Opus max. Re-license trigger: Anthropic restores Fable to subscription plans.

Opus 5 (claude-opus-5) is Anthropic's active top-tier model (released 2026-07-24): near-Fable intelligence at Opus price ($5/$25), 1M context, adaptive thinking on by default, SOTA on Frontier-Bench / GDPval-AA / ARC-AGI 3 / AutomationBench / OSWorld 2.0 cost-efficiency. Default for hardest Claude work, security/cyber/biology (Fable bio blocks now route here), and cross-lab review. Architecture → max effort. Keep claude-opus-5 only as the documented cyber-classifier fallback target.

Verified Transport — what actually serves what (as-of 2026-07-14)

Routing judgment (which model you want) and routing mechanism (whether the lane you dispatch to actually delivers that model) are different questions — this table is the second one, and it currently has a serious hole. Re-verify any row before a tier-sensitive decision leans on it; mechanisms drift faster than judgment.

Until the Agent-tool bug is fixed: any Agent-tool dispatch where the model tier is load-bearing (a cost claim, an eval arm, a "frontier vs cheap" comparison) needs a one-line self-report opening the brief, read back before trusting the result. One line catches a silent tier swap that otherwise bills or behaves as the wrong model.

Fable cost status is unreconciled, not merely unverified: the "$10/$50 metered" claim is press/pricing-page sourced; the llmx usage log shows claude-cli-transport Fable calls completing normally (large completions, zero errors) through 2026-07-12, after the cited cutoff, and the log has no cost/auth-mode field to say which billing path fired. Whether Claude Code's own OAuth entitlement is a separate pool from the claude.ai Pro/Max/Team plans the press covered is unverified (2026-07-12) — check actual Console billing before a batch decision hinges on "still $0" or "now expensive."

Default Routing

Judgment below assumes the lane you dispatch to actually delivers the named model — confirm that against Verified Transport above before trusting a routing choice for a tier-sensitive dispatch.

Quick Selection Matrix

For full score tables, read references/BENCHMARKS.md.

Selection trilemma (capability × calibration × efficiency)

Benchmark capability (Intelligence Index, SWE scores) and parameter count are weak proxies for real-world usefulness. They often invert on calibration — whether a miss is an abstention or a confident fabrication — and on efficiency — tokens/time to reach a correct or honest answer.

Settled ordering on calibration (AA-Omniscience, misses only, abstention invited): GLM-5.2 72% non-hallucination → Opus 4.8 64% → Grok 4.5 ~46% / Fable 5 45% → prior GPT class 14% (re-measure 5.6) → DeepSeek V4 ~6%. Capability ordering is nearly the reverse (Grok Intelligence Index 54, near Opus 56). A multi-trillion-parameter model can score at the top of an index and still be the worst choice when the task needs "I don't know" or detection of an impossible/contradictory spec.

Reasoning budget is not monotonic. On badly calibrated models, more reasoning often buys longer confident wrong answers, not better ones. Anecdotal corroboration (Shrimpton 2026-06-18, n=1, high effort, temp 1): an impossible asyncio event-loop spec — DeepSeek V4 Pro ~7.7k reasoning tokens, 3m52s, full wrong implementation; GLM-5.2 ~800 tokens, 12s, correctly flagged the paradox. Don't throw xhigh/max at poorly calibrated GPT or DeepSeek for epistemic guardrails; use Opus, GLM (opt-in), or deterministic impossibility checks.

Consumer rule: match model to the axis that matters for the task — capability for gated mechanical work with a verifier; calibration for unsourced facts, paradox detection, and "should we even do this?"; efficiency for throughput. Never select on size or index rank alone.

Transport facts (llmx — not judgment)

Read before dispatch: ~/.claude/cache/llmx-routing.json (regenerate: llmx info --write-mirror). Transport table, effort maps, exit classes live there — not in this skill.

Claude policy: NEVER anthropic-direct/API by default. Subscription only (llmx chat --subscription, claude -p with key stripped, Agent tool) unless the user explicitly requests metered API billing.

