On July 31, 2026, DeepSeek quietly pushed an update to its budget model that should have made the AI infrastructure world stop and think.
The official deepseek-v4-flash API entered public beta - same model string, same endpoint, same architecture - but with a fresh post-training pass that lifted agent scores clean past V4-Pro-Preview.
The budget tier beat the flagship. On every published agent benchmark. Without changing a single weight.
That result is worth sitting with, because the implication runs deeper than a price comparison.
What DeepSeek V4 Flash actually changed on July 31
The 0731 update is not a new model. DeepSeek-V4-Flash-0731 keeps the same model architecture and size as V4-Flash-Preview: 284B total parameters with 13B active in a Mixture-of-Experts setup, and a 1M token context window.
Only the post-training was redone.
The benchmark numbers from that retrain are stark. The most dramatic jump is DeepSWE, which rose from 7.3 on the April preview to 54.4 - a 645% improvement on the same model. DSBench-FullStack nearly doubled from 37.0 to 68.7, and Terminal-Bench 2.1 went from 61.8 to 82.7.
DeepSeek reports the official V4-Flash substantially exceeding V4-Pro-Preview across nine agent benchmarks. That is not a narrow win on one eval - it is a sweep across terminal tasks, software engineering, full-stack dev, and tool use.
The caveat is real and worth stating plainly: DeepSeek says the official V4-Flash was evaluated using DeepSeek Harness minimal mode at the max tier, with top_p = 0.95 and temperature = 1.0. Those settings matter: agent scores are extremely sensitive to the harness, so treat them as vendor-reported until independently reproduced.
Two of the suites - DSBench-FullStack and DSBench-Hard - are DeepSeek's own internal sets, not third-party comparables. They are useful as a signal of intent, not as a leaderboard.
The third-party signals that do exist point the same direction. Flash 0731 trails Claude Opus-4.8 by just 0.5 points on Agent Last Exam (25.2 vs 25.7). That is a rounding error in practical terms.
V4 Flash vs V4 Pro: a concrete comparison
DeepSeek V4 is a dual-model release built on a Mixture-of-Experts architecture. Both models support a 1 million token context window with a maximum output of 384K tokens, and both are released under the MIT license.
| V4 Flash | V4 Pro | |
|---|---|---|
| Total parameters | 284B | 1.6T |
| Active per token | 13B | 49B |
| API input (cache miss) | $0.14/M | $0.435/M |
| API output | $0.28/M | $0.87/M |
| Cache-hit input | $0.0028/M | $0.003625/M |
| Self-host (INT4) | ~140-158 GB VRAM | ~2.4 TB VRAM |
| Terminal-Bench 2.1 | 82.7 | 72.1 (preview) |
| DeepSWE | 54.4 | 8% (audit) |
| SWE-bench Verified | - | 80.6% |
Sources: V4-Pro: 1.6T total parameters, 49B active, $0.435/M input, $0.87/M output. V4-Flash: 284B total, 13B active, $0.14/M input, $0.28/M output. Both default to 1M-token context with 384K max output.
The split is clearer than the raw numbers suggest. The honest read is that V4 Pro leads on coding-style benchmarks like LiveCodeBench and Codeforces and is essentially tied with Claude Opus 4.7 on SWE-bench Verified. But SWE-bench Verified measures a different capability than agentic benchmarks like Terminal-Bench or DeepSWE - it is closer to a single-shot patch task than a multi-step agent loop. V4-Pro's 8% DeepSWE result is anomalously low relative to peers and to its own SWE-bench Pro score. The audit's read is that the gap reflects a real long-horizon-agent capability difference rather than a verifier artifact.
In short: reach for Pro when you are solving bounded coding problems. Reach for Flash when your agent has to run for more than one turn.
Self-hosting Flash versus calling the API
For teams that cannot send data to DeepSeek's China-based servers - regulated industries, anything with a BAA requirement - self-hosting Flash is the realistic path. DeepSeek stores API prompts on servers in China, where Article 7 of China's National Intelligence Law compels organizations to cooperate with state intelligence work. DeepSeek also trains on API data by default, publishes no SOC 2 report, and offers no BAA.
