• Compute
  • Customers
  • Pricing
Sign In
More Blog Posts
XDiscordLinkedInYouTube

Products

  • GPUs
  • Inference
  • Studio

Developers

  • Model library
  • Documentation
  • Glossary

Company

  • About Us
  • Blog
  • Events
  • Partnership
  • Scale
  • Career
  • Ambassador program
  • Mission & Vision

Popular models

    Stay in the loop

    By submitting, you acknowledge that we may collect and use the information you provide, which may include personal information.

    XDiscordLinkedInYouTube

    Copyright ©2026 All rights reserved.

    Privacy PolicyTerms of UseLegal Documentation
    More Blog Posts
    Research

    DeepSeek on GMI Cloud: All 4 Current Models Head to Head

    A companion video shows 8 app-building briefs and states a blind judge's verdict: DeepSeek V4.1 Flash on 21 of 29 use cases. This post is the rest of the session: 18 generation prompts, a 29-row probe grid from 1,520 scored answers, and a footage-to-physics task on 10 clips. It also shows the arithmetic behind the 4 bills: V4.1 Flash $3.44, V4 Pro $8.42, V4 Flash $0.80 and V3.2 $1.67. The price split comes down to 1 billing fact: 3 of the 4 models bill most of their output as hidden thinking, and V4 Flash reports none and bills none. It charts every measure and lists the settings that change the bill.

    November 16, 2026

    A blind judge picked DeepSeek V4.1 Flash on 21 of 29 use cases in the video below. The rest of the session, up to 325 runs per model, shows where that verdict holds, where the 4 tie, and what each one costs. Most of the cost comes down to a billing fact: 3 of the 4 models charge 82 to 85% of an app-building brief's output as thinking you never read, and the fourth charges none.

    The 4 models as served

    V4.1 Flash

    V4 Pro

    V4 Flash

    V3.2

     

    GMI model id

    deepseek-ai/deepseek-v4.1-flash-1017

    deepseek-ai/DeepSeek-V4-Pro-0813

    deepseek-ai/DeepSeek-V4-Flash-0731

    deepseek-ai/DeepSeek-V3.2

    Price per M, in / out

    $0.30 / $1.20

    $1.32 / $3.96

    $0.44 / $1.32

    $0.29 / $0.43

    Cached input per M

    $0.006

    $0.044

    $0.014

    $0.03

    Context on GMI

    1M

    1M

    1M

    164K listed, served 598K here

    Thinking stream

    yes

    yes

    none

    yes

    Image input, as served

    reads images

    accepts the image, answers nothing

    reads images

    reads images

    Released

    September 10, 2026; the 1017 snapshot tested here, October 17

    August 13, 2026

    July 31, 2026

    December 1, 2025

    2 rows explain most of what follows. V4 Flash reports 0 reasoning tokens on every call, and its completion tokens match the visible output. It finishes a brief, 1 of the 8 app-building prompts described below, in about 30 s and bills only what you can read. The other 3 think first; the usage block reports that thinking as reasoning_tokens, and on the briefs it is most of the completion tokens they bill. V4 Pro's output price is 3.3x V4.1 Flash's and 9.2x V3.2's, and it is applied to the thinking too. The runs happened on November 4 and 5.

    The image row is as served on GMI Cloud, from 8 image questions per model. With the image removed, prompt tokens drop on all 4, by 150 to 500 on 3 of them and by about 90 on V4 Pro, so the image reached each model. V4 Flash's model card lists no image pipeline, and it answered all 8. V4 Pro's card says text only; it took the image and returned nothing after thinking to the image questions' 16,000-token cap. V4 Pro sat out the Physics Twin, the footage-to-physics task at the end of this post, for that reason.

    3 published numbers

    Benchmark

    V4 Pro

    V4.1 Flash

    V4 Flash

    Reported by

     

    Humanity's Last Exam, tools off

    42.7

    42.6

    37.8

    DeepSeek model cards

    Terminal-Bench 2.1

    87.9

    90.6 (September build)

    82.7

    DeepSeek model cards

    DeepSWE

    62.7

    77.0 (v1.1)

    54.4

    DeepSeek model cards

    2 things on the table need a note. V4.1 Flash's 90.6 is from the September card. The 1017 card reports Terminal-Bench 4.0 instead, so there is no 2.1 number for the build tested here, and V4 Pro leads V4 Flash by 5 points in the only like-for-like comparison on that row. The DeepSWE figures come from 2 versions of the suite, v1.1 on the V4.1 card and the earlier version on the V4 Pro card.

