What changed on 31 July
DeepSeek pushed DeepSeek-V4-Flash-0731 into public beta on 31 July 2026. The model is a sparse Mixture-of-Experts with 284B total parameters and 13B active parameters per token, a context window of 1,048,576 tokens and a maximum output of 65,536 tokens. It sits in the V4 family alongside the Pro line, sharing that family's hybrid attention design — Compressed Sparse Attention paired with Heavily Compressed Attention, mHC connections and the Muon optimiser.
The most interesting engineering detail in the release note is what did not change. The 0731 build keeps the exact structure and size of V4-Flash-Preview. No new expert count, no re-shaped attention, no different tensor layout. DeepSeek redid the post-training and nothing else. For anyone who has already sized a serving cluster around Flash-Preview, that is a free upgrade: the weights drop into the same memory budget, the same tensor-parallel plan and the same kernel assumptions.
The headline claim is that Flash-0731 beats V4-Pro-Preview on all nine agent benchmarks DeepSeek reports, by margins of 8.7 to 41.6 points. A small model beating a big one from the same lab is a genuinely notable result, and it is also exactly the kind of claim that deserves the most scepticism, because it was measured entirely in-house.
Every score below was produced by DeepSeek on the DeepSeek Harness in minimal mode at the max tier, with top_p = 0.95 and temperature = 1.0. Agent benchmark scores are extremely harness-sensitive — the scaffold, retry policy, tool schema and step budget routinely move a result by more points than a model generation does. Until an independent party reproduces these on a neutral scaffold, treat all nine as vendor-reported marketing figures, not measurements you can plan a roadmap around.
The nine reported scores
Here is the full set, exactly as published. Read the caption before you read the column.
| Agent benchmark | Reported score | Status |
|---|---|---|
| Terminal Bench 2.1 | 82.7 | Vendor-reported |
| Cybergym | 76.7 | Vendor-reported |
| Toolathlon (verified) | 70.3 | Vendor-reported |
| DSBench-FullStack | 68.7 | Vendor-reported |
| DSBench-Hard | 59.6 | Vendor-reported |
| DeepSWE | 54.4 | Vendor-reported |
| NL2Repo | 54.2 | Vendor-reported |
| Agent Last Exam | 25.2 | Vendor-reported |
| Automation Bench (public) | 25.1 | Vendor-reported |
Two things stand out once you stop reading it as a leaderboard. First, the spread is enormous: 82.7 on Terminal Bench against 25.1 on Automation Bench. That is not a model being inconsistent, it is nine benchmarks measuring wildly different things at wildly different difficulty levels. A single average across them would be meaningless. Second, the benchmarks where Flash-0731 scores highest are the ones with the tightest, most mechanical task definitions — terminal command sequences, tool invocation — while the open-ended, long-horizon ones sit in the mid-twenties. That shape is consistent across the whole open-weight field, as we found when the open-weight leaderboard flipped between GLM, Kimi and DeepSeek earlier this year.
Why the harness caveat is the story, not a footnote
If you have never rebuilt an agent evaluation from scratch, the size of the harness effect is hard to believe. The same weights, on the same benchmark, will produce materially different numbers depending on how many steps the scaffold allows before it gives up, whether a failed tool call is retried or counted as a failure, how the file system is presented, whether the model can see its own previous stdout, and how the final answer is extracted. "Minimal mode at the max tier" tells you DeepSeek chose a lean scaffold with a generous compute allowance. Both choices are defensible. Neither is the scaffold you are running in production.
The temperature setting deserves a second look too. Running an agent benchmark at temperature = 1.0 is an unusual choice for a task class where determinism normally helps, and it means the reported figures are samples from a distribution rather than fixed values. Without variance bars or a stated number of runs, a 8.7-point margin over V4-Pro-Preview is not obviously outside the noise band on a benchmark with a few hundred tasks. The 41.6-point margin at the other end almost certainly is real in direction, if not in exact size.
None of this means DeepSeek is being dishonest. It means the numbers answer the question "how does this model perform inside DeepSeek's scaffold?" and you are asking "how does this model perform inside mine?" Those are different questions and only one of them affects your product.
Reproduce the claim on your own harness before you argue about it. Freeze your existing agent scaffold, swap only the model endpoint, and run your private task set twice — once at temperature = 1.0, top_p = 0.95 to match DeepSeek's settings, and once at your production sampling settings. The gap between those two runs is your harness sensitivity, and it is the single most useful number you will generate all week. Our walkthrough on benchmarking coding agents on your own repo has the full method, including how to build a task set from your own merged pull requests so it cannot leak into anyone's training data.
