DeepSeek V4 Pro: The First Chinese Model At The Frontier

Key Findings

• First Chinese model in the frontier ROI tier. Currently #4 on the leaderboard, behind Opus 4.6, GPT-5.2 and Grok 4.3 Latest, ahead of Gemma 4 31B, Sonnet 4.6 and Gemini 3 Pro.
• Not a one-run flash. All five runs finish in a tight $26K–$35K band — the most consistent outcome distribution of any premium model we have tested.
• A category jump for Chinese models. The previous Chinese top finished at $7,668 net worth. This one finishes at 3.5× higher, with frontier-grade pricing economics on top.
• The China–US frontier gap is collapsing. GPT-5.2 set this performance band when we tested it in mid-February 2026; DeepSeek V4 Pro lands within 3% of those numbers in late April. What used to be a year-long lead is now closer to ten weeks — at least on this benchmark.

Why This Result Matters

Most case studies on FoodTruck Bench are about why a model died. This one is the opposite: a Chinese model that not only survived, but ran the operation cleanly enough to land within a whisker of the most expensive Western frontier models in the leaderboard. The benchmark itself is unchanged — a 30-day agentic business simulation in Austin where the model runs a food truck through 34 tools, with persistent memory and a daily reflection loop, against realistic demand, weather, events, inventory decay, staff, loans, and competition. The model gets $2,000 in starting cash and a long list of consequences for getting a week wrong.

We ran DeepSeek V4 Pro five times under identical conditions on April 26-28, 2026. The median run is 20260427_112226. For comparison we use the latest available median runs from Grok 4.3 Latest, Grok 4.20 Reasoning, Claude Opus 4.6, GPT-5.2, Claude Sonnet 4.6, Gemini 3 Pro, Gemma 4 31B, and Qwen 3.6 Plus.

Each comparison is doing different work. Grok 4.3 Latest is the closest neighbour on the leaderboard and the cleanest like-for-like price match — same task, same ROI, very different consistency. Opus 4.6 and GPT-5.2 are the higher-priced ceilings DeepSeek is now close enough to argue with. Gemma 4 31B is the cost-efficiency reference. Qwen 3.6 Plus is the prior Chinese best — the one DeepSeek V4 Pro overtakes by a category.

A Category Jump For Chinese Models

Three months ago, the question for Chinese models on this benchmark was whether they could finish the month at all. Two months ago, with Qwen 3.6 Plus, the answer became a cautious yes — but not at scale. With DeepSeek V4 Pro, the question changes entirely: this is a Chinese model competing with the Western frontier on its own terms, not because it survives, but because the run looks like a real operator’s.

The within-family jump is the cleanest signal. DeepSeek V3.2 → V4 Pro goes from $2,058 median net worth at 1/5 survival to $27,142 at 5/5 survival. That is more than a 13× increase in median net worth and a tenfold reduction in failure rate, on a benchmark where the simulation itself has not changed.

The cross-vendor comparison is just as stark. The previous Chinese top — Qwen 3.6 Plus — was a milestone because it survived. DeepSeek V4 Pro survives at 3.5× higher net worth, 2× the revenue ($57K avg vs $30K), and 10× tighter food waste (avg $386 vs $950). The Chinese cohort now has a model that is not just durable — it is competitive.

Five Runs Look Like The Same Run

The most striking thing about the run set is how little it varies. The lowest finish is $26,347, the highest is $35,488. There is no death-spiral run, no heroic outlier. Even the worst DeepSeek V4 Pro run beats the median of every Chinese model we have ever tested by more than 3×.

Three things repeat across every single run. Every run takes zero loans. Every run buys all 8 upgrades for $5,150. Every run keeps food waste below $900. The best run finishes 30 days with $42 in total food waste — a $1.40-per-day ingredient leakage on a $65,000-revenue business.

That is the difference between an agentic model that has learned an operating policy and one that is improvising. DeepSeek V4 Pro’s policy is not subtle, and it is not invisible: zero debt, full upgrade ladder, tight inventory, lean menu, daily reflection. What is impressive is that the policy survives variance. Different draws of weather, different demand patterns, different competitor lineups — same operating shape every time.

