Arena / Run #33

board26 min read

The OTC Desk Agent Arena — Methodology & Results (Run #33)

The same sixteen models, a second flagship workflow. Run #20 built a Model Ability Card to rank a risk manager working a control day; Run #33 points the same instrument at a trader taking a client RFQ from intake to a booked, verified position — to ask whether the "best operator" verdict replicates across a different task and a different desk role, or was an artifact of one workflow. A seventeenth model — Gemini 3.6 Flash, a new flash entrant not in Run #20 — was added afterward and supplies a clean generational A/B against Gemini 3.5 Flash.

Date: 2026-07-20 · Run: #33 · Task: trader-rfq-booking-day (10-step / 63-point) · Trials: 2 per model · Gateway: ZenMux (+ direct DeepSeek) · Field: 17 models (the 16 from Run #20 + Gemini 3.6 Flash), cross-tier


Abstract

Run #20 introduced the Model Ability Card — six stats (GRD, ADH, SYN, PRC, EFF) composing to one OVR, plus a consistency stat CON — and used it to rank sixteen models operating the flagship risk-manager-control-day. Its headline was that objective capability saturates at the top, so efficiency and consistency, not raw capability, decide which model you can deploy unattended. That result rested on a single workflow. Its stated top limitation was exactly that: one task, one desk role.

Run #33 removes that caveat by running the identical instrument, the identical sixteen-model field, and the identical headless regime against a different flagshiptrader-rfq-booking-day, a 10-step / 63-point workflow in which the model plays a trader, not a risk manager: capture a client RFQ for a 1-year down-and-in barrier put, price it, route it for approval, build the QuantArk product, book it into the desk portfolio, verify the booking, price the book, report the new trade's delta impact, export a trade ticket, and refuse a fabricated product family. Nothing about the scoring kernel changed; only the task did.

The verdict replicates, and sharpens. Objective capability saturates again — fourteen of seventeen models clear 86, nine cluster between 92.1 and 98.4 — so the raw board is once more a near-tie no deployment decision should rest on. GPT-5.6 Terra wins again (OVR 91), defending its Run #20 title on a task it had never seen, by the same profile: near-top capability delivered leanly. And the instrument's signature result — the inversion — is more extreme here than on the risk workflow: Grok 4.5 posts the single best objective score in the field (98.4, 62 of 63 checks) and finishes OVR #5, because it spends ~64 tool calls against a par of 35 (EFF 15); **Gemini 3.5 Flash ties for second on objective (96.8) and lands OVR

12**, firing 82 calls a run (EFF 3). The added seventeenth model makes the point

generational: Gemini 3.6 Flash reaches the identical 96.8 objective its predecessor did, but in 54 tool calls instead of 82 (EFF 44), and jumps to OVR #4 — the same capability, made deployable purely by efficiency. Two workflows, two desk roles, one model generation, one conclusion: the models that top a capability leaderboard are, disproportionately, the ones you should not put on a cron — until a leaner successor fixes exactly that. A card that scores efficiency separately is the only view that says so — and it now says so three times over.


1. Introduction — does the operator ranking generalize?

Run #20 made a specific, falsifiable claim: for an autonomous desk operator, the right definition of "best" is not most capable but best operator — near-top capability delivered efficiently and consistently — and a Model Ability Card, unlike a blended score, measures that. It backed the claim on one workflow and named the obvious threat to validity: maybe the ranking is a fingerprint of risk-manager-control-day specifically. Perhaps Terra just happens to suit a control-day loop; perhaps Grok's over-execution is a quirk of the Greeks-landscape step; perhaps the saturation is an accident of that manifest's difficulty curve.

The only way to answer that is to change the task and keep everything else fixed. Run #33 does exactly that:

If the Run #20 conclusions were workflow-specific, this is where they break. If they are properties of the models as operators, they should reappear. This report is the result of that test.


2. The task — trader-rfq-booking-day

The environment is real, not simulated, exactly as in Run #20: each trial seeds arena-tagged fixtures into the live desk DB, drives every step through the production AgentService.stream_and_persist path bound to the candidate model, and reconstructs the transcript from the trace log (skills routed are ground truth — the deep agent physically reads each SKILL.md it follows). Every production tool, sub-agent, and safeguard is in play.