Probe subscription path before critique batches:

llmx chat --dry-run --subscription -m claude-opus-5 -e max
# or: uv run python3 ~/Projects/skills/critique/scripts/model-review.py --preflight

Mechanics and footguns: /llmx-guide.

llmx Cosigner / Dispatch Defaults (judgment — transport in mirror)

• Cosigner / critique / synthesis: gemini-3.5-flash (inverted from 3.1 Pro 2026-05-24, operator-empirical; re-confirmed 2026-06-13 — flash-3.5 ≈ GPT-high ≫ 3.1-pro on the ADR-0009 spine critique). Always in the 2G+2GPT mix — never the only reviewer. Probe flags invention on clean packets; orchestrator dispositions via --extract --verify (see agentlogs evidence).
• Cheap classification / mechanical audits: gpt-5.6-luna at low via codex-cli subscription ($0). gemini-3-flash / flash-lite — Gemini is critique-only since 2026-07-14 (metered gemini-* refuses outside the /critique engine; agent-infra decisions/2026-07-14-gemini-critique-only-policy.md).
• GPT-5.6 default effort is medium (suite supports max beyond xhigh) — pass -e high/xhigh for depth; reasoning bills as output.
• GLM-5.2 (Z.ai, NEW LAB) = opt-in review cosigner, NOT an extractor (2026-06-19). A 4th independent training lab (Zhipu) → real cross-lab diversity for critique; request explicitly --axes …,glm (glm_review profile, routed via OpenRouter). Calibration edge: 72% AA-Omniscience non-hallucination (2026-06-18 independent read) — best among commonly-routed large models, ahead of Opus 4.8 64%; strong on impossibility/paradox detection in anecdotal coding probes. Accepts ONLY high/xhigh reasoning (no low tier) → structurally expensive+slow → rejected for high-volume extraction/ingestion (cost-dominated, no quality gain; keep gpt-5.3/gemini-3-flash). Match reasoning floor to task: GLM for occasional thorough review and epistemic guardrails, not throughput. See agent-infra/decisions/2026-06-19-glm-5.2-integration.md, evals DECISIONS glm-5.2-extraction.
• Grok 4.5 is routable through Cursor subscription as of 2026-07-14. Use exact cursor-grok-4.5-{low,medium,high} or matching trailing--fast slugs. The opt-in critique grok axis pins cursor-grok-4.5-high in a read-only repo workspace and fails closed on registry or unrevealed repo-canary drift. The bare grok-4.5 xAI API lane remains separate.
• Gemini 3.6 Flash / 3.5 Flash-Lite (launched 2026-07-21) are REGISTERED, NOT ROUTED (2026-07-22). Live API ids gemini-3.6-flash, gemini-3.5-flash-lite (GA, no -preview suffix; verified against models.list, not guessed). Registered in llmx (652d1ed) purely so the spend guard stops refusing them as unpriced — the 2026-07-14 critique-only policy is unchanged and no default moved. Prices (verified at ai.google.dev/gemini-api/docs/pricing 2026-07-22): 3.6 Flash $1.50/$7.50, 3.5 Flash-Lite $0.30/$2.50, 3.1 Flash-Lite $0.25/$1.50. Same pass corrected two badly stale entries — gemini-3-flash was priced in llmx at $0.075/$0.30 against an actual $0.50/$3.00, so cost dashboards were understating Gemini ~7-10x. Note 3.5 Flash-Lite is 6x input / 12.5x output the price of 3.1 Flash-Lite — the "Lite" tier is no longer a rounding error. Effort ladders probed live: 3.5-Flash-Lite accepts minimal, 3.1-Flash-Lite rejects it (do not pin minimal on the older one).
• Do NOT reach for Flash-Lite as the cheap extraction lane — gpt-5.6-luna stays it. Luna is $0 on the ChatGPT subscription; Flash-Lite is metered under a policy that only permits /critique. A metered lane cannot beat a $0 lane on cost, so Flash-Lite would have to win big on quality, and our own screening probe says it does not (see below).
• Open, operator's call — 3.6 Flash as the /critique cosigner in place of 3.5 Flash. Strictly cheaper on the one lane Gemini is still allowed on: $7.50 vs $9.00 output and a vendor-claimed ~17% output-token reduction, i.e. roughly -30% on cosigner spend. NOT changed unilaterally — the gemini-3.5-flash cosigner default was set operator-empirical (2026-06-13, re-confirmed), and a vendor claim is not evidence that it reviews as well. Swap is one line in the critique axes.
• llmx vision is multi-provider as of 2026-07-22 (llmx 2b12289) — it used to be Gemini-only and off-ledger. It now routes through the normal dispatch path, so -m takes any vision-capable model id (gemini-3.6-flash, gpt-5.6-sol, gpt-5.6-luna, claude-opus-5), provider is inferred, and -e effort works. Three consequences worth knowing: (1) it is spend-guarded and policy-gated like everything else — a Gemini vision call now needs LLMX_GEMINI_OK=1, where it previously dispatched freely; (2) it writes real token counts to the usage ledger, so vision cost no longer has to be estimated (evals/figure_vision_bakeoff had been substituting a len(response)/4 proxy); (3) media fails loud rather than being dropped — video to an OpenAI-compat endpoint, an oversized inline upload, or any media sent through a CLI transport (claude-cli/codex/cursor) raises, because a model asked about a figure it never received invents an answer. Footgun retained for back-compat with the documented convention: in llmx vision, -p is the PROMPT, not --provider (use --provider to override the inferred one). This unblocks a cross-family vision judge, which the figure-vision eval previously could not have — its qualitative judge was Gemini-flash grading a Gemini-flash candidate, a same-family COI it documented as forced by the tool. Pass --judge gpt-5.6-sol there now.
• gemini-3.1-pro-preview is RETIRED as a routing option (2026-06-13, operator). Do not route here for critique/synthesis/review — flash-3.5 dominates and is cheaper/faster. (Benchmark records in references/BENCHMARKS.md are kept as evidence; this is a routing retirement, not a data scrub. Callable via explicit -m if a one-off ever needs ARC-AGI-2/GPQA/video, but it is not a default anywhere.)
• Cosigner calibration caveat (AA-Omniscience, 2026-06-11): both cosigner defaults are bottom-quartile abstainers — non-hallucination 39% (gemini-3.5-flash), prior GPT class 14% (re-measure Luna/Sol TBD), despite an abstention prompt. Critique output = adversarial pressure on reasoning, never a fact source; for fact-heavy review where calibration matters, verify novel specifics at primary and lean on a frontier model (Opus/GPT), not a cheap cosigner. Instruments: agent-infra research/2026-06-11-aa-benchmark-instrument-validity.md.

