On July 31, 2026, DeepSeek quietly swapped the model behind every deepseek-v4-flash API call. Same architecture. Same price. Same endpoint string. The new build - V4-Flash-0731 - now scores higher than DeepSeek's own V4-Pro-Preview on all nine agent benchmarks the company published, including a Terminal Bench 2.1 score of 82.7 and a DeepSWE score of 54.4.
DeepSeek's budget 284B model beat its own V4-Pro-Preview across agentic benchmarks - including a 645% DeepSWE jump - at the same $0.14 per million input tokens.
That last part is the thing worth sitting with. A model with 13B active parameters outscored a model with 49B active parameters on agentic work, and nobody changed the price or the architecture. The improvement came entirely from a fresh post-training run aimed at exactly the task class where agents fail: tool-calling loops, long-horizon planning, and self-correction.
If you are currently paying frontier prices because you assumed the budget tier couldn't handle autonomous work, this release is worth understanding in detail.
What DeepSeek V4-Flash-0731 actually changed
The V4-Flash-0731 checkpoint shipped on July 31, 2026 - the same 284B-parameter backbone, re-post-trained for agents, now outscoring the V4-Pro preview on every agentic benchmark DeepSeek publishes. What changed: a re-post-training pass focused on agentic capabilities, plus native Responses API support and explicit adaptation for Codex. What did not change: architecture, parameter count, context window, and pricing.
The model behind your deepseek-v4-flash API calls changed on July 31, 2026 - same name, same architecture, same price, and dramatically different agent benchmark scores.
Three specific things are new in 0731:
- Re-post-training for agentic tasks. The gains come from re-post-training, not a new design. The checkpoint ships with the DSpark speculative decoding module attached.
The DSpark paper reports 60-85% faster per-user generation on V4-Flash versus the MTP-1 baseline at matched aggregate throughput.
Native Responses API. On the API side,
deepseek-v4-flashnow natively supports the Responses API format and is adapted for Codex.Promoted weights on Hugging Face. Weights are out, same day: the DeepSeek-V4-Flash-0731 repository appeared on Hugging Face at 07:30 UTC, MIT-licensed, superseding the preview checkpoint.
For public-benchmark code-agent tasks, DeepSeek evaluated the official V4-Flash 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 benchmark suites are internal: DSBench-FullStack is DeepSeek's in-house full-stack development set and DSBench-Hard its hard coding-agent set. They are not third-party comparable - useful as a signal of intent, not as a leaderboard.
That's the honest version. The jump is real. Whether it holds on your specific tasks is a different question.
Why agent economics make the price gap matter more than it looks
The price difference between Flash and a frontier model is not 35x on one call. On agentic workloads, that multiplier compounds per step.
An AI agent planning a multi-step task sends its full accumulated context - including the system prompt, conversation history, and all intermediate results - to the language model at every step. By step 20, the agent has paid for the original context 20 times. Gartner estimates agentic models require 5-30x more tokens per task than chatbots as a result.
Agentic coding workflows average 1-3.5M tokens per task including retries and self-correction loops. At GPT-5.6 Sol's current output rate of $30/M tokens, a 3.5M-token coding task costs roughly $105 in output alone. At V4-Flash's $0.28/M output rate, the same task costs under $1. Even at a conservative 20-step agent loop on a shorter task, the math shifts fast.
The practical effect of cache pricing is large for agent workloads. A coding agent re-sends the same system prompt and tool definitions on every turn. Those tokens land as cache hits, so the marginal input cost of each turn drops toward the cache-hit rate.
That's the non-obvious lever most teams aren't pulling. If your system prompt and tool definitions are stable across turns - which they almost always are in a coding agent - you're paying $0.0028/M instead of $0.14/M on the bulk of your input tokens. Caching is automatic. If your prompt prefix matches a previous request, the API charges the cache-hit rate: $0.0028/M on V4-Flash versus $0.14/M cache miss - a 98% discount. System prompts and shared document context benefit the most. You do not need to enable caching; it happens by default on every request.
Beagle in action#eng-ops, 11:02amWhat the frontier price reversal means alongside this
Here's the context that makes the V4-Flash story sharper. Closed-frontier providers (Claude, ChatGPT, Gemini) all started between $7 and $30 per million tokens in November 2023, fell over two years, and reversed in 2026. On April 23, GPT-5.5 launched at $5/$30 per million tokens, doubling GPT-5.4 on both sides.
The cheapest API achieving GPT-4-class benchmark quality costs $0.14 per million input tokens as of July 29, 2026 (DeepSeek V4-Flash), versus $30.00 at GPT-4's March 2023 launch - a 214× reduction in 40 months. Meanwhile the frontier has moved back up. The frontier tier runs at $5.00 per million input tokens and $30.00 per million output tokens as of July 29, 2026, priced for GPT-5.6 Sol.