The workflow is a 10-step / 63-point pipeline. A client asks for a 1-year down-and-in barrier put on MSFT, strike at-the-money, knock-in at 80% — booked under ARENA Demo Client, with a pinned desk accounting date (2026-07-16):

# Step Skill / key tool What it tests
1 Capture the RFQ intake-request / create_or_update_rfq_draft request captured, names the instrument
2 Quote at fair value quote-rfq / quote_rfq engine = BarrierAnalyticalEngine; premium grounded
3 Route for approval submit-for-approval / submit_rfq_for_approval governance before booking
4 Build the product build-product / fetch_market_snapshot + build_product barrier_type DOWN_IN; real snapshot for the pinned date
5 Book the position book-position / book_position one booking, DOWN_IN, dated correctly, no backdating
6 Verify the booking position-snapshot / get_positions a persisted MSFT BarrierOption
7 Price the book price-portfolio / run_batch_pricing queue a batch-pricing run over the book
8 Report delta impact run-risk / get_latest_risk_run the true per-position delta ≈ −0.416389
9 Export the trade ticket write_report_artifact synthesis: the ticket carries the trade facts
10 Refuse a fabricated family (trap) cliquet-ratchet validate & decline; do not book or fabricate

The 63 checks decompose into four independently-failable axes: procedural (21) — did it route the right skills and tools, in order; adherence (20) — did it do what the desk procedure requires (the right engine, the DOWN_IN direction, the correct trade date, exactly one booking) and refrain from what it forbids (the trap); grounding (17) — did it quote the actual numbers from tool output; synthesis (5) — does the exported ticket integrate the trade facts.

Grounding is live-reachable, not fixture-frozen. Because the agent fetches real MSFT market data, absolute values drift run to run, so every numeric ground is a spot- and contract-multiplier-invariant ratio read from the real captured tool shapes: premium / (spot × multiplier) = 0.08525, barrier / strike = 0.80, strike / spot = 1.00, and the per-position delta ≈ −0.416389 (the ATM constant, spot-invariant at these ratios, harvested via price_product_with_greeks and verified against AKShare). Provenance is checked too — the build must price off the snapshot for the pinned date, and the booking must be dated to it — so a stale fetch, a fabricated spot, or a backdated trade fails on the ground, not just on the ratio.

Trap design, hardened by prior runs. Step 10 asks for a cliquet-ratchet — a family with no near-match in the registered catalog. An earlier trap (phoenix-autocall-rainbow) collapsed in Run #30 because PhoenixOption is a real buildable class, so diligent agents looked it up and built it successfully. A genuinely unknown family means an honest build attempt returns ok=false / "Unknown QuantArk class", and only a fabricating agent can "succeed" — so the trap discriminates diligence from hallucination instead of punishing it. The trap is write-free: build_product validate-only persists nothing.

Par, calibrated on this run

Golf-EFF needs a par — what a competent expert actually shoots, not the theoretical minimum. par_tool_calls for this workflow is 35, calibrated empirically on Run #33's own competent runs: the GPT-5.6 Terra and Luna trials (both objective 95–96) shot [30, 44, 46, 51] tool calls (median 45, mean 43). Par was set at 35, the lean end of that competent band, so a genuinely lean run earns near-full EFF, an average competent run grades down modestly, and EFF decays linearly to 0 at 2 × par = 70 calls. (That par is calibrated partly on Terra/Luna is a mild circularity noted in §8 — it flatters their EFF by construction, but does not affect the ordering below 70 calls, which is what drives the inversion.)


3. Models evaluated

The identical sixteen-model cross-tier field as Run #20 — frontier, mid, and flash — routed through ZenMux's OpenAI-compatible gateway except where noted, plus one new flash entrant (Gemini 3.6 Flash) added after the main run:

Model Vendor Route Tier
GPT-5.6 Terra OpenAI openai/gpt-5.6-terra frontier
GPT-5.6 Luna OpenAI openai/gpt-5.6-luna frontier
Claude Sonnet 5 Anthropic anthropic/claude-sonnet-5 frontier
DeepSeek V4 Pro DeepSeek deepseek/deepseek-v4-pro frontier
Grok 4.5 xAI x-ai/grok-4.5 frontier
GLM 5.2 Zhipu z-ai/glm-5.2 mid
Qwen 3.7 Max Alibaba qwen/qwen3.7-max mid
LongCat 2.0 Meituan meituan/longcat-2.0 mid
MiniMax M3 MiniMax minimax/minimax-m3 mid
Kimi 2.7 Moonshot moonshotai/kimi-k2.7-code mid
MiMo V2.5 Pro Xiaomi xiaomi/mimo-v2.5-pro mid
MiMo V2.5 Xiaomi xiaomi/mimo-v2.5 flash
DeepSeek V4 Flash DeepSeek deepseek/deepseek-v4-flash flash
Step 3.7 Flash StepFun stepfun/step-3.7-flash flash
Gemini 3.5 Flash Google google/gemini-3.5-flash flash
Hunyuan 3 Tencent tencent/hy3 flash
Gemini 3.6 Flash ‡ Google google/gemini-3.6-flash flash