Dispatch Economics (subagent executor tiers)

When dispatching subagents to execute work (Agent tool, headless claude -p, codex), the executor tier is set by how good the verifier in the brief is, not by how hard the task feels. Measured evidence: four preregistered evals, anim-workbench 2026-06-12 (anim-workbench/.claude/evals/2026-06-12-{dispatch-tier,effort-tier,codex-lane,effort-integration}/), all n=1 per arm (screening grade).

Every row above assumes the lane delivers the named model — false for Agent-tool pins as of 2026-07-12 (Verified Transport). Self-report-check any row where the tier is what's being measured.

Role → Lane (dispatch execution roles)

Codex lane mechanics (from codex-lane eval): codex exec --full-auto -C <out-of-repo-worktree> -c model_reasoning_effort="low" — the -c override is verified per-invocation (resolved effort confirmed in rollout logs). Gotchas: pre-install deps (the shell sandbox has no network); git commit fails inside worktrees (gitfile points outside workspace) — grade the dirty tree, commit from outside; require a final-message manifest (the -o empty-output gotcha).

Codex as a research/work subprocess (verified 2026-06-18): codex carries the same skills + MCP stack as Claude (~/.agents/skills/, ~/.codex/config.toml) — invoke a skill in the prompt via its $name keyword (single-quote the prompt). Network-backed MCP tools (research-mcp, exa, brave, scite) DO work under --full-auto — MCP servers are separate processes, so the shell-sandbox "no network" gotcha above does NOT apply to MCP calls. So a codex worker can do real (not training-memory) research and write its own memo, at $0 on the subscription. Full pattern — $skill invocation, canary-first discipline, stub-first/intern-rule briefs, the commit-word hook false-positive, benign MCP-teardown noise, llmx-is-not-the-vehicle: references/codex-subprocess-dispatch.md.

The conditioning rule: low effort doesn't mean less verification — both eval arms ran every gate because the gates were written in the brief. Self-initiated checking is what higher effort buys; an explicit verifier in the brief makes that purchase unnecessary. So the brief MUST carry: verification commands (exact, runnable), cleanup directives (worktree/scratch teardown), and a files-touched manifest requirement. A cheap executor on a gate-less brief is the worst quadrant.