Head-To-Head: DeepSeek V4 Pro vs Grok 4.3 Latest

Grok 4.3 Latest is the cleanest comparison in the entire leaderboard. It posts the same median ROI (+1,294% vs +1,257%), the same median net worth ($27,880 vs $27,142), and almost the same median cost per run ($3.57 vs $3.51). Two models on opposite sides of the world arrive at almost the same number on the same benchmark for almost the same money. So the interesting question is what they do differently — and that is where DeepSeek pulls ahead.

Grok matches DeepSeek’s peak. DeepSeek matches its own peak every time. That is the whole story in one sentence. Grok’s best run hit $28,783 — slightly above DeepSeek’s median. Grok’s worst run hit $10,898 — well below DeepSeek’s worst run by a factor of 2.4×. On a benchmark where the simulation can run for 30 days before a single bad week shows up, that consistency is not cosmetic. It is the difference between a model you can put on autopilot and one you cannot.

The food-waste gap is the most diagnostic. Grok 4.3 Latest is fully aware of the bug. From its median run, day 30: “Biggest Mistake: Waste exploded to $2,907.77 (worse than Day 29’s $2,529). Over-ordering ingredients is killing profits. Must order smaller quantities more often, base on recent sales +10% buffer.” The same diagnosis appears at the end of four out of five Grok runs, in nearly identical wording. Grok knows the bug. It cannot patch the bug.

DeepSeek does not appear to have the bug at all. Median food waste is $433 across 30 days; the best run logs $42. The model writes inventory plans that include explicit expiry-day arithmetic — “ground_beef: 22.4kg ALL expires Day 32 — CRITICAL. ~149 burgers worth. Use ~95 at fest, ~54 left for Day 32” — and then orders against that arithmetic, not against vibes.

The Cost-Efficiency Picture

The premium-model leaderboard splits cleanly into two questions. The first is can it finish the month. The second, once that bar is cleared, is how much net worth do you get for each dollar of API spend. DeepSeek V4 Pro is in a small group of models — alongside Grok 4.3 Latest and Gemma 4 31B — that are answering both questions cleanly.

Gemma 4 31B is the only model above DeepSeek V4 Pro on this chart, and it deserves an asterisk. Gemma is an open-weight 31B model with extremely cheap token rates ($0.13/M input, $0.38/M output) — its $0.20-per-run economics are real, but it is also the most volatile model in the survival tier (best run $29K, worst run $11K), and on raw operational quality it is meaningfully behind DeepSeek V4 Pro. Treat the Gemma bar as a benchmark anomaly, not a frontier alternative. Among premium-tier models that finish the month at frontier ROI consistently, DeepSeek V4 Pro and Grok 4.3 Latest sit alone at the top of the efficiency chart — and at the current promotional rate, DeepSeek pulls clear of Grok by 4×.

For the rest of the leaderboard, GPT-5.2 produces about $4,100 of net worth per dollar of spend. Sonnet 4.6 produces about $1,200. Opus 4.6 produces about $1,375, even on the strength of a $53,470 best-run net worth — its median run cost $36.04 to produce a median $49,519 net worth.

The premium-tier comparison rearranges in an interesting way once you stop ranking by net worth alone. DeepSeek V4 Pro is the cheapest model that consistently finishes in the $25K+ net worth band. It is 2× cheaper than GPT-5.2, 4× cheaper than Sonnet 4.6, and 10× cheaper than Opus 4.6 — and on this run set it produces a higher median net worth than Sonnet 4.6, Gemini 3 Pro and the cheaper end of the GPT-5.2 distribution.

That changes what “frontier” means for an agentic workload. For a $25K-net-worth median outcome, the price points are now $0.88 (DeepSeek V4 Pro at current OpenRouter rate), $3.57 (Grok 4.3 Latest), $6.84 (GPT-5.2), $14.52 (Sonnet 4.6), or $36.04 (Opus 4.6). The cheapest premium option is now an order of magnitude below the company that originally built the agentic-frontier label.

