Which model should each Harbour call run on, and which should the free tier run on?

8 November 2026 · 16 min read · Claude, for John Kershaw

On the catalog read 8 Nov 2026 11:14 UTC and the Library analysis of 23 Oct 2026, John's sense is right, for one model. OpenAI's GPT-6 Luna (openai/gpt-6-luna) costs $0.10 in and $0.50 out per million tokens, against $0.75 and $4.50 for today's default, GPT-5.4 Mini, and is ahead of it on every independent board that scores both. I recommend it for the free tier and for 14 of the 16 Harbour calls, each only after that call has been shown to come back whole on real tasks. Two calls stay put. The routing call (recommend) stays on GPT-5.4 Mini: it picks the stage every dispatch follows, it is the only call with a repeated dense-ticket pass (LIN-263), and it costs about two cents. The brief stays on GPT-6.1 Sol: nothing cheaper scores higher in either source, and Sol comes out whole. Claude Haiku 5.5 scores above Luna at the same base price in both sources, but every Haiku 5.5 option waits on the LIN-4352 brief replay (docs/areas/llm-and-prompts.md). Harbour's own calls are a small bill ($54 on October's 271 Done tickets, cheaper-capacity.md Method), so on John's key the case is quality; on the free tier, which bills the operator's shared key, price matters too.

Two benchmark sources, kept apart

Catalog (8 Nov). OpenRouter's /api/v1/models attaches Artificial Analysis figures (benchmarks.artificial_analysis). Of the 379 models that are not :batch variants, 145 carry at least one figure. 56 carry the intelligence index, 122 the coding index, and only 33 carry both. The catalog states no index version and no reasoning effort.

Library (23 Oct). cheaper-capacity-benchmarks.csv copied the AA leaderboard by hand. Its figures are at the "best listed effort", which is max for most rows, and it records the index as "v4.3.x per secondary sources". Its Vals rows (Terminal-Bench 4.0, Vibe Code Bench) are independent runs with one harness each.

The two AA sources agree at the top and drift apart lower down:

The gap is not one shift: near zero for frontier models, 6–16 points down for the cheaper ones above (Kimi K3, not shown: 27). That fits a revised index that is harder at the bottom, or figures taken below the Library's max effort. The catalog states neither, so I cannot tell which. It matters because the order among cheap models depends on the source: the Library puts DeepSeek V4.1 Flash (45) above Haiku 5.5 (43) and Luna (38); the catalog puts Haiku 5.5 (36.5) above DeepSeek (28.6), Luna (27.9) and GLM-5.3 Flash (27.8), which sit level. Both agree that Sol 6.1 and Sonnet 5.5 sit level near the top, that Haiku 5.5 is above Luna, and that both are far above Haiku 4.5 (catalog 7.4). The tables below use the catalog, with the Library figure beside it where one exists.

GPT-5.4 Mini has no intelligence figure in either source. The catalog gives only the older pair, coding 56.1 and agentic 17.9, so it cannot be put on Luna's scale directly. Three indirect readings:

• Independent boards (Library, Vals). Terminal-Bench 4.0: Mini 2.52, Luna 13.64 (22 Oct). Vibe Code Bench: Mini 47.97, Luna 81.65 (20 Oct).
• Luna's predecessor. GPT-5.6 Luna carries the older pair and beats Mini on both (coding 71.4, agentic 42.1), with an intelligence figure (23.8) below GPT-6 Luna's.
• Comparable models. Of ten non-free catalog models with coding 43–59 and an intelligence figure, nine score 4.4–18.9 and one (Solar Pro 4) 28.2. So Mini probably sits below Luna's 27.9: an inference, not a measurement.

Recommendation by call

Each call is a job in lib/llm-jobs.js; its model comes from the workspace's routing table (Settings › Models), falling back to DEFAULT_MODEL (openai/gpt-5.4-mini). "Now" is the code default; John's table may differ, and I could not read it. The brief's move to Sol on 7 Nov is a workspace setting, not code. "+ headroom" means lib/call-budget.js adds reasoning headroom from the catalog and, unless the call allows partial output, retries once and then refuses rather than return part of an answer. Prices are catalog $/1M.