Intern rule (gate-less delegation): exploratory, divergent, or conceptual dispatches (research sweeps, brainstorms, design options, synthesis) have no mechanical gate to put in the brief — so the coordinator's review IS the gate. Treat the return like an intern's draft: don't re-do the work, but spot-check it before adopting. Concretely: re-run 1-2 of its load-bearing probes/citations yourself, check one claimed source actually says what's claimed, run the completeness check (does every input appear in the output, are dropped items justified), and ask what the brief would have rewarded the agent for skipping. Scale the spot-check to stakes — a brainstorm needs a sniff test, a synthesis feeding a decision needs the citation check. Skipping this turns "delegate" into "launder": unverified subagent output adopted wholesale is the same failure as adopting cross-model critique without cosigning.

Effort knob mechanics: the Agent tool exposes only model:. Per-dispatch effort exists via (1) headless claude -p --model opus --effort low (verified working, CLI 2.1.175; background Bash + --output-format json for usage), or (2) .claude/agents/*.md frontmatter effort: (does NOT hot-register mid-session — usable only in later sessions). Codex/GPT cheap cosign via llmx --subscription is $0 — probe with --dry-run --subscription first; transport table in ~/.claude/cache/llmx-routing.json.

Agent-tool DEFAULT model is NOT the session model (2026-06-29). general-purpose/most subagents default to CLAUDE_CODE_SUBAGENT_MODEL (now claude-sonnet-5 — Sonnet 4.6 RETIRED 2026-07-07, never route to it), NOT the parent's Opus. A bare Agent(...) with no model: runs Sonnet 5 — fine for bounded work, a tier silently-wrong trap when the dispatch IS the measurement (an eval baseline, a "frontier agent" arm). The previously-recommended fix — pass model: explicitly, then grep '"model"' the transcript — is not proven sufficient as of 2026-07-12: the newer bug (Verified Transport) shows a pin can be requested and still not be served, and whether transcript-grep reflects the request or the actual serve is untested (ASSUMPTION: probably the request, since that would explain why grep-verification didn't already catch this). Require a first-line self-report instead — the one channel confirmed to reflect the true served model. Second footgun, same 2026-06-29 session: an open-ended "be exhaustive" prompt to general-purpose triggered sub-delegation + stall (6 children spawned, "I'll pause here," 72K tokens burned, nothing delivered) — for bounded research dispatches, explicitly forbid delegation.

External validity: all four evals are regime-1 (clear mechanical verifiers — tsc, deterministic scripts, numeric oracles) and screening-grade (n=1/arm). Only within-eval contrasts are clean — cross-eval comparisons are confounded by task, brief density (briefs improve as the author learns, flattering later arms), and harness (codex carries MCP servers + sandbox; opus arms ran bare). Every cheap-lane verdict is conditional on the dispatch-time classification "fully-briefed + mechanically gated" being honest — nothing here licenses cheap lanes for judgment-shaped or incomplete-spec work. The greenfield→integration replication trigger from the morning run is SATISFIED (effort-integration, port shape); the standing revocation trigger replaces it.

Reasoning escalation guard (calibration × effort): the cheap-lane evals show less reasoning is fine when the verifier is in the brief. The inverse also holds outside regime-1: escalating effort on poorly calibrated models (GPT family until re-measured, DeepSeek V4) on paradox/impossibility or unsourced-fact tasks tends to produce more confident fabrication, not more abstention — see Selection trilemma. Effort buys depth only where calibration is already adequate (Opus, GLM for review).

Claude Opus 5 - "Near-Fable daily driver" (primary Claude)

Use for: all active Claude frontier work — hardest autonomous runs, codebase-scale migrations, architecture, code review, security/cyber/biology, professional analysis, legal/financial reasoning, long autonomous loops, and cross-lab critique. Keeps Fable-tier routing by default: Fable is metered+paid and unreachable via the Agent tool (Verified Transport), so Opus is the practical default even where Fable might otherwise win on capability.

Operational specs: claude-opus-5, 1M context (default = max), 128K max output (300k batch beta), $5/M input and $25/M output (same as 4.8). Fast mode ~2.5× speed at 2× price ($10/$50). Adaptive thinking on by default; effort default high on API/Code. Knowledge cutoff May 2026 (training). Subscription-routable (lite_allowed_models). Cyber-classifier refusals can auto-fallback to claude-opus-4-8; bio refusals on Fable now route here.