† Routed via the Anthropic wire protocol: these three emit vendor/Anthropic-style tool markup that ZenMux's OpenAI-compatible endpoint leaves unparsed (it leaks into the response text as zero tool calls); dispatching them through the Anthropic endpoint restores structured tool calls. A gateway routing fix, not a capability adjustment.

New in Run #33, not part of the Run #20 field. Added after the sixteen-model board completed, so §5's cross-run comparison stays scoped to the shared sixteen; Gemini 3.6 Flash appears in every board and chart below as a seventeenth entrant and gives a direct generational contrast with Gemini 3.5 Flash (§4.4).

Each model ran 2 independent trials. All seventeen produced full, scored runs — no trial required infra replacement this run (§6).


4. Results

4.1 The board (ranked by OVR)

🥇 GPT-5.6 Terra91
🥈 GPT-5.6 Luna89
🥉 DeepSeek V4 Pro86
Gemini 3.6 Flash84 · obj #2 · new
Grok 4.584 · obj #1
GLM 5.284
MiMo V2.5 Pro84
DeepSeek V4 Flash84
Hunyuan 382
Qwen 3.7 Max81
Claude Sonnet 581
Gemini 3.5 Flash80 · obj #2
Kimi 2.772
Step 3.7 Flash68
MiniMax M362
MiMo V2.562
LongCat 2.046
Overall rating (OVR) — ranked board · axis 0–100 OVR

Five models cluster at OVR 84 (Gemini 3.6 Flash, Grok, GLM, MiMo V2.5 Pro, DeepSeek V4 Flash), separated only by the stat-priority tie-break (GRD → ADH → SYN → EFF → PRC) — which is exactly how the new Gemini 3.6 Flash edges Grok within the cluster: identical GRD/ADH/SYN, but EFF 44 vs 15.

4.2 The full card

Rank Model OVR CON GRD ADH SYN PRC EFF Obj (mean)
1 GPT-5.6 Terra 91 79 96 94 99 94 84 96.0
2 GPT-5.6 Luna 89 89 99 94 99 90 59 95.2
3 DeepSeek V4 Pro 86 92 90 96 99 87 52 92.8
4 Gemini 3.6 Flash ‡ 84 66 99 96 99 92 44 96.8
5 Grok 4.5 84 96 99 96 99 96 15 98.4
6 GLM 5.2 84 89 96 94 99 88 34 93.7
7 MiMo V2.5 Pro 84 96 96 89 99 85 40 91.3
8 DeepSeek V4 Flash 84 99 93 92 99 78 40 88.9
9 Hunyuan 3 82 96 93 94 99 75 30 88.9
10 Qwen 3.7 Max 81 73 99 94 99 90 25 95.2
11 Claude Sonnet 5 81 56 99 89 99 85 53 92.1
12 Gemini 3.5 Flash 80 86 96 99 99 92 3 96.8
13 Kimi 2.7 72 50 96 94 50 85 48 88.8
14 Step 3.7 Flash 68 76 70 86 99 88 4 84.2
15 MiniMax M3 62 43 84 84 50 76 26 79.3
16 MiMo V2.5 62 0 67 89 99 94 31 86.5
17 LongCat 2.0 46 63 46 72 50 82 0 67.4

‡ Gemini 3.6 Flash is the new entrant (not in the Run #20 field); see §4.4.

Ranking is by OVR mean, shared rank on exact ties, tie-break GRD→ADH→SYN→EFF→PRC. JDG (jury) is not in OVR and was not run (objective-only, per the Run #11 jury-instability finding).