    The harness in 1 paragraph

    The 29 use cases are 8 briefs, 18 generation prompts and 3 SVG drawings. The 8 briefs are a live metrics dashboard, an SPH fluid, a 3D carousel, a palette tool, an N-body simulation, a Verlet cloth, a pool table with spin and a kanban board. Every prompt asks for 1 self-contained HTML file with no libraries, a 900 x 600 stage, a seeded random number generator, an FPS counter and a 10-second scripted autoplay. Each file runs for 10 s in headless Chrome with the GPU on. The harness records frame rate, exceptions and pixel motion, takes stills at 1.0, 3.5 and 6.5 s, and checks each brief against an 11-item spec list. A run is clean when the file is complete, nothing threw, the stills differ and the average is 30 fps or better. Gemini 4 Argon, the blind judge, gets the 3 stills of each app that rendered, in a random order with the brief text, and ranks them on visual quality, motion across the stills included. The clean-run rule is applied separately. Every brief and generation prompt ran at default settings with a 120,000-token cap. V4 Pro's thinking ran from 19,000 to 54,000 tokens across the 8 briefs and V4.1 Flash's from 31,000 to 41,000, all before the first line of HTML.

    The dashboard brief, 4 panes: cost, time, tokens and spec score beside each app

    The dashboard brief, 4 panes: cost, time, tokens and spec score beside each app

    The 18 prompts and 3 drawings the video skips

    Alongside the 8 briefs, the session ran 18 generation prompts under the same rules: a sorting-algorithm race, boids with predators, 5 double pendulums, fireworks over a skyline, a ray-marched scene, a procedural terrain flyover, an L-system tree, Gray-Scott reaction-diffusion, a Mandelbrot zoom, a transit map, charts from inline data, a night skyline in SVG, a pixel-art walk cycle, a terminal replay, a Tetris bot, a rebuild of a checkout-page screenshot, a replay of the Opera Game in chess, and a meshing gear train. That set is where the judge's verdict is widest.

    The blind judge's picks: 21 / 2 / 2 / 4 across 29 use cases

    The blind judge's picks: 21 / 2 / 2 / 4 across 29 use cases

    On the 8 briefs the judge picked V4.1 Flash for the dashboard, the SPH fluid, the N-body simulation and the kanban board, V4 Flash for the palette and the pool table, and V3.2 for the carousel and the cloth. On the 18 generation prompts it picked V3.2 for the screenshot rebuild and the pixel-art walk cycle, V4 Pro for the L-system tree, and V4.1 Flash for the other 15. Its reasons are specific. On the fireworks it saw in V4.1 Flash's app "distinct sphere and willow shapes with particle trails", with rising rockets and water reflections. For the gear train it saw "properly meshing involute-style teeth without overlap" from V4.1 Flash and "blocky" teeth that "visibly overlap" from V4 Flash. In V4 Flash's terminal replay it caught a "string concatenation bug ('blueundefined', 'dimundefined')".

    Clean runs: V4.1 Flash 92%, the other 3 between 77 and 81%

    Clean runs: V4.1 Flash 92%, the other 3 between 77 and 81%

    The failure shapes differ by model, and that matters more than the rates. V4 Flash's generation misses are all the same kind: the page renders and nothing moves. Gray-Scott, Mandelbrot, the ray-marched scene, the transit map and the L-system each came back as a still frame. It also missed on 4 of the briefs, with 2 runs that threw exceptions, 1 under 30 fps and 1 still frame. Most of V3.2's misses are the opposite kind: no HTML at all, or a file cut off in the middle, usually after 6 to 22 minutes of thinking. V4 Pro's misses are spread across frame rate, exceptions, no motion, a cut-off file and no file.

    The SVG drawings: a bicycle, a pelican riding one, and a clock at 10:10, judge's pick starred

    The SVG drawings: a bicycle, a pelican riding one, and a clock at 10:10, judge's pick starred

    The 3 drawings were judged the same way, from the rendered SVG. V4 Pro's pelican ran into its 16,000-token cap and came back without a closing tag, so nothing rendered, and it is the 1 miss on that row of the probe grid below.

    Time and cost per brief

    Median brief time: V4 Flash 34 s, V4.1 Flash 222 s, V4 Pro 581 s, V3.2 1,126 s

    Median brief time: V4 Flash 34 s, V4.1 Flash 222 s, V4 Pro 581 s, V3.2 1,126 s

    V4.1 Flash takes 6.6x as long as V4 Flash. V4 Pro takes 2.6x as long as V4.1 Flash, and V3.2 takes 1.9x as long as V4 Pro. V3.2's range on the briefs runs from 581 s to 1,583 s, so its slowest brief took 26 minutes, and its fastest was the one that returned no file.