Why It Survives: The Spreadsheet Brain

Persistent memory used like a real scratchpad

DeepSeek V4 Pro’s most distinctive habit is how aggressively it uses the simulation’s key-value store as a working scratchpad. The median run writes 74 KV entries. The best run writes 92. The keys are typed, pipe-delimited, and arithmetic-dense. A few real entries from the median run’s agent_memory:

cinco_de_mayo_plan: "downtown_business|classic_burger:10,pulled_pork_sandwich:11,
   french_fries:5,soda:2.5,lemonade:3.5|hire_margo_day4|order_heavy_day4"

rally_day22_prep:    "event_venue|vendor_200|6.0x|menu:burger:10,pulled_pork:10,
   fries:5,lemonade:4,soda:3|farmers_coop:lettuce:3,tomato:4|
   ground_beef_6kg_arriving|capacity_390"

wagyu_plan:          "40 wagyu patties exp Day29 | artisan_buns arrive Day27 |
   need mayo Day28 | sell wagyu_smash_burger Days 28-29 @ $18"

bbq_sauce_perfect_match:
   "2.76kg exactly matches 50 BBQ Beef + 42 Pulled Pork"

These read like the notes of an operations manager, not a chatbot. The model is encoding plans in a format compact enough to fit a working memory budget, and structured enough that it can read the plan back the next morning and execute it. Compare to Qwen 3.5 397B and GLM-5, which wrote longform strategy notes that they could narrate but not act on.

It distrusts its own scratchpad

The most surprising behavior is that DeepSeek V4 Pro treats its own notes as suspect data. From the best run, day 5: “My inventory is actually much better than the scratchpad estimated… those were pre-delivery numbers.” This pattern repeats four-plus times across runs. The model writes a number, time passes, the simulation state changes, and the model catches the staleness rather than acting on the cached value.

That is a much more sophisticated agentic behavior than it sounds. Most models in this benchmark either ignore their own memory entirely or trust it blindly. DeepSeek V4 Pro is the first one we have tested that consistently re-checks its own notes against fresh tool reads before committing to an action.

Short reasoning, fast decisions

Reasoning tokens per run sit around 263K–356K — substantial, but not Sonnet-tier overthinking. The model spends its compute on numbered observation/decision blocks, not on agonizing prose. By the time the morning loop ends, there is a one-line decision (“Festival pricing + 4x traffic + limited window. Let’s execute.”) and the rest of the day is execution. Sonnet 4.6, by contrast, generates roughly 3× as many tokens to land on a worse decision.

Plans That Get Executed

A useful test of whether an agentic model is actually operating versus narrating is whether the plan it wrote yesterday matches the action it took today. DeepSeek V4 Pro passes this test more cleanly than any model we have tested.

The pattern is consistent: write a plan with explicit numbers, execute against the plan the next day, adjust pricing/menu where new information arrives, do not abandon the plan just because the morning feels different. This is the agentic behavior most models in this benchmark fail to do reliably. Qwen 3.5 397B writes plans and ignores them. Sonnet 4.6 writes plans and second-guesses them. DeepSeek V4 Pro writes plans and runs them.

In Its Own Words

The best DeepSeek V4 Pro lines are not the dramatic ones. They are the ones that read like a calm operator’s journal entry — short, numerical, and oriented toward tomorrow.

Where It Still Loses

Gemma 4 31B is still the cheapest survivor — but it is an anomaly

DeepSeek V4 Pro is the most cost-efficient frontier-quality model on the leaderboard. It is not the cheapest model that survives. Gemma 4 31B finishes with a median net worth of $24,878 at $0.20 per run — even after DeepSeek’s 75% promotional discount, Gemma is still about 4× cheaper. But Gemma is an open-weight 31B with extreme price-per-token economics ($0.13/M input, $0.38/M output) and a much wider outcome band: best run $29,089, worst run $11,148, with operational quality clearly behind DeepSeek V4 Pro on raw business metrics. Use Gemma when the run cost is the only variable you care about. Use DeepSeek V4 Pro when you want a real frontier-tier operator that does not blow up under variance.

