DeepSeek vs OpenAI: V4-Flash Matched Every OpenAI Tier at $0.13 per 1,000 Tasks (and Took 4x Longer)

We ran two DeepSeek models and four OpenAI models through the same executed nine-task Python harness. DeepSeek V4-Flash scored 9/9 at a measured $0.13 per 1,000 tasks. Every OpenAI model we ran also scored 9/9 — GPT-5 mini at $1.53, GPT-5.4 at $1.69, GPT-5.6 Sol at $4.98, GPT-5.5 at $8.83. That is 13x between V4-Flash and GPT-5.4, and 68x between V4-Flash and GPT-5.5, with the score column flat. The catch is real: V4-Flash averaged 14.5 s per task against GPT-5.4's 3.6 s, because it spends 568 reasoning tokens where GPT-5.4 spends zero. So this is a trade, not a winner — DeepSeek buys cost, GPT-5.4 buys latency, and on nine short Python functions they tie on correctness.

Most DeepSeek-versus-OpenAI comparisons are built from two vendor announcements and a price list. This one is built from a run. Six models — two DeepSeek, four OpenAI — went through the same executed nine-task Python harness, with the same prompts, the same grader and the same settings. Here is what came back.

The measured answer

DeepSeek V4-Flash scored 9/9 at a measured $0.13 per 1,000 tasks. Every OpenAI model we ran scored 9/9 too, and the cheapest of them cost $1.53. On correctness, this harness could not separate them. On cost it separated them by more than an order of magnitude, and on latency it separated them the other way.

The three numbers worth carrying away:

• 13x — V4-Flash's $0.13 against GPT-5.4's $1.69 per 1,000 tasks, same 9/9.
• 68x — V4-Flash's $0.13 against GPT-5.5's $8.83, same 9/9.
• 4.0x — GPT-5.4's 3.6 s per task against V4-Flash's 14.5 s, going the other direction.

That last line is why this page does not declare a winner. If your workload is batched and nobody is waiting, DeepSeek V4-Flash did the same job for a thirteenth of the money. If a human is watching a cursor blink, GPT-5.4 finished in a quarter of the time. Both statements come from the same run.

Read this first: what nine functions can prove

The harness runs nine short, self-contained Python functions. The model gets a function signature and a prose spec. It returns code. The code executes against assertions it never saw, in an isolated subprocess. Every assertion passes or the task fails — no partial credit, no human grader, no LLM judge.

Nine bounded functions are a fair test of one thing: can this model turn a clear specification into working code on the first try. They are not a test of long-context reasoning, multi-file refactoring, agentic tool use, or ambiguous requirements — and those are precisely the axes OpenAI and DeepSeek both market their top tiers on. This harness touches none of them.

So read the result narrowly and it is still useful: for bounded, clearly specified code generation, DeepSeek V4-Flash did what four OpenAI tiers did, for between a twelfth and a sixty-eighth of the money. Read it widely and you will be wrong. A tie on nine short functions is evidence that nine short functions do not separate these models — not evidence that nothing does.

We are putting this above the numbers because a cheap-model recommendation is exactly where a benchmark like ours gets over-read. The harness design is on the methodology page, and the full multi-model run is in the coding cost benchmark.

Six models, one harness

Two DeepSeek models and four OpenAI models have been through this harness. Rows are ordered by measured cost, cheapest first. Cost is measured token counts multiplied by that model's list price on the date in the last column — a derived measured cost, not a bill anyone sent us.

Five of the six scored 9/9. The one miss belongs to DeepSeek, and it is the expensive DeepSeek. V4-Pro failed parse_csv_line — quoted fields with embedded delimiters — while V4-Flash, at $0.13 against V4-Pro's $0.74, passed it. Within DeepSeek's own lineup, the cheaper model was both cheaper and more correct on this set, which we go into in V4-Pro against V4-Flash.

Note the mixed price dates. V4-Flash, V4-Pro and GPT-5.5 ran in our original 13-model sweep and are priced at 2026-07-17 list rates; GPT-5 mini, GPT-5.4 and GPT-5.6 Sol ran later on the same harness and are priced at 2026-07-29. For the OpenAI rows that mixture is harmless — none of those list prices moved in between. For V4-Flash it is not harmless: the $0.09 / $0.18 in its row is the rate that produced the $0.13, and it is no longer the rate you would pay. We deal with that below.

The cost gap: 13x and 68x

Here is the same table as a picture, because the ratios are hard to feel in a column of numbers.

The DeepSeek bar is 7.8 pixels wide on the same scale that gives GPT-5.5 a bar 530 pixels wide. That is the whole argument in one image.

Put it in annual terms, because ratios stop being abstract when they have a currency sign on them. A thousand tasks of roughly this size per day, every day for a year, projects to about $47 on V4-Flash, about $558 on GPT-5 mini, about $617 on GPT-5.4, about $1,818 on GPT-5.6 Sol and about $3,223 on GPT-5.5. Those are projections from measured per-task numbers at the stated list prices, not invoices. Wrap any of them in an agent loop that makes twenty model calls per run and the same multiple applies twenty times over — the compounding mechanism we walk through in what agents actually cost.

