Aug 6, 2026

DeepSeek V4-Flash-0731: The Silent Update That Beats V4-Pro on Nine Agent Benchmarks at $0.14/M Tokens

By AICoderScope Team · 11 min read

deepseekcursorclineclaude-codeapilocal-llmcostbenchmark

TL;DR: DeepSeek retrained V4-Flash and shipped the result on July 31, 2026 as V4-Flash-0731 — same 284B/13B MoE architecture, same $0.14/$0.28 per-million pricing, but it now outscores DeepSeek’s own V4-Pro-Preview on all nine agent benchmarks the company published. If you already call deepseek-v4-flash, you got the upgrade automatically. The catch: every one of those numbers came from DeepSeek’s own harness.

	V4-Flash-0731	DeepSeek V4-Pro	Claude Sonnet 5
Best for	Agentic loops in Cline, Claude Code, Codex	Long multi-step reasoning (a shrinking niche)	Max instruction fidelity, vision
Input / Output per 1M	$0.14 / $0.28	$0.435 / $0.87 (valley, 2× at peak)	$2 / $10 intro through Aug 31, then $3 / $15
Terminal-Bench 2.1	82.7 (vendor harness)	72.1 (Pro-Preview)	—
Context	1M tokens	1M tokens	1M tokens
Open weights	MIT, on Hugging Face	MIT	No
The catch	Vendor-run benchmarks, 2 of 9 in-house	3× Flash’s price for less agentic skill	14–36× pricier per token

Honest take: If you route DeepSeek traffic between Flash and Pro, this update ends the debate — send the agentic work to Flash and let your Pro spend wither. But treat the headline 82.7 as DeepSeek’s number, not an independent one, until third-party harnesses reproduce it.

What actually shipped on July 31

DeepSeek took V4-Flash out of preview on July 31, 2026 under the build name DeepSeek-V4-Flash-0731. Nothing about the model’s shape changed: it is still a Mixture-of-Experts design with 284 billion total parameters and 13 billion active per token, a 1-million-token context window, and up to 384K output tokens. What changed is post-training — the phase that shapes how a model applies what it already knows. DeepSeek reran that phase with a heavy emphasis on multi-step agent behavior: tool calling, terminal work, long task chains.

The economics did not move. Input stays at $0.14 per million tokens on a cache miss, $0.0028 on a cache hit, output at $0.28 per million, with a stated concurrency ceiling of 2,500 parallel requests. That pricing was already the reason we recommended Flash as a Cline backend back in June; now the same money buys a substantially more capable agent.

The weights shipped the same day. deepseek-ai/DeepSeek-V4-Flash-0731 is on Hugging Face under an MIT license — the repository weighs in at 304B parameters because it bundles DSpark, a speculative-decoding draft module, alongside the 284B model itself.

Nine benchmarks, one asterisk

Here is the full set DeepSeek published, with the two comparison points that matter — the Flash preview build it replaces and the larger V4-Pro-Preview it embarrasses:

Benchmark	Flash preview	V4-Pro-Preview	V4-Flash-0731
Terminal-Bench 2.1	61.8	72.1	82.7
Cybergym	—	—	76.7
Toolathlon (verified)	—	—	70.3
DSBench-FullStack*	—	37.0	68.7
DSBench-Hard*	—	—	59.6
DeepSWE	7.3	12.8	54.4
NL2Repo	—	—	54.2
Agent Last Exam	—	—	25.2
Automation Bench (Public)	—	—	25.1

*DSBench-FullStack and DSBench-Hard are DeepSeek’s in-house suites.

The margins over V4-Pro-Preview run from 8.7 to 41.6 points across the nine, per DeepSeek’s release notes. The DeepSWE jump is the wildest line in the table: 7.3 to 54.4 in one post-training pass, on a benchmark where the 1.6-trillion-parameter Pro-Preview managed 12.8.