The spread that results is striking. The price spread across 130 tracked LLM APIs is approximately 2,625× - from Ministral 3 3B at the floor to o1-pro at the ceiling - as of July 24, 2026.
An agent is a loop: it reads context, calls tools, evaluates results, and iterates - often dozens of steps per task, each step carrying the accumulated context of everything before it. As enterprises move from chat assistants to agentic workflows, token consumption per task grows by one to two orders of magnitude. The bill that looked negligible in the pilot becomes material at production scale - not because anything went wrong, but because the architecture multiplied and nobody was watching the meter.
Uber disclosed in May 2026 that its AI budget for the full year was consumed in four months, with the company's COO publicly questioning whether the link between agent spend and consumer-facing output was visible enough to justify the rate.
The two dynamics - frontier prices reversing upward, agentic token consumption multiplying - collide exactly where most teams are building right now. A model that can handle agentic work at $0.14/M input is no longer a fallback. It's a structural cost decision.
What to watch and what to be careful about
Why does a 13B-active model beat the 49B-active V4-Pro preview on agentic benchmarks? Because the Pro preview has not been re-post-trained for this release, and the official Flash is a new post-training run aimed at exactly the agentic task class. Agentic competence is largely policy, and the evaluation is harness-conditional - treat the margin as real but workload-dependent.
A few practical notes for teams considering a migration:
- Test on your own harness first. The benchmarks use DeepSeek's own evaluation framework. Your agent's tool schemas, retry logic, and context structure will produce different results.
- Pro is still unretrained. The V4-Pro API and the app and web models were not updated.
DeepSeek says the V4-Pro official release "will follow soon," with Responses API and Codex support for Pro expected in early August 2026. When Pro gets the same treatment, the Flash advantage on benchmarks will likely narrow.
- Legacy aliases are gone.
DeepSeek retired the legacy
deepseek-chatanddeepseek-reasoneraliases after July 24, 2026. If you're still calling those identifiers, your requests are failing silently. - Peak-hour pricing is coming. DeepSeek has announced upcoming peak/off-peak pricing, with peak-hour rates set to double once an effective date is confirmed. Build budget headroom for this.
- Multi-agent routing beats single-model all-or-nothing. For multi-agent systems, use a hierarchical architecture: budget models for worker agents, frontier models only for the lead orchestrator. This can achieve 97.7% of full-frontier accuracy at ~61% of the cost.
A teammate like Beagle, routing Slack questions to the right model tier on each call rather than treating every lookup as a frontier task, is exactly the kind of setup that compounds these savings across a team over a month. The operational decision isn't Flash versus frontier - it's whether you've built the routing layer to use each at the right moment.
DeepSeek V4-Flash for agents: common questions
What benchmarks did V4-Flash-0731 actually improve on?
DeepSeek reports V4-Flash-0731 exceeds V4-Pro-Preview across nine agentic and coding benchmarks, posting Terminal Bench 2.1 at 82.7 and DeepSWE at 54.4. Two of the suites are internal DeepSeek sets, not third-party evaluations, so treat the margin as directionally real but reproduce it on your own workload before making routing decisions.
Why would re-post-training beat a bigger model on agent tasks?
Agentic performance is largely about policy - how a model decides to call tools, handle errors, and manage multi-step context - rather than raw parameter count. The Flash model was specifically re-post-trained on agentic task classes the V4-Pro-Preview hadn't been optimized for. A smaller, well-trained model can outperform a larger, less-tuned one on the specific task it was trained against.
How does DeepSeek V4-Flash cache pricing work for agents?
Cache hits on V4-Flash cost $0.0028 per million input tokens versus $0.14 per million on a cache miss - a 98% discount. Caching is automatic: when your prompt prefix (system prompt, tool definitions, shared context) matches a recent request, those tokens bill at the cache-hit rate. Agent loops that resend the same system prompt on every step benefit most. You don't need to enable it.
Is V4-Flash genuinely good enough to replace a frontier model for agentic work?
For many coding, extraction, and tool-calling tasks: probably yes, with verification. For multi-step reasoning requiring deep judgment on ambiguous inputs: likely not yet. The pragmatic approach is hierarchical routing - budget models like Flash for worker agents and repetitive steps, frontier models for orchestration decisions and edge cases where quality genuinely diverges.
What changed for developers migrating from the old DeepSeek API?
The deepseek-chat and deepseek-reasoner aliases were retired on July 24, 2026. Existing integrations using those identifiers will fail. Migrate to deepseek-v4-flash and deepseek-v4-pro explicitly. The endpoint is otherwise unchanged, and the 0731 build is served automatically when you call deepseek-v4-flash.