One clarification on where the ratio comes from, because it is not the sticker price. Our $0.13 was computed at V4-Flash's 2026-07-17 rate of $0.09 / $0.18 per 1M; on list price alone, that against GPT-5.4's $2.50 / $15 is 27.8x on input and 83.3x on output. Our measured gap is 13x — well below either — because V4-Flash emits 568 reasoning tokens per task at the output rate and GPT-5.4 emits none. The token counts give back most of DeepSeek's sticker advantage on this workload. That is the kind of thing a price table cannot show you and a run can.

The catch: 14.5 s against 3.6 s

DeepSeek V4-Flash averaged 14.5 s per task. GPT-5.4 averaged 3.6 s. Four times the wall clock, for the same nine correct answers. This is the honest counterweight to the cost number and it deserves its own chart.

The mechanism is visible in the reasoning-token column. GPT-5.4 emitted 0 reasoning tokens across all nine tasks; V4-Flash emitted 568 per task. A reasoning trace has to be generated before the answer arrives, so it costs wall-clock time and it bills at the output rate. DeepSeek pays for its correctness with thinking; GPT-5.4 answers directly.

There is a second-order consequence that a mean hides. A model at zero reasoning tokens has no cost variance from prompt to prompt. A model that thinks has a per-call cost that tracks how awkward your prompt is rather than how long it is, which makes budgeting harder even when the average is low. If you are sizing a bill in advance, that predictability is worth something the mean does not show. Our cost calculator takes your own token mix if you want to run the arithmetic on your workload rather than ours.

So the decision rule falls out cleanly. Batch work: DeepSeek, and it is not close. Nightly codegen, bulk migration, test scaffolding, anything where the answer is consumed by a machine. Interactive work: GPT-5.4, and it is also not close. Inline completion, a chat that has to feel alive, a lint-and-fix loop with a human in it. Fourteen and a half seconds is a long time to watch a spinner; thirteen dollars where the other model charged one is a long time to look at a bill. Pick the axis your workload actually cares about. If latency is your constraint, the latency guide covers what else moves it.

GPT-5 mini loses on both axes

One row in the table is worth pulling out because it is the only genuinely one-sided result on this page. GPT-5 mini measured $1.53 per 1,000 tasks at 15.2 s per task. DeepSeek V4-Flash measured $0.13 at 14.5 s. Both scored 9/9. V4-Flash is 11.8x cheaper and 0.7 s faster.

That matters because GPT-5 mini is the model most people reach for when they want OpenAI on a budget, and it is the natural head-to-head against V4-Flash: two cheap tiers, two similar reasoning-token profiles (555 against 568 per task), two similar latencies. On this workload it is not a trade at all. The cheap OpenAI tier does not buy speed back, because it thinks about as much as DeepSeek does.

Inside OpenAI's own lineup, GPT-5.4 is the interesting model, not GPT-5 mini. GPT-5.4 costs $0.16 more per 1,000 tasks than GPT-5 mini and returns the answer in 3.6 s instead of 15.2 s — a 4.2x latency improvement for about a 10% cost increase, on identical scores. If you are on GPT-5 mini for latency-sensitive work, that is a straightforward upgrade; we compare the mini tiers in GPT-5 mini against GPT-5 nano and the flagship tiers in GPT-5.4 against GPT-5.5.

The other OpenAI observation, since we have all four in one table: the two most expensive OpenAI models on this harness bought nothing it could detect. GPT-5.6 Sol at $4.98 and GPT-5.5 at $8.83 scored the same 9/9 as GPT-5.4 at $1.69, and both were slower. That is not a criticism of either model — it is what happens when a test sits below the level the top tier was built for. The individual runs are in the GPT-5.6 Sol review and the GPT-5.5 review.

Our $0.13 is a July 17 price, and the price moved

This is the caveat that most affects the headline, so it gets its own section rather than a footnote.

DeepSeek V4-Flash's list price rose between our run and this writing. On 2026-07-17 it was $0.09 input / $0.18 output per 1M tokens. On 2026-07-29 it was $0.14 / $0.28. That is a 1.56x increase on both sides. Our measured $0.13 was computed at the older price.

We have not re-run V4-Flash at the new rate. Because both sides of the price moved by the identical factor, rescaling is exact arithmetic rather than an estimate — but it is still arithmetic, and we label it as such. At about $0.20 per 1,000 tasks, V4-Flash is 8.4x below GPT-5.4 and 43.7x below GPT-5.5 instead of 13x and 68x. Smaller gaps. Still enormous gaps. The conclusion does not move; the number does.

This is not a DeepSeek-specific problem. Forty-nine of roughly 396 catalogue models changed price in the twelve days to 2026-07-29. Any cost comparison you read — including this one — is true as of its pricing date and should be recomputed against current rates before you act on it. We wrote up how fast the floor moves in LLM price volatility, and the live per-token rates for DeepSeek specifically are in the DeepSeek API pricing guide and for OpenAI in the GPT-5 API guide.