Three caveats before you re-plan your stack around this table:

DeepSeek ran the evaluations itself, using DeepSeek Harness in minimal mode at the max tier, with top_p 0.95 and temperature 1.0. Vendor harnesses flatter vendors. When Terminal-Bench’s own leaderboard picks up 0731 under a neutral harness, expect the number to move — our Terminal-Bench 2.1 analysis covers why harness choice swings scores by whole points.
Two of the nine suites are internal. DSBench-FullStack and DSBench-Hard have no external baselines, so 68.7 and 59.6 are unfalsifiable until someone else can run them.
The comparison target is V4-Pro-Preview, not the GA Pro. DeepSeek has not published the same nine numbers for the production Pro build. The gap is probably directionally real — Pro’s post-training predates this agentic pass — but “beats its own flagship” comes with that footnote.

What survives the caveats: the preview-to-0731 deltas are enormous and consistent across benchmarks DeepSeek does not control (Terminal-Bench, DeepSWE), and the price stayed flat. Even discounted for vendor optimism, that is the best capability-per-dollar move any coding backend made in July.

If you already use Flash, you have already been upgraded

DeepSeek swapped the model behind the existing API name. Same endpoint, same key, same deepseek-v4-flash identifier — the 0731 build is what answers now. There is no config to touch, which also means there is no way to pin the old preview build on the official API. If your Cline prompts were tuned around the preview’s weaker tool calling (retry scaffolding, “one tool per turn” instructions), those workarounds are now dead weight you can delete.

You can confirm what you’re talking to in one call:

$ curl -s https://api.deepseek.com/models \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" | grep flash

{"id": "deepseek-v4-flash", "object": "model", "owned_by": "deepseek"}

The name won’t say 0731 — the build designation only appears in the release notes and on the Hugging Face repo. Any request against deepseek-v4-flash from July 31 onward hits the new build.

Cursor and Cline wiring is unchanged from our June setup guide: Cursor takes it as an OpenAI-compatible override (base URL https://api.deepseek.com, model deepseek-v4-flash), Cline has a native DeepSeek provider. The June guide’s warning still applies — leave thinking mode off in Cline, and remember Cursor’s Tab autocomplete never routes through custom backends.

New in 0731: it speaks Codex and Claude Code natively

The quieter half of this release is protocol support. Alongside the retrained model, DeepSeek shipped Responses API support — OpenAI’s newer request format, which DeepSeek describes as adapted specifically for Codex — and it continues to serve the Anthropic message format it has offered since V3.1. That means the two big terminal agents now talk to Flash without a translation proxy in the middle.

For Claude Code, point the Anthropic-format endpoint at DeepSeek:

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=sk-your-deepseek-key
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-flash
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash

Launch claude and every request — main loop and subagents — bills at DeepSeek rates. DeepSeek publishes official integration references for Claude Code, Codex, Cline, OpenCode, and GitHub Copilot Chat, so the env-var names above are the vendor’s own, not community reverse-engineering.

Should you? Terminal-Bench 82.7 is exactly the benchmark that predicts terminal-agent competence, and it is the score that made this release news. A Claude Code session that burns 500K input / 60K output tokens costs about $0.09 on Flash versus roughly $1.60 on Sonnet 5’s intro pricing. For unattended bulk work — test generation, lint sweeps, migration grinds — that 18× gap is the whole argument. For interactive pairing where one wrong edit costs you more time than the tokens saved, the frontier models still earn their premium.

The routing advice from July just flipped

Our V4 Pro analysis three weeks ago landed on a split: route routine work to Flash, reserve Pro ($0.435/$0.87 valley, double at peak) for multi-step agentic tasks. The 0731 numbers invert the second half. On every published agent benchmark, the cheap model is now the better agent — DeepSWE 54.4 vs 12.8 isn’t a gap you route around, it’s a different class.

What’s left for Pro is the work that leans on raw model depth rather than agentic reflexes: dense single-shot reasoning, subtle spec interpretation, the SWE-bench-Verified-style single-patch tasks where Pro’s 80.6% still leads Flash’s 79.0%. That is a real but narrow lane, and it no longer justifies a default. New default: everything goes to Flash at $0.14/$0.28, and you escalate to Pro (in the valley window, never at peak) only when Flash demonstrably fails a task twice.

