WATCH — DeepSeek V4-Pro official release: re-verify the planner's DEFAULT model (pricing, id, the tool_calls workaround, Responses support) when it ships
Trigger
DeepSeek publishes the official (non-preview) deepseek-v4-pro. Do not start this card before then — there is nothing to verify until it ships.
Where things stand (verified 2026-08-04)
deepseek-v4-pro is still DeepSeek-V4-Pro-Preview. The 0731 changelog entry is explicit: "This update only upgrades the DeepSeek-V4-Flash API. The DeepSeek-V4-Pro API and the APP/WEB models are unchanged" — closing with "The official release of DeepSeek-V4-Pro will follow soon." No date is committed. Third-party posts claiming "Pro in early August 2026" are not corroborated by DeepSeek's own docs.
Current preview specs, for the diff: 1.6T total / 49B active MoE, 1M context, 384K max output, $0.435 in (cache miss) / $0.003625 in (cache hit) / $0.87 out per MTok.
Why this one is not routine housekeeping
deepseek-v4-pro is motir-ai's default planner model — src/llm/gatewayClient.ts:23, default: 'deepseek-v4-pro'. Every plan Motir generates runs on it. Motir's primary product path is therefore pinned to a preview build today, and the release will change that build under us without an id change (that is exactly what Flash's 0731 did).
What to check, in order
- Model id. Flash's official build kept the id
deepseek-v4-flash— no code change. Confirm Pro does the same; if it introduces a dated or renamed id,relay/adaptor/deepseek/constants.goand the pin atgatewayClient.ts:23both move. - Pricing. Re-read api-docs.deepseek.com/quick_start/pricing against
relay/billing/ratio/model.go—"deepseek-v4-pro": 0.435 * MILLI_USD(input, cache-miss) and the output multiplier0.87 / 0.435. These are hand-curated entries that deliberately override the OpenRouter catalog (PR #10); if DeepSeek moves the numbers, the curated entries must move with them, and motir-ai'sModelCreditRateseed (aligned to the gateway catalog by MOTIR-1324) has to be re-checked in the same pass or the ledger silently drifts. - The tool_calls reliability workaround.
gatewayClient.ts:251documents a V4-Pro caveat: the model lands tool-calls in non-thinking mode and can emit a reasoning-only turn with notool_callsfield (deepseek-ai/DeepSeek-V3#1244), and the planner loop reads around it. Flash's 0731 build was re-post-trained specifically for agentic/tool-calling behaviour. If Pro gets the same treatment, re-test this caveat — the workaround may be removable. Removing dead defensive code is the point; keeping it after the underlying bug is fixed makes the loop harder to reason about forever. - Responses API + Codex adaptation. Flash's official build added both. If Pro's does too, say so on MOTIR-2114 — it changes that decision's inputs (today Pro cannot be the first Responses consumer because it does not speak it).
- Benchmarks vs Flash. The 0731 Flash build already beat V4-Pro-Preview on agent benchmarks (Terminal Bench 2.1: 82.7 vs 72.1). If official Pro does not clearly reclaim the lead for planning-shaped work, the "Pro is the default, Flash is the cheap lane" split in
gatewayClient.ts:23-24deserves a fresh look — Flash is ~3× cheaper on input and ~3× on output.
Deliverable
A short written verdict on this card: what changed, what did not, and which follow-on cards are needed. Any code consequence is a separate card per repo (one subtask = one repo = one PR) — this card verifies and reports, it does not ship the change. Expected follow-ups, if triggered: a motir-gateway pricing/id card, a motir-ai workaround-removal card, and possibly a model-choice decision.
Acceptance criteria
- The verdict names the exact DeepSeek changelog entry (date + build identifier) it is based on — not a third-party summary.
- Gateway pricing is either confirmed unchanged against the live pricing page, or a follow-up card exists with the new numbers.
- The
gatewayClient.ts:251caveat is either re-confirmed as still necessary (with the observation that showed it), or a removal card exists. - MOTIR-2114 is updated with Pro's Responses status either way.
- If nothing needs changing, that is recorded explicitly and the card is closed — a "no action" result is a valid, useful outcome, not a reason to leave it open.
Raised 2026-08-04 from the V4-Flash 0731 release check, which found the gateway already correct for Flash but the Pro release still pending.