V4-Flash at INT4 is the realistic entry point for self-hosting. It fits on a single H100, a pair of A100s, or a four-card RTX 4090 rig - which puts a frontier-class model within reach of a single well-equipped workstation.
V4-Pro is a different category: at FP8 it needs around 2.4 TB of VRAM, which means a multi-node cluster of roughly sixteen H100s.
The economics of going self-hosted are harder than they look at first. On production on-demand infrastructure, DeepSeek V4-Flash's crossover sits at roughly 1.45 billion tokens per day on a 4x H100 cluster. Below that threshold, the API is simply cheaper. Use context caching aggressively - constant system prompts drop input cost by roughly 97%. A V4-Flash cache hit costs $0.0028/M against the $0.14/M cache-miss rate. If your agent repeats the same system prompt and tool definitions on every call - and it almost certainly does - you are not really paying $0.14/M in practice.
Beagle in action#eng-platform, 10:41amWhat the post-training story means for agent routing
The larger point here is not really about DeepSeek specifically. The Flash 0731 update suggests DeepSeek found a training methodology that unlocks agentic capability without scaling up compute requirements. If a single post-training pass can add 645% on DeepSWE with no new hardware, then the models you are using today for agent work may be substantially undertrained for that task - regardless of who made them.
The benchmark hierarchy for agentic tasks and for single-turn coding tasks are not the same list. A model that ranks #2 on SWE-bench Verified can rank below a cheaper, smaller variant on Terminal-Bench. The larger V4 Pro set the ceiling with a score of 80.6% on SWE-bench Verified - the top open-weights score. But it is Flash that broke through, because it captures most of that agentic capability at a price that is on the Pareto frontier of performance and cost.
Teams that are currently routing all LLM work through a single frontier model are leaving something on the table - not just money, but capability. Flash 0731 is faster and cheaper than Pro, and it outperforms Pro on the tasks that actually define agents: multi-step planning, tool use, terminal commands, and long-horizon execution. A teammate like Beagle, running agent-assisted workflows in Slack, benefits directly from that kind of routing precision - cheap, fast, and capable where it counts.
The caveat about vendor benchmarks stands. Until the DeepSeek Harness is open-sourced and independent teams reproduce these numbers, treat the DeepSWE figures as directional. But directional at 645% improvement is still directional in a very clear direction.
DeepSeek V4 Flash agent benchmarks: common questions
Is DeepSeek V4 Flash actually better than V4 Pro for agents?
On the benchmarks DeepSeek published July 31, yes - Flash 0731 beat V4-Pro-Preview on all nine agent evals, including a 14.7-point gap on Terminal-Bench 2.1. V4 Pro still leads on SWE-bench Verified (80.6%), which measures single-patch coding more than multi-turn agent loops. Route accordingly.
What changed in the DeepSeek V4 Flash 0731 update?
Only the post-training. The architecture, parameter count (284B total, 13B active), and context window (1M tokens) are identical to the April preview. DeepSeek redid the post-training pipeline and published benchmark improvements across nine agent tasks. The weights are the same; the training alignment is not.
How much does DeepSeek V4 Flash cost per million tokens?
The published API rate is $0.14 per million input tokens on a cache miss and $0.28 per million output tokens. Cache hits drop the input rate to $0.0028/M - a 98% discount for repeated prefixes like system prompts and tool definitions. Verify the current rate at DeepSeek's official pricing page before budgeting.
Can I self-host DeepSeek V4 Flash?
Yes. Flash at INT4 quantization needs roughly 140-158 GB of VRAM, which fits on a single H100, two A100s, or four RTX 4090s. FP8 roughly triples the memory requirement. Self-hosting makes cost sense only above roughly 1.45 billion tokens per day on a 4x H100 cluster - below that, the API is cheaper, and the data-control question is the real reason to self-host.
Are DeepSeek V4 Flash's agent benchmarks independently verified?
Not yet. DeepSeek measured the July 31 scores using its own DeepSeek Harness, which has not been open-sourced. Third-party results on overlapping evals (like Agent Last Exam) are consistent with the trend. Treat the DeepSWE and DSBench figures as vendor-reported until external teams replicate them with the open harness.