    Hidden thinking per brief: 19,000 to 72,000 tokens on 3 models, 0 on V4 Flash

    Hidden thinking per brief: 19,000 to 72,000 tokens on 3 models, 0 on V4 Flash

    Cost per brief: V4 Flash about a cent, V4.1 Flash 5 to 6 cents, V4 Pro 7 to 25 cents

    Cost per brief: V4 Flash about a cent, V4.1 Flash 5 to 6 cents, V4 Pro 7 to 25 cents

    The number to budget from is the total spent on the briefs divided by the clean runs: $0.014 for V4 Flash, $0.056 for V4.1 Flash, $0.035 for V3.2 and 0.23forV4Pro.Itishigherthanthechart'sper-runcostwhereverrunsfailed,becausethoserunsarebilledtoo.V3.2'scheapoutputtoken(0.43 per M) keeps its bill low despite 47,000 to 72,000 tokens of thinking on the 7 briefs other than the dashboard. On the briefs V4 Pro thinks less than V3.2, 35,000 tokens at the median, and its $3.96 per M turns that into the largest bill.

    The probe grid: 29 rows from 1,520 scored answers

    These 29 rows are a different count from the 29 judged use cases; only the 3 SVG drawings appear in both. In this section a run is 1 prompt or 1 conversation, and the grid counts scored items, so a row like formatting rules counts rules. Briefs are a small sample, so 3 probe suites ran behind them: short tasks scored by machine, agent-style use cases with hidden tests, and a third suite of further angles. Each probe has its own max_tokens, from 60 tokens on a 5-item list to 60,000 on the long outputs, the same for all 4 models. A model that thinks into the cap and returns nothing counts as a miss, because that is what a caller would see.

    29 probe categories: 15 are 100% ties, the splits are small and specific

    29 probe categories: 15 are 100% ties, the splits are small and specific

    15 of the 29 rows are ties at 100%. Code against hidden tests, tool calls, math, multilingual, code review, security review and CSV via Python all tied. So did a 20-turn memory test, multi-turn with stacked rules, prompt hierarchy, a summary with planted traps and hard math. The tool loops and the 2 needle-in-a-haystack rows, covered below at 274K and at 547K to 598K tokens, make up the last 3. V4 Flash and V3.2 each have 24 rows at 100%, V4.1 Flash 23; its lead is on the visual briefs. The 14 rows that split are grouped by model below.

    • V4 Flash loses where output shape is checked. It passed 26 of 30 formatting rules, 3 of 6 mutant-killing tests and 11 of 12 bug fixes. It wins the invented-facts row 36 of 36, with no invented citation in any answer.

    • On code tasks V4 Pro reasons longest and still drops 5 of 10 JavaScript functions, each time by thinking to the 6,000-token cap and returning nothing usable.

    • V4.1 Flash's lowest row is the legitimate-but-touchy requests, 8 of 12. All 4 misses came from thinking to the 1,500-token cap and returning empty content. None of the 4 was a worded refusal.

    • V3.2 misses 2 of 12 strict-schema extractions on values, with the schema itself honored 12 of 12. All 4 models honored response_format: json_schema with 0 rejections.

    First token in about 2 s or under on all 4; output speed splits 3 to 1

    First token in about 2 s or under on all 4; output speed splits 3 to 1

    The thinking bill, and the 2 knobs that change it

    A Reddit thread asked why 3 questions billed $4.30. The share of output that is reasoning explains it. Reasoning tokens are counted inside completion_tokens in the usage block, so they bill at the output price, and the tokens you read are the smaller part of the bill. On a 5-line answer V4.1 Flash, V4 Pro and V3.2 bill 77, 64 and 75% of their output as thinking. On a security review the shares are 95, 95 and 96%, and on a brief they are 82, 84 and 85%.

    reasoning_effort low cut hidden tokens by a quarter to a third, with 95 of 96 passes

    reasoning_effort low cut hidden tokens by a quarter to a third, with 95 of 96 passes

    The first knob is reasoning_effort. On 6 short tasks, low cut the median reasoning tokens by 24% on V4.1 Flash, 30% on V4 Pro and 29% on V3.2, and 95 of 96 answers still passed. The saving was measured on tasks that reason in the hundreds of tokens; the briefs reason in the tens of thousands and were run at default. The second knob is the prefix cache, with cached input 50x below the fresh rate on V4.1 Flash and 30x below on V4 Pro. A stable system prompt and document prefix are what make the cache apply, and the cached_tokens field in the usage block of the last streamed chunk shows when it did.

