DEEPSEEK V4 FLASH 0731 — CONFIRMED SPECS (JULY 31, 2026)
● Model string: deepseek-v4-flash (replaces deepseek-v4-flash-preview)
● Architecture: 284B MoE, 13B active per token — unchanged from preview
● Context window: 1M tokens
● What changed: Re-post-training on agent-focused data only — not a new architecture
● Pricing: $0.14/M input · $0.28/M output
● Terminal-Bench 2.1: 82.7% (up from 61.8% preview, up from 72.1% V4-Pro-Preview)
● NL2Repo: 54.2% (up from 39.4% preview)
● DeepSWE: 54.4% (up from 7.3% preview)
● Cybergym: 76.7% (up from 38.7% preview)
● API format: Responses API native. Codex-compatible (no adapter needed).
● Weights: MIT license — on Hugging Face at deepseek-ai/DeepSeek-V4-Flash-0731
● Deprecation: deepseek-chat and deepseek-reasoner legacy strings deprecated — end in 3 months (October 24, 2026)
● What did NOT change: V4-Pro API, app, and web models unchanged. V4-Pro official release "coming soon."
The Result That Matters: Small Model Beats Large Model on Agent Work
According to Digital Applied's analysis, the key finding in V4 Flash 0731 is not the absolute benchmark scores — it is the sequencing. DeepSeek chose to productionise the 284B/13B model first, while the 1.6T V4-Pro remains in preview. And in its own changelog, DeepSeek states that the 0731 checkpoint's agent scores "far exceed V4-Pro-Preview." A 284B model — the cheaper, faster, lower-VRAM option — outperforming the 1.6T flagship on the benchmarks DeepSeek cares most about is the result worth internalising. According to OfficeChai's benchmark breakdown, V4 Flash 0731 scores 82.7 on Terminal-Bench 2.1 — ahead of Z.AI's GLM-5.2 at 81.0 and within striking distance of Claude Opus 4.8's 85.0, at a fraction of Opus 4.8's price.
Benchmark Comparison — Full Agent Suite
| Benchmark | V4 Flash Preview | V4 Flash 0731 | V4 Pro Preview |
|---|
| Terminal-Bench 2.1 | 61.8% | 82.7% | 72.1% |
| NL2Repo | 39.4% | 54.2% | Not published |
| DeepSWE | 7.3% | 54.4% | Not published |
| Cybergym | 38.7% | 76.7% | Not published |
| Toolathlon-Verified | 49.7% | 70.3% | Not published |
All benchmark scores from DeepSeek's official model card and changelog. NL2Repo, DeepSWE, Cybergym, and Toolathlon are DeepSeek's own benchmark suite — the DeepSeek Harness used for evaluation will be released publicly to allow third-party verification. Per the HuggingFace model card, evaluations used minimal mode, max effort, temperature 1.0, top_p 0.95.
$0.14/M — The Price-Performance Case
At $0.14/M input and $0.28/M output, V4 Flash 0731 is the cheapest model in this analysis by a meaningful margin. According to AI/TLDR's pricing breakdown, that puts it at roughly a third of Laguna S 2.1's $0.10/M input (wait — actually Laguna is slightly cheaper on input at $0.10 vs DeepSeek's $0.14), but DeepSeek's output rate of $0.28/M is significantly cheaper than Laguna's $0.20/M. For agentic workloads that generate substantial output — tool calls, code, multi-step reasoning — DeepSeek V4 Flash 0731 is the cheapest capable option with verified Terminal-Bench 82.7% performance.
| Model | Input /1M | Output /1M | Terminal-Bench 2.1 |
|---|
| DeepSeek V4 Flash 0731 | $0.14 | $0.28 | 82.7% |
| Laguna S 2.1 | $0.10 | $0.20 | 70.2% |
| Grok 4.5 | $2.00 | $6.00 | 83.3% |
| OpenAI Codex (GPT-5.6) | $5.00 | $30.00 | 88.8% |
| Claude Opus 4.8 | $5.00 | $25.00 | 85.0% |
What Developers Need to Do Now
Deprecation deadline — October 24, 2026: Legacy API strings deepseek-chat and deepseek-reasoner will be discontinued in three months. Migrate to deepseek-v4-flash and deepseek-v4-pro now — do not wait until October.
OpenCode integration: According to Developers Digest, V4 Flash 0731 runs directly in OpenCode. Set model to deepseek-v4-flash in your OpenCode config.
Responses API and Codex compatibility: Existing harnesses built for OpenAI's Responses format work with V4 Flash 0731 without an adapter. No SDK changes required for teams already using OpenAI tooling.
Benchmark caveat: All benchmark scores are from DeepSeek's own evaluation suite using DeepSeek Harness (to be released publicly). The harness's "minimal mode, max effort" configuration is not directly comparable to how other labs run the same benchmarks. Wait for independent third-party verification on the DeepSWE and Cybergym numbers before routing production workloads based on those specific scores.
Sources: DeepSeek API changelog (official) · HuggingFace model card · OfficeChai benchmark breakdown · Digital Applied analysis · AI/TLDR · Related: DeepSeek vs Moonshot risk profiles → · Laguna S 2.1 review →