DeepSeek-V4-Flash-0731: Codex Support and $0.14/$0.28 Pricing

Chinese model providers keep shipping API upgrades that undercut Western frontier pricing while chasing agent-tooling compatibility. On July 31, 2026, DeepSeek pushed the pattern further: the DeepSeek-V4-Flash API went live in public beta with what DeepSeek calls a "massive upgrade" to agent capabilities, native support for OpenAI's Responses API format, and full Codex compatibility.

Update — August 3, 2026: OpenCode reported Flash processed 8 trillion tokens on August 1 (5T free + 3T Go). Coverage: DeepSeek Flash’s 8T-token day.

Update — August 8, 2026: ARC Prize independently verified this release at 89.0% on ARC-AGI-1 ($0.02/task) and 61.4% on ARC-AGI-2 ($0.04/task) at max reasoning effort. Full benchmark breakdown: DeepSeek V4 Flash 0731 on ARC-AGI.

This follows DeepSeek's mid-July 2026 official V4 launch, which introduced peak-hour pricing for both deepseek-v4-pro and deepseek-v4-flash. The July 31 release is narrower and specific: it only touches the Flash API tier, but it's the one DeepSeek is positioning as its agent-coding workhorse.

TL;DR

What Actually Shipped

DeepSeek's own announcement draws a careful line around scope. Three things happened on July 31:

1. DeepSeek-V4-Flash official API entered public beta.
2. Agent capabilities were "massively upgraded" — DeepSeek says benchmark scores now "far surpass" the earlier V4-Pro-Preview, without publishing a full third-party benchmark table alongside the tweet.
3. The API gained native Responses API format support and full Codex adaptation.

A same-day follow-up post narrowed the scope further: the model architecture and size are identical to the preview version — this is a capability/serving upgrade to the Flash API specifically, not a new model. DeepSeek-V4-Pro's API and the consumer App/Web products are explicitly unchanged. The official DeepSeek-V4-Pro release remains a separate, later milestone — worth watching alongside the peak-pricing mechanics explainx.ai broke down for the mid-July V4 launch.

Why Responses API + Codex Support Matters

Most agent harnesses and coding tools built around OpenAI's newer tooling — including Codex — assume a backend that speaks the Responses API shape, not just the older Chat Completions format. A model provider that only supports Chat Completions requires an adapter layer to plug into that tooling; one that speaks Responses API natively is a much closer drop-in.

DeepSeek publishing configuration steps specifically for Codex integration (api-docs.deepseek.com/quick_start/agent_integrations/codex) signals where DeepSeek expects developers to actually use V4-Flash: as a lower-cost backend swapped into existing Codex-based agent workflows, not just chat.

This mirrors a pattern explainx.ai has tracked across Chinese open-weight providers chasing coding-agent compatibility rather than chatbot market share — see how to run GLM 5.2 across agent harnesses and Cline's $9.99/mo bundled access to GLM-5.2 and DeepSeek. The strategy is consistent: undercut on price, match on tool-calling compatibility, and let harness authors do the rest.

Pricing: $0.14 / $0.28 per Million Tokens

Community figures circulating alongside the launch put V4-Flash-0731 at roughly $0.14 per million input tokens and $0.28 per million output tokens — described by developer Hassan (@nutlope) as "the best model intelligence vs cost out of any model," landing near GLM 5.2 and GPT Luna on capability while undercutting both on price.

Treat that specific framing as launch-day enthusiasm rather than an audited benchmark — it's one developer's post-launch read, not a controlled eval. DeepSeek's own pricing page is the source of truth, and this provider has changed pricing structure mid-cycle before: the mid-July V4 launch introduced 2× peak-hour multipliers (9:00–12:00 and 14:00–18:00 Beijing time) on top of baseline rates for both Pro and Flash tiers. Confirm whether V4-Flash-0731's quoted price is an off-peak baseline or a flat rate before budgeting production spend.

Independent Benchmark: Artificial Analysis

Within hours of launch, Artificial Analysis published an Intelligence, Performance, and Price analysis for V4-Flash-0731 — the first third-party read on DeepSeek's own claims, and it landed on Hacker News with 56+ points and active discussion.

The core finding: at max reasoning effort, V4-Flash-0731 costs roughly $0.03 per task at an intelligence index around 50 — putting it directly against OpenAI Luna, which needs high-to-xhigh effort ($0.03-$0.04) to hit index 46-49, or max effort ($0.07) to clear index 51. Read plainly by HN commenter spwa4: "OpenAI Luna [is] between 2x and 3x the price of DeepSeek Flash [for] what you get," while Luna runs 2-5x faster per token.

That speed-for-price tradeoff matters for how you'd actually deploy it:

• Throughput: ~93 tokens/sec on OpenRouter, per community reporting — usable for agent loops, not the fastest option if raw latency is the bottleneck.
• No multimodal support. HN commenter qtalen flagged this directly: V4-Flash still doesn't accept image or other non-text input, which caps its value versus a multimodal competitor like GPT 5.6 SOL for any workload that needs vision.
• Reasoning-length variance is model-specific, not a V4-Flash issue — a separate HN thread noted Artificial Analysis's "Output Tokens per Intelligence Index Task" metric can undercount how long some models (e.g. Kimi K3) reason relative to hands-on experience, a reminder that any single index number is a proxy, not ground truth.

The practical read: if your workload is cost-sensitive, text-only, and tolerant of a few hundred milliseconds more latency per token, V4-Flash-0731's price/intelligence ratio is now independently verified, not just vendor-claimed. If you need vision input or the fastest possible token stream, it's not (yet) the right fit.

What to Verify Before Switching Traffic

DeepSeek's announcement is a vendor tweet, not an independent benchmark suite. Before routing production agent traffic to V4-Flash-0731:

• Run your own eval set anyway. Artificial Analysis's index confirms the price/intelligence claim in aggregate, but it's still a proxy metric — test against your actual task distribution before switching production traffic.
• Confirm current pricing directly — Chinese-provider API pricing has shifted mid-quarter before (peak-hour multipliers being the most recent example).
• Check quota and rate limits for the beta tier — public beta releases commonly ship with tighter or less-documented rate limits than GA releases.
• Test the Responses API path specifically if you're integrating via Codex or a Responses-API-native harness — native support is new as of this release and worth a smoke test before depending on it.
• Rule out multimodal needs upfront — if your agent pipeline touches images or other non-text input at any step, V4-Flash isn't a fit yet.

Related Reading

• DeepSeek V4 Flash 0731 verified on ARC-AGI: 89% at $0.02/task
• DeepSeek Flash hit 8T tokens in a day — OpenCode Aug 1
• Musk: 99.99% AI compute in space — Calacanis quotes DeepSeek Flash
• DeepSeek V4 Official Release: Peak-Hour Pricing Explained
• DeepSeek DSpark V4: Speculative Decoding Guide
• How to Run GLM 5.2 Across Coding Plans and Agent Harnesses
• Cline's $9.99/mo GLM-5.2 Plan: Bundled Open-Weights Access
• Top Chinese AI Companies and Startups Guide
• GLM 5.2: Zhipu's China AI Response to the Fable 5 Ban

Sources: DeepSeek official announcement on X, DeepSeek API docs (api-docs.deepseek.com/quick_start/agent_integrations/codex), Artificial Analysis DeepSeek-V4-Flash analysis, Hacker News discussion, community commentary from Hassan (@nutlope) and Unsloth AI. Pricing and benchmark claims accurate as of July 31, 2026 — verify against official docs before production use.