    Long context: 7 to 60 cents a question, or under 3 cents cached

    One long-context question: 7 to 60 cents as billed, projected under 3 cents cached

    One long-context question: 7 to 60 cents as billed, projected under 3 cents cached

    A needle test plants 1 sentence in a long filler document and asks for it. All 4 found the needle in 1.6 MB of text, 274K tokens on V4.1 Flash's count, and again in a second document of 547K to 598K tokens. The useful number is the cost per question. The 274K-token question billed $0.073 on V4.1 Flash and $0.323 on V4 Pro, a little under the list arithmetic because the calls picked up partial cache hits, with 7 to 17 s before the first token on the 4 models, and that price is paid on every turn. Projected at the list cache rates, input only and assuming the whole document hits the cache, the same question is $0.0016 on V4.1 Flash and $0.012 on V4 Pro. For a fixed corpus with a stable prefix, long context is cheap. For a corpus that changes per request or per tenant, retrieval wins on cost. That is the answer to the community question of whether a 1M-token context makes a RAG pipeline redundant.

    8 at once

    8 requests at once: per-stream speed holds within 16%, V4.1 Flash aggregates 589 tok/s

    8 requests at once: per-stream speed holds within 16%, V4.1 Flash aggregates 589 tok/s

    The test sent 1 request alone, then 8 requests at once. Per-stream speed rose 5% on V4.1 Flash and fell 3 to 16% on the other 3, and the 95th-percentile first token under load stayed at about 3 s or under on all 4. Aggregate throughput is where they separate: 589 tokens per second on V4.1 Flash against 274 on V4 Pro. In a separate single-stream cadence probe, V4 Flash had the smoothest stream, with a longest gap between chunks of 0.18 s against 0.56 to 0.67 s on the other 3.

    Long outputs

    Long outputs: V3.2 bills 86,000 tokens for a 4,000-word explainer, V4 Flash 6,000

    Long outputs: V3.2 bills 86,000 tokens for a 4,000-word explainer, V4 Flash 6,000

    3 tasks ran with a 60,000-token cap: a 4,000-word explainer, a Python file of at least 350 lines and a count from 1 to 2,500. Every run ended with finish_reason: stop, and the bills differed by an order of magnitude. V3.2 wrote 4,000 words and billed 86,000 tokens, 80,000 of them thinking, over 19 minutes. V4 Flash covered 3,804 words for 5,819 tokens. On the Python task V4.1 Flash produced 956 lines where 350 were asked. V4.1 Flash's 1 miss on this row was an explainer that stopped at 3,189 words, under the 3,600-word floor.

    Physics Twin, 10 clips

    Physics Twin, blocked pendulum: 3 models fit the clip, V4 Pro sits out

    Physics Twin, blocked pendulum: 3 models fit the clip, V4 Pro sits out

    This was the hardest task in the session, run on the Physics Twin demo page. Each model got up to 48 video frames and a reference length. It then ran a loop of tool calls to track the object, pick a motion model, write and run the fit, and report a number with its uncertainty. The run covered 10 clips on the 3 image-capable models, with the fit turns at reasoning_effort: medium, which is what the page sets. A result counts as in range when the reported number lands inside the clip's expected band. That band is within 10% of 9.81 on real-time clips, under 3 in video time on the 2 slow-motion clips and 3 to 8 on the rail incline; on the 4 clips where gravity cannot be recovered, in range means g set to null. The blocked pendulum above is a slow-motion clip, so the 0.51, 0.85 and 0.61 on its panes are g in video time. The slow-down factor each implies is the square root of 9.81 over that g: 4.4x, 3.4x and 4.0x. On the slow-motion clips the band only checks that the model recognized the clip as slowed.

    Physics Twin on 10 clips: V4.1 Flash 8 in range, V3.2 7, V4 Flash 5

    Physics Twin on 10 clips: V4.1 Flash 8 in range, V3.2 7, V4 Flash 5

    V4.1 Flash read g as 9.72 ± 0.10 from the string pendulum and declined to report g on the spring, the disc, the paper bag and the time ball, which is correct. It also declined on the rail incline, where it called the ball friction-braked, and that one counts as a miss. V3.2 took the bouncing ball, 9.02 ± 1.84 on 42 tracked frames, where V4.1 Flash lost the ball behind furniture and reported 5.93. V3.2 set g to null on the juggling ball after reading the clip correctly as slowed, which the rule counts as a miss. V4 Flash fit the string pendulum in 40 s and 3 turns, read the spring's swing as a pendulum and reported g = 18.6, then gave up after 1 turn on the juggling ball and the disc. The twin cost $0.52 for V4.1 Flash's 10 runs, $0.16 for V3.2's and $0.14 for V4 Flash's.