What 13B active at 1M context does to your bill
This is where the release is genuinely consequential, and where the maths does not depend on anyone's harness. A model that activates 13B parameters per token is cheap to serve relative to its capability class, and DeepSeek has priced it accordingly.
| Source | Input (cache hit) | Input (cache miss) | Output |
|---|---|---|---|
| DeepSeek official API | $0.0028 | $0.14 | $0.28 |
| OpenRouter listing | — | $0.09 | $0.18 |
Read the cache-hit column again: $0.0028 per million tokens. That is fifty times cheaper than the cache-miss input price on the same API. For an agent that carries a large stable prefix — a repository map, a coding standard, a product schema, a long system prompt — the caching behaviour is not an optimisation, it is the entire cost structure. Design your prompt so the stable portion sits first and never moves, and the 1M window costs almost nothing to re-read on every step.
The OpenRouter listing has no cache-hit tier in the same shape, and its headline input price is lower than DeepSeek's cache-miss price. If you are quoting a number in a board deck or a pricing model, attribute it to the specific source, because the two structures reward completely different prompt designs. A team optimising for OpenRouter's flat rate will build a very different agent from a team optimising for DeepSeek's cache tier.
Price per task beats price per token
A cheaper model that needs three attempts is more expensive than a dearer model that lands first time. This is why the leaderboard framing is the wrong one for anybody actually costing a product. The number you want is cost per successful task, which folds in retries, failed tool calls, wasted context and the human minutes spent cleaning up. We laid out the full method in our guide to LLM unit economics and cost per task, and it applies directly here: a 25.1 on Automation Bench means roughly three-quarters of attempts fail, and each of those failures still bills you for input tokens.
Work an example. Suppose an agent step carries a 180k-token stable prefix and generates 4k tokens of output. On DeepSeek's API with a warm cache, the input side costs about $0.0005 and the output side about $0.0011 — call it a sixth of a US cent per step. On a cold cache, the same step costs about $0.025 on input alone, a fiftyfold jump. A twenty-step task therefore lands somewhere between three cents and fifty cents depending entirely on cache discipline. That range is far wider than any benchmark difference between Flash-0731 and its Pro sibling.
Self-host versus API, recalculated
The 13B active figure changes the self-host arithmetic more than the benchmark table does. Total parameters still govern memory — 284B of weights has to live somewhere — but active parameters govern throughput, and throughput is what determines whether your GPU hours beat an API bill. A sparse model with a low activation ratio gives you far more tokens per second per accelerator than a dense model of comparable quality, which pushes the break-even point down towards a volume that mid-sized teams can actually reach.
For a team in Bengaluru weighing reserved capacity in the AWS Mumbai region against DeepSeek's API, and for a London team comparing the AWS London region under UK data-residency commitments, the calculation now has three inputs rather than two: your sustained tokens per second, your cache-hit rate, and how much you value keeping customer code inside your own VPC. Data residency is often the deciding factor rather than price — a UK health-tech or financial services team frequently cannot send customer context to a third-party endpoint at all, at which point the API price is irrelevant and only the serving cost matters. We ran the full break-even model for the larger sibling in self-hosting DeepSeek V4-Pro on 8× H100, and the structure of that analysis carries over; the Flash numbers shift the crossover point, not the method.
If you are going down the self-host route, budget engineering time for expert routing and memory layout rather than for raw GPU count. Serving a sparse MoE well is a different discipline from serving a dense model, and the failure modes are unfamiliar to teams coming from a Llama-shaped stack.
What to actually do this week
The disciplined sequence, in order, is short.
- Do not switch anything on the leaderboard. Nine vendor-harness numbers are a reason to run an experiment, not a reason to migrate a production agent.
- Stand up a shadow lane. Mirror a slice of production traffic to Flash-0731 without acting on its output. A week of shadow traffic tells you more than any public benchmark, and it costs almost nothing at these prices.
- Measure cache-hit rate first. If your prefix is unstable, you are paying $0.14 per million rather than $0.0028, and every downstream cost estimate you have made is fifty times optimistic.
- Score on your own task set. Build it from your own merged pull requests, your own support tickets, your own data pipelines. Contamination is not a hypothetical problem with a model this recent.
- Re-run the self-host maths only after the above. Serving infrastructure is a months-long commitment; benchmark enthusiasm is a week-long mood.
The wider pattern is worth naming. Open-weight releases have moved from "close to frontier on paper" to "cheap enough that the paper hardly matters", and the analytical work has shifted with them — from reading leaderboards to auditing harnesses and modelling unit economics. That shift is visible across the last several DeepSeek releases, including the V4 open-weight launch and its coding benchmark claims. The teams getting real value out of these models are not the ones reading the tables fastest. They are the ones who built a private evaluation harness eighteen months ago and can answer, in an afternoon, whether a new release helps them.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →The short version
DeepSeek-V4-Flash-0731 is a post-training refresh that leaves the architecture untouched, ships nine impressive but entirely self-measured agent scores, and lands at a price point where cache discipline matters more than model choice. The capability is real enough to test. The ranking is not yet real enough to plan around. Run it through your own harness this week, measure cost per successful task rather than tokens, and let your own numbers decide.