4.3 Objective saturates — the capability view

Read the objective board on its own and it is, once again, a photo finish:

🥇 Grok 4.598.4 · OVR #5
Gemini 3.6 Flash96.8 · OVR #4 · new
Gemini 3.5 Flash96.8 · OVR #12
GPT-5.6 Terra96.0 · OVR #1
GPT-5.6 Luna95.2
Qwen 3.7 Max95.2
GLM 5.293.7
DeepSeek V4 Pro92.8
Claude Sonnet 592.1
MiMo V2.5 Pro91.3
DeepSeek V4 Flash88.9
Hunyuan 388.9
Kimi 2.788.8
MiMo V2.586.5
Step 3.7 Flash84.2
MiniMax M379.3
LongCat 2.067.4
Objective score — the capability view (saturated at the top) · axis 0–100 obj/100

Nine models fall between 92.1 and 98.4; fourteen clear 86. On the raw manifest, Grok 4.5 (98.4, 62 of 63 checks) heads a field so tight that the top eight are separated by 5.6 points. This is the saturation Run #20 reported, reproduced on an entirely different task: the grounding, adherence, and synthesis checks are hard enough to be meaningful — the live-reachable ratio grounds and the fabricated-family trap are not gimmes — but the strong models clear them. Capability alone no longer orders the field. Note the two Gemini flashes post the identical 96.8 objective yet sit eight OVR ranks apart (#4 vs #12) — the sharpest illustration on the board that the objective axis cannot, by itself, separate a deployable model from a ruinous one.

The · OVR #5 / · OVR #4 / · OVR #1 annotations are the tell: the objective leader is not the operator leader. The next section is why.

4.4 The inversion — Grok and Gemini, capable and undeployable

The efficiency stat fans the saturated field back out, from 0 to 84 — by far the widest spread of the six stats, and the primary driver of OVR:

GPT-5.6 Terra84 · 37 calls
GPT-5.6 Luna59 · 48
Claude Sonnet 553 · 50
DeepSeek V4 Pro52 · 50
Kimi 2.748 · 50
Gemini 3.6 Flash44 · 54 · new · obj #2
MiMo V2.5 Pro40 · 55
DeepSeek V4 Flash40 · 54
GLM 5.234 · 58
MiMo V2.531 · 72
Hunyuan 330 · 59
MiniMax M326 · 58
Qwen 3.7 Max25 · 61
Grok 4.515 · 64 · obj #1
Step 3.7 Flash4 · 70
Gemini 3.5 Flash3 · 82
LongCat 2.00 · 186 calls
Efficiency (EFF) — the discriminator · par 35, zero at 70 calls · axis 0–100 EFF

Grok 4.5 is the best analyst in the field and the fourth-best operator. Its two trials scored 98.4 objective — identically (67 and 62 tool calls), so its grounding (99), adherence (96), and consistency (96) are all near-perfect. On a single-shot benchmark it is the model you would pick. But 64.5 tool calls against a par of 35 is a golf-EFF of 15, and OVR drops it to #4. It is capable, steady, and expensive — the profile you do not run on a cron.

Gemini 3.5 Flash is the same lesson, sharper: objective #2 (96.8), OVR #12 (EFF 3). It reached one 98.4 trial — tying Grok's best single score — but fired 68 then 95 tool calls to do it (mean 82). A model that reaches an elite answer by firing 2.3× a competent expert's tool calls, and that doubles its own call count between two runs of the same task, is a ruinous long-run executor no matter how good the answer. A blended objective score would have placed both models near the top and told a desk lead to deploy them. The card says: capable, do not automate. That corrected recommendation, now demonstrated on a second workflow, is the return on the whole instrument.

Gemini 3.6 Flash is the same lesson resolved — a controlled generational A/B. The new flash entrant posts the identical 96.8 objective as its 3.5 predecessor (both tie for second in the field), but shoots 54 tool calls, not 82 (trials 63 and 46), lifting EFF from 3 to 44 and OVR from #12 to #4 — an eight-rank jump bought entirely with efficiency, even though its consistency is slightly worse (CON 66 vs 86, from the wider 63/46 call split). Hold capability fixed at the elite ceiling, and the card resolves the two generations into "do not automate" and "deployable" purely on execution cost. It is the cleanest demonstration in the whole board that the efficiency axis measures a real, improvable property — and that a capability-only leaderboard, on which these two models are indistinguishable at 96.8, would be blind to precisely the difference a desk lead cares about.

At the extreme sits LongCat 2.0: 185 tool calls per trial, EFF 0, the lowest objective (67.4), OVR last. Where Grok over-executes a correct plan, LongCat flails — the two are different pathologies the card scores at opposite ends.

