DeepSeek-V4-Flash-0731 Goes GA: Responses API Is Flash-Only and the Concurrency Gap Just Became a Routing Decision

DeepSeek's July 31 changelog entry does not announce a new model — it announces a routing inflection point. The deepseek-v4-flash API alias now points to V4-Flash-0731, a re-post-trained checkpoint that clears every major agent benchmark above V4-Pro-Preview. At the same time, DeepSeek made the Responses API format exclusive to Flash, leaving Pro without it until early August. For teams using Codex-native workflows or OpenAI's Responses API shape, there is currently no Pro fallback. That is the structural fact every routing configuration must absorb today.

What happened

DeepSeek published the V4-Flash-0731 public beta in their official API changelog. The update:

• Promotes deepseek-v4-flash to the new checkpoint without requiring any model-specifier change in existing integrations
• Confirms the same 284B/13B MoE architecture as the preview; only post-training changed
• Declares Responses API support exclusive to deepseek-v4-flash — the pricing table footnote states Pro support is coming in early August 2026
• Notes that V4-Flash-0731 is "specifically adapted for Codex" with documentation at the Codex integration guide
• Posts agent benchmark results far exceeding V4-Pro-Preview across every public agentic test set

The benchmarks are sharp. On tasks that matter to coding-agent and autonomous-operator workflows — Terminal Bench 2.1 (82.7), NL2Repo (54.2), DeepSWE (54.4), Toolathlon verified (70.3), Cybergym (76.7), and DSBench-Hard (59.6) — Flash-0731 leads V4-Pro-Preview by material margins. DeepSeek's own note: results used the DeepSeek Harness minimal mode at max effort, topp=0.95, temperature=1.0.

Why this matters for AI engineering teams

The agent benchmark reversal is the first surprise. V4-Flash was positioned as the cost-efficient tier, roughly equal to Pro on simple agent tasks; it now leads Pro on hard coding-agent tasks. If your routing logic defaults Pro as the "quality" tier for code generation or agent pipelines, the assumption needs to be tested against current numbers.

The more immediate operational fact is the Responses API split. DeepSeek's pricing page makes the current state explicit:

The 5× concurrency differential between Flash and Pro (2500 vs 500 concurrent requests) was already significant for burst-heavy agent workflows. Combined with Responses API exclusivity and 3× lower pricing, Flash is now the dominant tier for any workflow that accepts the quality tradeoff — and the benchmark data suggests that tradeoff is narrower than it appeared in April.

The third structural change: DeepSeek announced incoming peak/off-peak pricing — 2× regular rates during peak hours (09:00–12:00 and 14:00–18:00 Beijing time, UTC+8). An exact effective date has not been announced. Teams running high-volume inference during those windows should model the billing impact before it becomes a budget incident.

The router/operator angle

Routing implication 1 — Responses API lock-in window. If your agent harness uses the Responses API format — which an increasing number of Codex-native and tool-loop workflows do — the only available DeepSeek endpoint today is Flash. Routing to Pro will return an error for Responses API requests until early August. Fallback chains that treat Flash and Pro as interchangeable for this request shape will fail silently if Pro is in the first position.

Routing implication 2 — Codex native harness. DeepSeek explicitly adapted V4-Flash-0731 for Codex. The integration path in their docs exposes configuration parameters that the preview checkpoint did not formally support. If you are routing DeepSeek through a Codex-compatible harness, the -0731 checkpoint is the first version with a documented, stable integration path — not just a compatible one.

Routing implication 3 — Concurrency headroom vs Pro. 2500 concurrent requests for Flash versus 500 for Pro is a 5× burst headroom difference. For agent swarms that fan out many parallel tasks — code review loops, multi-file diff generation, agentic search — Flash is the natural tier under burst conditions even before considering cost. Pro's lower concurrency limit creates a hard capacity ceiling that no amount of fallback routing can abstract away.

Routing implication 4 — Peak-hour cost planning. The announced 2× peak-hour multiplier converts Flash from $0.28/1M output to $0.56/1M during Beijing business hours. That is still lower than Pro's off-peak rate of $0.87/1M, but the gap narrows. Teams with globally distributed users who absorb Beijing-hour load should model this into provider cost forecasts now, before the effective date is announced. Building time-of-day routing logic (route to a different provider during Beijing peak hours) is worth evaluating while there is still lead time.

Routing implication 5 — Migration path alignment. V4-Flash-0731 retains the same API interface as V4-Flash-Preview. There is no required migration for existing integrations calling deepseek-v4-flash. The checkpoint upgrade is transparent. Teams that pinned -0731 explicitly in their model specifier for the preview should verify whether they need to remove the suffix or if the alias absorbs it — the changelog indicates the alias is stable.

What to watch

The Pro GA announcement is the next event to track. DeepSeek's changelog forecasts V4-Pro GA "soon," which will likely include Responses API support and potentially updated benchmarks. At that point the current asymmetry resolves, and the routing decision shifts to pure cost-vs-quality optimization rather than API format availability. Until then, Responses API workflows must route through Flash exclusively.

Teams evaluating the incoming peak/off-peak pricing should watch the official DeepSeek announcements channel and price table for an effective date. The window to adjust routing time-of-day policy before the change hits is open now.

For anyone integrating via AI gateways — including TheRouter — the Flash concurrency expansion (2500 concurrent) means upstream burst headroom is substantially higher than the Pro tier. Gateway-side rate limit policies set against V4-Flash should be revisited to reflect the actual provider headroom.