A month of heavy agentic use — say 40M input / 6M output tokens — now prices out at roughly $7.30 on Flash, against ~$22.60 on Pro valley rates and ~$140 on Sonnet 5 intro pricing. Numbers like that are why the inference margin collapse keeps being the story of 2026.

Running it locally: MIT weights, honest RAM math

The 0731 weights hit Hugging Face the day of release, and Unsloth had GGUF quantizations up fast. The sizes are the reality check:

Quant	File size	Needs (RAM+VRAM)
UD-Q8_K_XL (near-lossless)	162 GB	~170 GB
UD-Q4_K_XL	155 GB	~168 GB
UD-IQ3_XXS (Unsloth’s recommended)	103 GB	~110 GB
1-bit tier	~85 GB	~96 GB

Because only 13B parameters activate per token, a MoE this size runs tolerably on high-RAM machines with llama.cpp offloading — a 128GB Mac Studio or a 96GB-VRAM workstation clears the 3-bit and 1-bit tiers, and early Hugging Face discussion reports usable agentic behavior even at IQ1_S. That is a different proposition from June, when we wrote Flash off as “enterprise-grade multi-GPU only.” It is still not a 24GB-GPU model and never will be; for what actually fits a single consumer card, see our Mellum 2 and Laguna XS 2.1 writeups, and the hardware-tier breakdown at runaihome.com.

The API remains the sane way to consume this model. At $0.14/M input you would need to push a genuinely absurd token volume before self-hosting 103GB of weights beats just paying DeepSeek — privacy and compliance, not cost, are the reasons to run it at home.

Where this leaves your stack

The pattern to watch: DeepSeek got these gains from post-training alone, on a 13B-active model, without touching price. Moonshot did something similar with Kimi K3’s verbosity retrain. The frontier labs’ pricing umbrella keeps getting punctured from below, and agentic capability — the thing that was supposed to justify premium tokens — is turning out to be the most post-trainable skill of all.

Concrete moves this week:

Already on Flash: delete your preview-era prompt workarounds, rerun the agentic tasks you’d written off, and check whether your Pro routing still earns its 3× premium.
On Sonnet 5 or GPT-5.6 for bulk agent work: run one real workload through Flash via the Anthropic endpoint before your next invoice. The quality floor moved.
Waiting for local: the 27B-class release to watch is Qwen3.8-27B, promised as open weights within days. Flash-0731 at home is a 110GB commitment.

FAQ

Do I need to change my model name to get the 0731 build? No. deepseek-v4-flash on the official API has served the 0731 build since July 31. There is no way to request the old preview build through the API.

Does V4-Flash-0731 beat the GA version of V4-Pro too? Unknown. DeepSeek’s nine-benchmark comparison targets V4-Pro-Preview. No equivalent numbers exist for the production Pro yet, though nothing suggests GA Pro closed a 41-point DeepSWE gap.

Is the 82.7 Terminal-Bench score trustworthy? It is DeepSeek’s own measurement using DeepSeek Harness (minimal mode, top_p 0.95, temperature 1.0). Treat it as an upper bound until the official Terminal-Bench leaderboard lists 0731 under a neutral harness.

Can I use it with Codex CLI? Yes — 0731 shipped with Responses API support that DeepSeek says is adapted for Codex, plus official integration docs. You point Codex at DeepSeek as a third-party model provider with your DeepSeek key.

What hardware runs it locally? Unsloth’s recommended 3-bit GGUF is 103GB and wants ~110GB of combined RAM+VRAM; the 1-bit tier squeezes into ~96GB. Nothing consumer-single-GPU. The MIT license permits all of it, including commercial serving.

Sources

Last updated August 6, 2026. Pricing and benchmark claims verified against the sources above on the day of writing; model pricing and capabilities change frequently — verify current state before committing budget.

Was this article helpful?