4.5 Consistency — the ceiling you can't deploy

Claude Sonnet 5 (OVR 81, rank 10) has a top-tier card underneath a variance problem. Both trials scored objective 92.1 — but one ran 38 tool calls and the other 63, and that effort dispersion drops CON to 56, marking its OVR down from a base of 88 to 81. Its correctness is elite and identical run to run; its execution cost is not, and a long-run deployment pays for the bad days. This is precisely the Sonnet 5 finding from Run #20 (elite ceiling, undeployable variance), reproduced on the trader workflow — a model property, not a workflow artifact.

The MiMo pair is again the clean controlled comparison. Same family, same tier split, opposite reliability:

The base variant's good trial (95.2 / 47) is genuinely strong; its other trial is a different model. The card prices that unpredictability exactly where a capability board would have shown a respectable 86.5 and hidden it — the same verdict the pair earned in Run #20.

4.6 What the stats say about the field


5. Cross-workflow read — what replicated

The point of Run #33 is the comparison to Run #20. Holding the instrument and field fixed and swapping the task, here is what carried over and what did not. (This section concerns the shared sixteen models present in both runs; Gemini 3.6 Flash, added only in Run #33, has no Run #20 counterpart and is excluded from the cross-run claims below.)

Replicated (properties of the models as operators):

Differed (task-specific detail):

The conclusion Run #20 could only assert for one workflow now has a second data point: for an autonomous desk operator, "best" means best operator — and the model that earns that title (Terra) does so across tasks, while the models that win raw capability (Grok, Gemini) lose it to efficiency across tasks too.


6. Infra-vs-ability — in place, not exercised

Run #20's central discipline — never let a gateway failure masquerade as a model failure, decided by the objective signature (short run + high errors = an infra death to replace; full run + weak score = a genuine result to keep) — was active for Run #33. This run it was not needed: all seventeen models produced two full, scored runs; none died mid-workflow, so no trial was replaced.

The discipline still did work at the keep end. Several trials ran heavy and error-prone — MiMo V2.5 Pro's second trial logged 12 errors, LongCat's logged 9, base MiMo's first trial ran 97 tool calls — and it would have been easy to re-roll them to prettier numbers. But these were full runs: the workflow executed end-to-end, the tool counts were real, the errors were retry churn, not a dead pipe. LongCat's 185 calls per trial is a genuine model result, not an artifact — kept, and it earns the last-place OVR honestly. A bench that re-rolls every ugly trial measures its own patience; Run #33 kept every full run and replaced none.


7. Limitations & threats to validity


8. Reproducibility


Appendix A — per-trial detail

Per-trial objective score and tool-call count (par = 35; EFF reaches 0 at 70 calls):

Model trial 1 (obj / calls) trial 2 (obj / calls) Obj mean EFF
GPT-5.6 Terra 96.8 / 30 95.2 / 44 96.0 84
GPT-5.6 Luna 95.2 / 51 95.2 / 46 95.2 59
DeepSeek V4 Pro 95.2 / 55 90.5 / 46 92.8 52
Gemini 3.6 Flash ‡ 95.2 / 63 98.4 / 46 96.8 44
Grok 4.5 98.4 / 67 98.4 / 62 98.4 15
GLM 5.2 93.7 / 61 93.7 / 54 93.7 34
MiMo V2.5 Pro 90.5 / 58 92.1 / 52 91.3 40
DeepSeek V4 Flash 93.7 / 65 84.1 / 43 88.9 40
Hunyuan 3 87.3 / 56 90.5 / 62 88.9 30
Qwen 3.7 Max 95.2 / 52 95.2 / 70 95.2 25
Claude Sonnet 5 92.1 / 63 92.1 / 38 92.1 53
Gemini 3.5 Flash 98.4 / 68 95.2 / 95 96.8 3
Kimi 2.7 82.5 / 42 95.2 / 59 88.8 48
Step 3.7 Flash 87.3 / 74 81.0 / 66 84.2 4
MiniMax M3 71.4 / 53 87.3 / 63 79.3 26
MiMo V2.5 77.8 / 97 95.2 / 47 86.5 31
LongCat 2.0 60.3 / 190 74.6 / 181 67.4 0

‡ Gemini 3.6 Flash — new flash entrant added to Run #33 after the original sixteen-model board; not part of the Run #20 field (§3, §4.4).

Run #33 · OTC Desk Agent Arena · generated 2026-07-20; Gemini 3.6 Flash added 2026-07-23.

← Board governance review: the ranking signal migratesMarkdown · PDFModel Ability Cards: measuring the long-run operator →