Gates:

• W (whole). Set the job's model for John's workspace only (Settings › Models), run it on three real tasks, one dense (LIN-4352 or LIN-344), and read the LLM call log: every call ends finishReason=stop (first attempt or the one retry), nothing is refused, and the output reads right against GPT-5.4 Mini's for the same task. Only the brief has a replay script, so for the rest the routing table and the call log are the test.
• T (tools). W, plus one turn that calls a read tool and one that proposes a follow-up, both whole. Add openai/gpt-6-luna to SUGGESTED_MODEL_IDS first: on a cold catalog supportsTools() falls back to that list, and a model missing from it runs without tools.
• R (roadmap). One full generation with every layer ending stop, on a warm catalog. Since LIN-4358 the reasoning split comes only from the catalog (an explicit bound when it says mandatory or default_enabled: true, as it does for Luna; headroom otherwise); on a cold cache these calls send a bare prose cap and may show a cut-off layer, because they take partial output and are not retried.

Haiku 5.5 could replace Luna on any row: the same base price ($0.10/$0.50; $0.50/$2.50 over 100k prompt tokens) and higher in both sources (36.5 / 43; Terminal-Bench 4.0 36.36 against 13.64). Only if the LIN-4352 replay passes (block [2] WHOLE, finishReason=stop, any retry row accepting the reasoning bound), and then with its own dense-ticket check, because Haiku 4.5 is the model LIN-263 saw mis-read dense tickets three times in three.

The free tier

Today. Free-tier calls bill the operator's shared OPENROUTER_FREE_TIER_KEY. resolveFreeTierModel() (lib/openrouter.js) returns OPENROUTER_FREE_TIER_MODEL if it is in FREE_TIER_MODEL_IDS, else DEFAULT_MODEL; the clamp (lib/workspace-preferences.js) ignores workspace preferences and per-request overrides. So one model serves 15 of the 16 calls on the free tier (observer-pass resolves without the clamp). I could not read the production environment; if the variable is set, the tier may already run on another of the ten allowed ids.

Recommendation: openai/gpt-6-luna.

• Better on the evidence. Terminal-Bench 4.0 13.64 against 2.52; Vibe Code Bench 81.65 against 47.97 (Vals, Library, 20–22 Oct). Catalog AA intelligence 27.9 against none; its predecessor beats Mini on the catalog's coding and agentic fields.
• Same family as the validated default. Same reasoning field (optional, on by default at medium); 1.05M context (Mini 400k); 128k output; tools. Listed since 22 Sep 2026, no expiry date.

Price. $0.10 in and $0.50 out per 1M ($0.20/$0.75 above 272k prompt tokens), against Mini's $0.75/$4.50.

That is 7.5–8.5 times cheaper per call. The 8,300-token brief input is the ticket discussion's figure; this paper did not measure it.

What has to change (none of it done here): add openai/gpt-6-luna to FREE_TIER_MODEL_IDS, whose comment names this gate; add it to SUGGESTED_MODEL_IDS for the cold-catalog tool fallback; set OPENROUTER_FREE_TIER_MODEL=openai/gpt-6-luna and check that getFreeTierModelConfigWarning() reports nothing at startup. Moving the free tier also moves free-tier recommend, which this paper otherwise keeps on Mini.

What must be shown whole before switching. Every free-tier job, on Luna, through a free-tier session: recommend (its gate above); brief (replay-brief.mjs on LIN-4352 with openai/gpt-6-luna, block [2] WHOLE); recap, scan and scan-retire (W, JSON parsing); run-summary, session-summary, run-paragraph and feedback-title (W); task-chat and flight-companion (T); the three roadmap calls and ship-biscuit (R). The tier has one model, so if any job fails it stays where it is.

Disposition (8 Nov 2026, John's rulings on LIN-4358). The free tier stays on GPT-5.4 Mini. FREE_TIER_MODEL_IDS stays an allowlist, not a price ceiling, and pickers get no benchmark column (ruling lin4358-free-tier-and-quality-source, 12:41 UTC). GPT-6 Luna stays this paper's recommendation, evidence pending: no free-tier job has been run on it, and LIN-4358 closed without that run (ruling lin4358-luna-evidence, option stay-mini, 12:54 UTC). The move needs the whole-job run above, done with an OpenRouter key, followed by the changes listed under What has to change. The per-call rows above are still recommendations, each behind its own gate. No call moved with this paper.

Alternative, conditional. Claude Haiku 5.5 has the same base price and scores higher in both sources. It is not in the free-tier allowlist either: lib/openrouter.js says it was deliberately left out. It qualifies only after the LIN-4352 replay passes and the same job list is shown whole. Above 100k prompt tokens its price rises fivefold.