What third-party benchmarks say

Our harness answers one narrow question. For the broad ones, the public benchmarks still carry the load, and they are somebody else's numbers — we label them as such.

On independent SWE-bench Verified, GPT-5.5 scores 82.6% (vals.ai) against DeepSeek V4's vendor-reported 80.6% — a few points apart on real-world repository bug-fixing, with the caveat that one figure is independent and the other is not. OpenAI also leads on agentic and terminal benchmarks. Flip to competitive and algorithmic coding — LiveCodeBench, Codeforces — and DeepSeek sits at or near the top of the field.

That pattern is consistent with what we saw. Short, well-specified, algorithm-shaped problems are DeepSeek's home ground, and our nine tasks are exactly that shape. Long-horizon repository work is where OpenAI's lead shows up, and we did not test it. The two sources agree because they are measuring different halves of the same question. Broader cross-vendor context is in the best coding LLM roundup and the three-way V4 against GPT-5.5 against Opus 4.8.

Open weights vs ecosystem

The structural difference survives the benchmark. DeepSeek publishes V4-Pro and V4-Flash as open weights under the MIT licence per DeepSeek's own release, so you can self-host, fine-tune and keep data in-house. OpenAI's GPT-5 models are closed and API-only, with the deepest third-party tooling ecosystem of any vendor.

Two caveats we would rather state than skip. First, open weights and open source are not the same thing, and the distinction matters more than the marketing suggests — we unpack it in open weights vs open source. Second, every number on this page came from an API-hosted model. We have measured nothing about self-hosted DeepSeek: not its quality, not its latency, not its cost per task. Local deployment changes all three, and running DeepSeek locally covers the hardware reality rather than the benchmark one.

Migration between the two is not the hard part. Both expose an OpenAI-compatible API, so switching is a base_url and model-id change, not a rewrite.

Which to pick

The one thing we would not do is pick on the flagship comparison. Neither vendor's top tier earned its price on this harness — GPT-5.5 at $8.83 scored what GPT-5.4 scored at $1.69, and DeepSeek V4-Pro at $0.74 scored a point below V4-Flash at $0.13. Both ladders are pricing structures, and on bounded work neither is a quality gradient. If your workload looks like ours, start at the bottom of both and make the expensive tier prove itself. For the wider sub-$1 field, see the cheap coding model roundup; for the individual runs, the V4-Flash review and the V4 family review.

How these numbers were produced

Nine executed Python tasks: two_sum, valid_parentheses, merge_intervals, roman_to_int, lcs_len, flatten, top_k_words, token_bucket, parse_csv_line.

The model receives a function signature and a prose spec. It never sees the assertions. The returned code runs against hidden assertions in an isolated python3 -I subprocess with a 12-second timeout. Every assertion passes or the task fails — no partial credit, no human grader, no LLM judge.

Temperature 0, max_tokens 4000. One scored attempt per task. The harness retries only on an API error, never on a wrong answer, which is why DeepSeek V4-Pro's miss on parse_csv_line stayed a miss.

Cost is derived, not invoiced. It is the token counts the API reported multiplied by that model's list price on the date shown in the table. Reasoning tokens bill at the output rate and are included.

Calls go through OpenRouter's OpenAI-compatible endpoint, deliberately not through the DataLLM Lab gateway. Nothing on this page depends on our infrastructure, and you do not have to be our customer to reproduce it.

Our original sweep was 13 models run in one sitting; 10 of those 13 scored 9/9, and the spread within that sweep was $0.10 to $8.83, or 88x. Ten further models have since run on the same harness under the same settings, bringing the total to 23, of which 20 scored 9/9. DeepSeek V4-Flash, V4-Pro and GPT-5.5 come from the core sweep; GPT-5 mini, GPT-5.4 and GPT-5.6 Sol are among the later runs. The core sweep was and remains 13.

What we did not measure

Not measured at all: long-context reasoning, multi-file refactoring, agentic or multi-turn tool use, any non-Python language, vision, and translation quality. The harness is single-turn and API-hosted. It does not run agents and it does not use tools. Almost everything DeepSeek and OpenAI differentiate their top tiers on commercially sits in that list.

Nothing here is a self-hosted result. Every figure came from an API call. We have run no local, Ollama or on-premise deployment of any DeepSeek model, so the open-weights advantage is a property we describe, not one we benchmarked.

A known scoring artefact: the 4,000-token ceiling can truncate a very verbose model mid-answer, and a truncated answer scores as a miss. That penalises verbosity, which is a real production cost but is not the same thing as being wrong.

Nine tasks is nine data points. A 9/9 is a clean result on this set, not a claim about a distribution. We did not run any model twice, we did not vary the prompts, and we did not buy anyone a retry. V4-Pro's 8/9 is one failed function, not a verdict on the model.

Not tested, and never claimed as ours: GPT-5 nano, Grok 3, Grok 4, Grok 4.20, grok-build-0.1, VibeThinker, Claude Fable 5, any locally-run model, and any Gemini other than 3.6 Flash and 3.1 Pro. If a first-party number for any of those appears anywhere on this site, it is an error.

FAQ