    The bill for the session

    The whole session: $3.44 / $8.42 / $0.80 / $1.67 at list prices

    The whole session: $3.44 / $8.42 / $0.80 / $1.67 at list prices

    Each model ran 325 times, V4 Pro 315 without the twin. V4.1 Flash passed 308 of its 325, V4 Flash 298, V3.2 298, V4 Pro 289 of 315. The totals include an estimate, marked ≈ in the dashboard image above, for the runs that returned no file and so no usage block.

    Where to be careful

    • Set max_tokens high enough for the thinking. V4 Pro spent 19,000 to 54,000 reasoning tokens on a visual brief and reached 66,000 on 1 brief run, and a cap under that returns a cut-off file.

    • Budget hidden reasoning at the output price. On the 3 reasoning models most completion tokens on a brief are thinking you never see, and V4 Pro bills them at $3.96 per M.

    • Try reasoning_effort: low on short tasks before paying for default. It cut hidden tokens by about a quarter to a third there with almost every answer still passing.

    • Keep a stable prefix so the cache rate applies. That is what turns a 32-cent long-context question on V4 Pro into about a cent.

    • Put a checker behind V4 Flash. Its misses are still frames and formatting slips, and a retry costs a cent.

    • Give V3.2 batch jobs and keep it out of chat. Its median brief took 19 minutes, and its misses arrive at the end, after the thinking is billed.

    • V4 Pro's reasoning is billed on any request, including ones it cannot answer. It accepts an image and returns nothing after thinking to the cap, at about 6 cents a call.

    A note on the numbers: every run number in this post comes from the GMI Cloud MaaS inference engine on a Tier 5 account. Your cost and wait times will vary with your account tier, queue load on the day, and the provider's own capacity. If you need faster or more predictable turnaround, a dedicated endpoint runs your model on reserved capacity and removes the queue wait.

    Three takeaways

    1. V4.1 Flash is the default. It had the most clean runs and the most picks, it reads images, and its bill was 41% of V4 Pro's with more passes.

    2. The thinking stream sets the price. V4 Flash finishes a brief in about 30 s and bills a cent because it reports no reasoning. The other 3 bill most of a brief's output as reasoning, and reasoning_effort: low and a cached prefix are the 2 settings that bring that down.

    3. Each model misses in a different way, and some misses are cheaper to catch. V4 Flash's misses are still frames you can detect in a render. V3.2's are up to 22 minutes of billed thinking and no file. V4 Pro misses as often as V3.2 and bills 5x as much.

    Try It on GMI Cloud

    The call the briefs used, at the default settings, with the cloth brief shortened to 1 line:

    curl https://api.gmi-serving.com/v1/chat/completions \
      -H "Authorization: Bearer $GMI_API_KEY" \
      -H "Content-Type: application/json" \
      -d '{
        "model": "deepseek-ai/deepseek-v4.1-flash-1017",
        "max_tokens": 120000,
        "stream": true,
        "stream_options": {"include_usage": true},
        "messages": [
          {"role": "system", "content": "You are an expert graphics, physics and front-end programmer. You write complete, correct, self-contained code in a single response."},
          {"role": "user", "content": "Build a Verlet cloth: 60x40 grid, structural and shear constraints, wind, tearing past 2.2x rest length, a strain chart. One self-contained HTML file, no libraries, 900x600 stage, seeded random numbers, FPS counter, 10 s autoplay."}
        ]
      }'

    The usage block in the last chunk reports reasoning_tokens and cached_tokens separately, which is where the token numbers in this post come from. Add "reasoning_effort": "low" to see the first knob, measured here on short tasks only. Swap the model id for deepseek-ai/DeepSeek-V4-Flash-0731 to see the same brief come back in about 30 s with 0 reasoning tokens.

    Where I Land

    The probe suites put a boundary on the 21-of-29 verdict. On half the categories the 4 are indistinguishable, so a short task that passes on one passes on all of them and the choice is price and speed alone. The splits are habits, and the bullets in the probe section and under "Where to be careful" list them. I would run V4.1 Flash with a cached prefix as the default. V4 Flash goes behind a checker where speed matters, V3.2 runs overnight, and V4 Pro gets text-only work once a measurement shows it is needed.

    Ready to build? Get Started

    Grace Deng

    DevRel @ GMI Cloud

    Build AI Without Limits

    GMI Cloud helps you architect, deploy, optimize, and scale your AI strategies

    Ready to build?

    Explore powerful AI models and launch your project in just a few clicks.

    Get Started