Jomoq-1 everywhere: frontier-grade chat in 1.4 s on 21× fewer tokens

Two weeks ago we shipped Jomoq-1 in the builder: a verified fast pass that drafts most code edits deterministically-checked in under a second, escalating to a frontier model only when verification fails. This post reports what happened when we applied the same architecture to chat, and the finalized numbers for both surfaces.

The design premise is unchanged: most turns don't need a frontier model: they need the right context assembled fast. In chat that means three things: classify the turn before any model call; fetch structured data (FRED economic series, search results) in parallel rather than letting the model discover it over sequential tool round-trips; and send mode-sized scaffolding: a research turn gets 4 tool schemas and a 600-token prompt, not the 21 KB monolith (9.1 KB system prompt + 23 tool schemas ≈ 4,000 tokens) that a conventional harness re-sends on every loop iteration.

Chat results

Seven tasks, comprising three live economic-data questions (multi-series FRED), two live-web research questions, and two conversational, run through four configurations. Three of them use what we call the conventional harness: Jomoq chat's own pre-Jomoq-1 pipeline (the full 9.1 KB system prompt and all 23 tool schemas on every loop iteration, sequential tool round-trips). In other words, we benchmark the new architecture against our own previous product, with three different models inside it. Every configuration executes real tools (live FRED and Tavily calls) and every reply is judged by Claude Sonnet 5 against ground truth fetched independently at judging time.

The headline case is the three-series question: "Pull the latest CPI, core CPI, and the Fed funds rate from FRED and compare each to a year ago." A conventional harness turns this into sequential tool round-trips; the frontier configuration spent 50 seconds and 62K tokens on it (and 159 seconds on the single-series Treasury question, where it spiralled through repeated searches). Jomoq-1 classifies the intent client-side, fetches all three series from FRED in parallel, and makes exactly one model call with the data attached: 1.4 s, 2,228 tokens, judged 10/10 against independently fetched ground truth, as were both other economic-data tasks.

Builder results (finalized)

The builder numbers below consolidate the full 24-task regression suite (text edits, styles, elements, refactors, generative sections, understanding-required edits on 1.3–2.1K-line files), judged by Claude Opus 5, plus live dogfooding traces on a real 2.1K-line project.

How this compares to the industry

The closest published work on dual-model inference reports a 1.25× realistic end-to-end speedup from transferring KV-cache state between a small and a large model, with 61–82% output agreement, restricted to open-weight model pairs. We took the opposite bet: operate at the task level, not the tensor level. When the fast lane ships, fidelity is 100% by construction: the output either passes deterministic verification or the turn escalates with full context. That trade is what turns 1.25× into the multiples above: 21× on brand renames, 45× on coordinated large-file edits, ~45× on app builds, 5.6× on frontier-versus-Jomoq chat latency (35× on the three-series economics question), because the win compounds: fewer model calls × smaller calls × a faster engine.

Against conventional assistant harnesses, the token economics compound the same way. The 21 KB scaffolding a typical tool-calling loop re-sends every iteration is invisible in a demo and dominant at scale; mode-sized context plus parallel prefetch is why a judged Jomoq-1 chat turn costs a median 850 tokens while the same engine in the pre-Jomoq harness costs 9,090: same model, same tools, 10.7× difference from architecture alone.

Method & honesty notes

• Chat: n=7 tasks × 4 configurations, single run, real live tools (FRED, Tavily). Judge: Claude Sonnet 5 with ground truth fetched independently at judging time for economic tasks (year-ago values fetched by date window, correct for daily and monthly series alike), and the run's actual tool results as the grounding reference otherwise. All 28 rows were scored. The judge shares a model family with the frontier configuration, a bias we note rather than pretend away; it would favor the frontier row, not ours.
• Live-web news tasks scored low (1–6) for every configuration: composition quality there is capped by search-snippet quality, not by harness. We report it rather than dropping the tasks.
• The frontier baseline runs in the conventional harness: that is the point of the comparison (architecture, not model quality), but it means these numbers say nothing about Sonnet 5 in an optimized harness.
• Builder: 24-task suite, judge Claude Opus 5; cost multiple computed on matched tasks against a full-rewrite baseline at list prices. The 2-second build is a judged 6/10 first draft, smaller than a frontier model's 90-second first shot; the product bet is that sub-second iteration beats a bigger first draft.
• Benchmark numbers come from runs of the checked-in suites (bench/run.mjs, bench/chat-bench.mjs) with results committed alongside. The three builder before/after traces are live dogfooding sessions recorded on 21 August 2026 (the "~90 s" baseline is approximate from the session log); they are illustrations of the suite's measured effect, not suite rows.

Jomoq-1 chat lanes ship behind the same verification philosophy as the builder: structured answers cite the series they were computed from, research turns carry their sources, and four independently-verified recovery layers keep free-tier models from ever ending a turn with leaked tool syntax. Try it: pick Jomoq 1 in the chat model menu.