Launch routing line (2026-07-24): near-Fable capability at half Fable's price; Anthropic claims SOTA on Frontier-Bench, GDPval-AA, and best cost-efficiency on OSWorld 2.0 / AutomationBench / ARC-AGI 3 (≈3× next-best). Efficiency at low/medium effort is a real lever — re-sweep effort defaults. Most aligned of recent Claude models on Anthropic's automated behavioral audit (misalignment score 2.3). Prompting deltas vs 4.8: longer default verbosity (prompt for concision), stronger self-verification (remove redundant "verify again" scaffolding — it over-verifies), more subagent-eager (cap delegation), thinking-disabled capped at high effort. Digest: references/opus-5-system-card.md. Prior 4.8 card kept for calibration history: references/opus-4-8-system-card.md.

Two shape-changes the launch line misses (system card §8.12 / §8.2 / §2.2 — read 2026-07-25, agent-infra research/2026-07-25-opus5-arc-agi-generalization.md):

• Tools beat effort — spend the budget there first. §8.12 verbatim: "agentic tool-use is generally a more cost-effective method of scaling test-time compute than adaptive thinking by itself." Before raising a dispatch one effort tier, give it a verification command / probe / read-back tool instead — cheaper and stronger. This is a cost lever, not only a quality one.
• Attach the image. SWE-bench Multimodal 38.4→59.4 (+21pp) is the single largest coding delta in the release; OSWorld 2.0 +15pp. Screenshot-the-render / plot / broken-UI and hand it the source, instead of describing the visual defect in prose. Applies to rendered frames, QC plots, dashboards, CAD.
• Its exploration gain is verifier-conditioned — this is the liveness rule. ARC-AGI-3 1.5→30.2 (20×, dense per-action score) sits in the same card as §2.2, where two Opus 5 arms of a 24h autonomous design campaign delivered nothing and one went silent for its final 8 hours in self-verification loops (no in-loop verifier). Give any long autonomous run a per-step check it can score against, and bind completion to an advancing artifact — a live PID proves the process runs, not that it progresses.

Prompting and API rules:

• Use XML tags; adaptive thinking explicit (thinking:{"type":"adaptive"}); no manual budget_tokens.
• Default effort high; max for architecture/design/high-reasoning critique (operator 2026-06-20); xhigh for serious coding/review/long agentic work; low for gated mechanical dispatch (see Dispatch Economics).
• Measured effort curve (Artificial Analysis, 2026-07-25) — max is a poor default: AA Intelligence Index by effort — low 51 ($556 / 12M out-tok), medium 56 ($1,115 / 29M), high 59 ($1,974 / 52M), max 61 ($3,836 / 100M). low→max = +10 points for 6.9× cost; high→max = +2 points for +$1,862. At max AA flags it "very verbose" (100M vs 63M median); at high, "fairly concise." With §8.12 (tools scale test-time compute better than thinking), the rule is: default high, escalate to max only for architecture/irreversible calls, and spend the delta on an in-loop verifier instead of the top tier. AA's task mix ≠ ours — a strong prior, not a workload-specific verdict.
• Calibration warning (AA, independent): AA-Omniscience Index 31 — below Fable 5's 40, despite Opus 5 leading the Intelligence Index at 61. It leads on intelligence and trails on confident-wrongness — keep provenance tagging and claim verification on for factual work. Direction corroborated by the card itself (§6.5: "hallucinates factual claims slightly more than Opus 4.8, despite being more accurate overall"). ⚠ A widely-quoted "hallucination +14pp → 50%" figure is UNVERIFIED — it traces to a search-engine summary of an @ArtificialAnlys X post, and AA's own pages do not publish per-model accuracy/hallucination for Opus 5. Cite the Index gap (verified), not the 50%.
• Mid-conversation role:"system" messages supported immediately after a user turn — use for permission/budget/environment updates without rebuilding the prompt.
• No non-default temperature/top_p/top_k (400 on 4.7+); no assistant prefill; min cacheable prompt 1,024 tokens.
• Put long documents first and the query/instructions last.

Full guide: references/PROMPTING_CLAUDE.md.

Claude Sonnet 5 - "The Cost Tier" (added 2026-06-30)

Use for: cost-sensitive coding and agentic work with a mechanical gate (tests, typecheck), mechanical no-gate dispatch (rename sweeps, boilerplate), and anything where untrusted tool output

…(truncated)