DeepSeek-V4-Flash Enters Public Beta with Major Agentic Benchmark Improvements and Codex Integration

Models7 sources·Jul 31

benchmarks-and-evals agents coding-tools deepseek deepseek-v4-flash

Summary

• DeepSeek-V4-Flash API entered public beta July 31, 2026, with benchmark scores far exceeding V4-Pro-Preview across all nine agentic tasks.
• Top scores: Terminal Bench 82.7, Cybergym 76.7, DeepSWE 54.4, DSBench-FullStack 68.7, and NL2Repo 54.2 — all well ahead of the preview version.
• V4-Flash approaches Anthropic's Opus-4.8 on several benchmarks: Agents' Last Exam 25.2 vs 25.7, AutomationBench 25.1 vs 27.2, DSBench-FullStack 68.7 vs 71.6.
• New release adds native Responses API support, full Codex integration, and a speculative decoding module; all gains came from re-post-training alone with no architectural changes. API pricing adjustments are reportedly expected amid rising demand.

Adjust signal

Updates

Aug 7

Type	Key Point	Context
Industry Update	API Pricing Adjustments Reportedly Expected	Reports from social media sources indicate a significant API price increase for DeepSeek V4-Flash is expected amid rising demand. Details unconfirmed; no official announcement from DeepSeek.

Industry Update

API Pricing Adjustments Reportedly Expected

Reports from social media sources indicate a significant API price increase for DeepSeek V4-Flash is expected amid rising demand. Details unconfirmed; no official announcement from DeepSeek.

Details

Type	Key Point	Context
Product Launch	V4-Flash API Now in Public Beta	DeepSeek officially released the V4-Flash API into public beta on July 31, 2026, replacing its preview version under model ID deepseek-ai/DeepSeek-V4-Flash-0731.
Stat	Terminal Bench 2.1 Score: 82.7 (vs Opus-4.8's 85.0)	V4-Flash scored 82.7 on Terminal Bench 2.1, well ahead of V4-Pro-Preview (72.1) and approaching Anthropic's Opus-4.8 (85.0).
Stat	DeepSWE 54.4, Cybergym 76.7, NL2Repo 54.2	Additional agentic benchmarks show broad capability gains across coding, repository navigation, and tool-use tasks compared to both the preview version and V4-Pro-Preview.
Stat	Near-Parity with Opus-4.8 on Three Benchmarks	V4-Flash scores 25.2 vs Opus-4.8's 25.7 on Agents' Last Exam, 25.1 vs 27.2 on AutomationBench, and 68.7 vs 71.6 on DSBench-FullStack — despite being a far smaller model.
Tech Info	Same Architecture, Gains from Re-Post-Training Only	V4-Flash-0731 retains the exact same model architecture and parameter count as the preview; all benchmark improvements come from re-post-training, with no size increase.
Tech Info	Speculative Decoding Module Attached	V4-Flash shares its structure with DeepSeek-V4-Flash-DSpark, meaning it includes a speculative decoding module for faster inference out of the box.
Product Launch	Native Responses API + Full Codex Integration	V4-Flash now supports the Responses API natively and includes full Codex integration, making it directly usable in autonomous coding agent pipelines without additional configuration.
Strategy	Official V4-Pro API Release Announced as Coming Soon	DeepSeek signaled the forthcoming official release of the V4-Pro API, indicating a continued staged rollout strategy building toward a premium tier.

Product Launch

V4-Flash API Now in Public Beta

DeepSeek officially released the V4-Flash API into public beta on July 31, 2026, replacing its preview version under model ID deepseek-ai/DeepSeek-V4-Flash-0731.

Stat

Terminal Bench 2.1 Score: 82.7 (vs Opus-4.8's 85.0)

V4-Flash scored 82.7 on Terminal Bench 2.1, well ahead of V4-Pro-Preview (72.1) and approaching Anthropic's Opus-4.8 (85.0).

Stat

DeepSWE 54.4, Cybergym 76.7, NL2Repo 54.2

Additional agentic benchmarks show broad capability gains across coding, repository navigation, and tool-use tasks compared to both the preview version and V4-Pro-Preview.

Stat

Near-Parity with Opus-4.8 on Three Benchmarks

V4-Flash scores 25.2 vs Opus-4.8's 25.7 on Agents' Last Exam, 25.1 vs 27.2 on AutomationBench, and 68.7 vs 71.6 on DSBench-FullStack — despite being a far smaller model.

Tech Info

Same Architecture, Gains from Re-Post-Training Only

V4-Flash-0731 retains the exact same model architecture and parameter count as the preview; all benchmark improvements come from re-post-training, with no size increase.

Tech Info

Speculative Decoding Module Attached

V4-Flash shares its structure with DeepSeek-V4-Flash-DSpark, meaning it includes a speculative decoding module for faster inference out of the box.

Product Launch

Native Responses API + Full Codex Integration

V4-Flash now supports the Responses API natively and includes full Codex integration, making it directly usable in autonomous coding agent pipelines without additional configuration.

Strategy

Official V4-Pro API Release Announced as Coming Soon

DeepSeek signaled the forthcoming official release of the V4-Pro API, indicating a continued staged rollout strategy building toward a premium tier.

DeepSeek V4-Flash public beta release details, benchmark results vs Opus-4.8, new developer integrations, and reported pricing developments — July–August 2026

What This Means

DeepSeek-V4-Flash's public beta marks a meaningful shift in the agentic AI competitive landscape: a smaller, cost-efficient model is now approaching Anthropic's frontier Opus-4.8 on multiple autonomous agent benchmarks, including Agents' Last Exam, AutomationBench, and DSBench-FullStack. That these gains came purely from re-post-training — with no architectural changes — underscores how much competitive value labs can still extract from the optimization process alone. The addition of native Responses API support and full Codex integration makes V4-Flash immediately actionable in autonomous coding pipelines, accelerating developer adoption. With V4-Pro's official release still to come and pricing adjustments reportedly on the horizon amid rising demand, DeepSeek's trajectory signals continued pressure on both cost and capability benchmarks for all frontier model providers.

Sources

Update history (2)

Aug 7Linked speculative pricing article (Grok Sweep/Twitter source): reports indicate upcoming API price increase for DeepSeek V4-Flash amid rising demand. Source is unverified (t.co link, content unavailable). Added as row 9 in tier3 deep dive and noted in tier1 and what_this_means with explicit unconfirmed caveat. No architectural or benchmark content changed.

Aug 3Added Codex integration and native Responses API support (new from Alpha Signal). Added full benchmark comparison table vs Anthropic Opus-4.8 and GLM-5.2, revealing near-parity on Agents' Last Exam (25.2 vs 25.7), AutomationBench (25.1 vs 27.2), and DSBench-FullStack (68.7 vs 71.6). Added speculative decoding module detail. Updated title, tier1_scan, tier2_understand, tier3 (8 rows), key_facts, and what_this_means accordingly.

Similar Events

DeepSeek Releases V4.1 Flash with 4x KV Cache Compression for Long-Context Agent Workflows

Sep 17

DeepSeek V4 Flash (304B) Runs in Production on Single AMD MI300X at 168 tok/s

Aug 4