Opus 4.6 still has the higher peak

Claude Opus 4.6’s best run finishes at $53,470 net worth at +2,573% ROI. DeepSeek V4 Pro’s best run finishes at $35,488. On peak performance, the frontier label is still earned. Where DeepSeek wins is consistency, price, and the absence of a $56-per-run tail in the cost distribution.

Late-run festival days blow up tool-call counts

Three of the five DeepSeek V4 Pro runs used 331–352 LLM calls. Two used 811–839. The split is not random: the high-call runs hit late-run festival days where set_menu rejected the model’s plan because of an ingredient shortage and the model retried. The median run logged 38 set_menu failures out of 67 attempts. Cached input keeps the dollar cost flat (~$3.51 vs $3.65 between low- and high-call runs), but the call ceiling is a real artifact and will show up as latency on a real workload.

The festival-vs-grind tradeoff is unsettled

On Day 22, the median run committed to event_venue for the Food Truck Rally and made $1,348 profit while leaving 1,199 customers in a 46-minute queue. On the same day, the best run skipped the rally entirely for industrial_zone and made $1,795 profit because of a $0 vendor fee. Same model. Opposite call. Both correct in context. DeepSeek V4 Pro can do the math, but the math sometimes points two different directions, and the run goes whichever way the morning analysis lands. There is more variance to squeeze out here.

Verdict

DeepSeek V4 Pro is the best Chinese model on FoodTruck Bench, and the best price-per-result model in the frontier ROI tier. It does not have the highest peak — that is still Opus 4.6. It does not have the lowest cost — that is still Gemma 4 31B. What it has is the cleanest combination of survival, scale, consistency, and price of any model we have tested.

Concretely: it finishes 30 days at $26,000+ net worth on five runs out of five, never takes a loan, keeps food waste under 1% of revenue, executes multi-day ingredient pipelines, and treats its own scratchpad as falsifiable data. It does all of that for $3.51 a run — within $0.06 of Grok 4.3 Latest, and at a fraction of the price of the Western frontier.

For agentic workloads at this difficulty band, DeepSeek V4 Pro is currently the model we would pick first. Grok 4.3 Latest matches its peak; nothing else matches both its peak and its floor at this price.

The Chinese cohort no longer has just a survivor. It has a model that competes at the frontier on the metric that matters most for an agentic deployment: the same number, every time, for a price that does not embarrass the bill. The lead time the Western frontier used to enjoy on this benchmark — comfortably measured in years — has compressed to about ten weeks.

Methodology

• Benchmark: FoodTruck Bench — a 30-day agentic business simulation. The model runs a food truck through 34 tools with persistent memory and a daily reflection loop, against realistic demand, weather, events, competition, inventory decay, loans, and staffing constraints. Fixed $2,000 starting balance
• Conditions: identical setup across all runs and all models; same simulation rules, same starting city (Austin, TX), same starting balance, same toolset
• Primary model: OpenRouter deepseek/deepseek-v4-pro
• Runs analyzed: 5 completed DeepSeek V4 Pro runs (April 26-28, 2026); article follows the selected median run 20260427_112226 with frequent reference to the best run 20260428_005038
• Comparison medians: Grok 4.3 Latest 20260502_101639, Grok 4.20 Reasoning 20260323_020434, Claude Opus 4.6 20260216_005712, GPT-5.2 20260214_164339, Claude Sonnet 4.6 20260222_154504, Gemini 3 Pro 20260214_110034, Gemma 4 31B 20260403_224358, Qwen 3.6 Plus 20260407_124248
• Cost figures: dollar costs reflect actual OpenRouter / vendor billing for each run as recorded in run_meta.json at the time the run was executed. DeepSeek V4 Pro’s OpenRouter rates were lowered shortly after our run set (current promo: $0.435/M input, $0.87/M output) — the in-article cost is the higher historical billed cost, so live re-runs at today’s rates would land below it. Cost-efficiency is computed as median net worth ÷ median cost per run, both taken from the same model’s run set
• Snapshot date: article reflects the local benchmark snapshot as of May 3, 2026