Candidates considered

From the live catalog (8 Nov): tool-capable models with an AA figure, plus those Harbour uses now. Reasoning is mandatory/optional, default_enabled (? where absent), default effort. Terminal-Bench 4.0 is Vals via the Library (22 Oct).

Why not the cheaper ones. GLM-5.3 Flash, DeepSeek V4.1 Flash and MiniMax M3.1 Flash cost less than Luna and, on the catalog, score level with it or above it on coding and agentic. I do not recommend them now. GLM and MiniMax make reasoning mandatory at max effort, a latency and cut-off risk on 60–400-token calls, and Harbour has run none of the three on these calls. LIN-263's cheap open-weight models failed: DeepSeek V4-Flash, an earlier build, mis-routed the dense leaf twice in three, and Qwen3.7-Plus gave unparseable action lines at 40–90 seconds a call. Luna comes from the family Harbour's call layer was built around. They are the next candidates once Luna's gates have run. GPT-6.1 Terra costs half Sol's price and scores 3.3 points below it; it is the brief's cheaper option, but was listed on 26 Oct and has no Library or Vals figure.

What this cannot tell you

• Harbour's own tasks. AA indexes and Terminal-Bench measure general reasoning and agentic coding. Harbour's calls read one dense ticket and write fixed-shape prose or JSON. LIN-263 found failures no benchmark predicted: Haiku 4.5 said "no plan exists" with the plan in the ticket body; Gemini 3.5 Flash wrote its reasoning before the recap JSON and left the recap empty; the newest model, Qwen3.7-Plus, fitted worst. The catalog shows which model is likely to be better, not that it will hold a call's shape, so every row above has a gate.
• Reasoning behaviour under Harbour's caps. The catalog's reasoning flags are not reliable. On 7 and 8 Nov it said default_enabled: false for GPT-5.4 Mini, which Harbour has measured spending hidden reasoning (LIN-1000, LIN-1665). It gives no default_enabled for Haiku 5.5. Why Haiku 5.5 stopped mid-brief on 7 Nov is still unproven.
• Which index. The catalog states no AA version or effort, and for cheap models its figures sit 6–27 points below the Library's, so rankings among them are soft.
• Latency, uptime and provider. None of these is in /api/v1/models. OpenRouter may serve one model from several providers, and this paper did not check the endpoints.
• Volumes. Token counts were read for recommend (1 Oct baseline) and one Flight Companion turn (e37d20e0). The other calls' input sizes were not measured.

How to re-run

curl -s https://openrouter.ai/api/v1/models > models.json
python3 -I candidates.py models.json 1.25   # optional ceiling on output $/1M

import json, sys
CURRENT = {'openai/gpt-5.4-mini', 'openai/gpt-6.1-sol', 'anthropic/claude-haiku-5.5',
           'anthropic/claude-sonnet-5', 'openai/gpt-5.6-terra', 'openai/gpt-5.6-sol'}
ceiling = float(sys.argv[2]) if len(sys.argv) > 2 else 99
rows = []
for m in json.load(open(sys.argv[1]))['data']:
    if m['id'].endswith((':batch', ':free')):
        continue
    aa = (m.get('benchmarks') or {}).get('artificial_analysis') or {}
    idx = [aa.get(k) for k in ('intelligence_index', 'coding_index', 'agentic_index')]
    tools = 'tools' in (m.get('supported_parameters') or [])
    p = m['pricing']
    pin, pout = float(p.get('prompt') or 0) * 1e6, float(p.get('completion') or 0) * 1e6
    if m['id'] not in CURRENT and not (tools and any(v is not None for v in idx) and pout <= ceiling):
        continue
    r = m.get('reasoning')
    rs = '-' if r is None else '%s, on=%s, %s' % (
        'mandatory' if r.get('mandatory') else 'optional', r.get('default_enabled', '?'), r.get('default_effort'))
    rows.append((idx[0] or 0, m['id'], pin, pout, m.get('context_length'),
                 (m.get('top_provider') or {}).get('max_completion_tokens'), tools, rs, *idx))
for row in sorted(rows, reverse=True):
    print('%-40s %7.3f %7.3f %9s %7s %-5s %-28s %5s %5s %5s' % row[1:])

Long-context tiers are in pricing.overrides, keyed by min_prompt_tokens. The Library column comes from cheaper-capacity-benchmarks.csv, not from this script.