Coding Horizon

Open Models Are 14x Cheaper Per Task. Here Is The Catch

Every figure quoted in this video, where it came from, and the arithmetic behind it. All prices are US dollars per million tokens unless stated otherwise. All benchmark scores and prices are as published in late July 2026; both move often, so treat the dates as part of the figure.

Sources

What Where
SWE-bench Verified leaderboard, 29 July 2026 https://benchlm.ai/benchmarks/sweVerified
SWE-bench Verified, open weights flags and vendor prices https://llm-stats.com/benchmarks/swe-bench-verified
SWE-bench Pro leaderboard https://llm-stats.com/benchmarks/swe-bench-pro
DeepSeek API pricing, verified 25 July 2026 https://deepseek.ai/pricing
DeepSeek V4 Pro specs and pricing https://openrouter.ai/deepseek/deepseek-v4-pro
DeepSeek V4 licence, parameters, weight size, hardware https://localaimaster.com/models/deepseek-v4
Anthropic API pricing, batch and cache rates https://benchlm.ai/anthropic/api-pricing
Cost per task method and comparison table https://ssojet.com/blog/cheapest-ai-coding-models
Measured cost per task on a real harness https://github.com/Vexp-ai/vexp-swe-bench
Token usage per task type, cache read share https://tokenade.net/en/stats/token-usage-by-task-type
H200 cloud pricing across 33 providers https://getdeploying.com/gpus/nvidia-h200
Kimi K3 pricing and weights status https://www.digitalapplied.com/blog/open-weight-model-wave-july-2026-momentum-tracker

The benchmark

SWE-bench Verified is 500 human validated GitHub issues taken from real repositories. A model is scored on whether the patch it produces makes the repository’s own tests pass, so the score is a resolved rate rather than a similarity judgement.

Scores on SWE-bench Verified, 29 July 2026:

Model Score Weights
Claude Opus 5 96.0 closed
Claude Mythos 5 95.5 closed
Claude Fable 5 95.0 closed
Ornith 1.0 397B 82.4 open
Claude Opus 4.6 80.8 closed
DeepSeek V4 Pro (Max) 80.6 open
Gemini 3.1 Pro 80.6 closed
MiniMax M3 80.5 open
Kimi K2.6 80.2 open
GPT-5.2 80.0 closed
DeepSeek V4 Flash (Max) 79.0 open

Two gaps are quoted from that table:

Gemini 3.1 Pro at 80.6 is level with DeepSeek V4 Pro to the tenth of a point, and GPT-5.2 sits at 80.0, which is why the video says closed models are also clustered at that level rather than the cluster being an open weights problem.

DeepSeek V4 Pro Max and DeepSeek V4 Flash Max are the highest reasoning effort settings of those two models. Lower settings score lower on the same board: V4 Pro (High) 79.4, V4 Flash (High) 78.6.

SWE-bench Pro

SWE-bench Pro is the harder variant, built from longer multi step tasks that need extended reasoning rather than a single localised fix. Scores:

Model Score
Claude Fable 5 80.0
GLM 5.2 62.1 (top open weights)
MiniMax M3 59.0
Kimi K2.6 58.6
DeepSeek V4 Pro Max 55.4

80.0 - 55.4 = 24.6 points. That is the figure behind the claim that the gap widens as tasks get closer to real work: 13.6 points on Verified becomes 24.6 points on Pro for the same pair of vendors.

Model and licence facts

DeepSeek V4 shipped on 24 April 2026 with weights and code under a single MIT licence, which permits commercial use, fine tuning and redistribution.

Kimi K3 is quoted as a counterexample to open meaning cheap: $3.00 input and $15.00 output, with open weights announced for late July 2026 and not published as of the figures above.

Prices per million tokens

Model Input Cache read Output
DeepSeek V4 Pro $0.435 $0.003625 $0.87
DeepSeek V4 Flash $0.14 $0.0028 $0.28
Kimi K3 $3.00 $0.30 $15.00
Claude Opus 5 $5.00 $0.50 $25.00

Anthropic cache reads are 10 per cent of the base input rate, so $5.00 x 0.10 = $0.50. Anthropic’s batch API takes 50 per cent off both directions, giving Claude Opus 5 batch rates of $2.50 input and $12.50 output, with results returned within 24 hours.

Per token multiples between DeepSeek V4 Pro and Claude Opus 5:

Cost per resolved task, fixed attempt model

The comparison uses one fixed token budget per attempt for every model, so the only things that vary are the published prices and the published solve rate. The budget is 50,000 input tokens for repository context and conversation history, and 12,000 output tokens for reasoning and the patch. Cost per resolved task is the cost of one attempt divided by the solve rate, which prices in the fact that a failed attempt still costs a full attempt.

DeepSeek V4 Pro, one attempt:

input   0.050M x $0.435 = $0.021750
output  0.012M x $0.870 = $0.010440
attempt                 = $0.032190
per resolved task = 0.032190 / 0.806 = $0.039938   -> $0.040

Claude Opus 5, one attempt:

input   0.050M x $5.00  = $0.250000
output  0.012M x $25.00 = $0.300000
attempt                 = $0.550000
per resolved task = 0.55 / 0.96 = $0.572917        -> $0.573

0.572917 / 0.039938 = 14.34, quoted as 14.3 times. The solve rate advantage of the more expensive model is already inside that number, because it is the divisor.

Claude Opus 5 on the batch API:

input   0.050M x $2.50  = $0.125
output  0.012M x $12.50 = $0.150
attempt                 = $0.275
per resolved task = 0.275 / 0.96 = $0.286458       -> $0.286

0.286458 / 0.039938 = 7.17, quoted as about 7 times the open model’s standard price.

As a sanity check on the modelled figure, an open harness measured $0.67 per task running Claude Code over a 100 task subset of SWE-bench Verified at a 73 per cent pass rate. That is the same order as the $0.573 the fixed budget model predicts.

What the fixed attempt model hides

Over 100 tickets, using the same attempt cost and the published solve rates:

Claude Opus 5    100 x $0.550000 = $55.00   resolves 96.0
DeepSeek V4 Pro  100 x $0.032190 = $ 3.22   resolves 80.6
saving                           = $51.78
extra unresolved  96.0 - 80.6    = 15.4 tickets
saving per extra failure  51.78 / 15.4 = $3.3624   -> $3.36

At a chosen rate of $100 per hour for engineer time, $100/60 = $1.6667 per minute, so 3.3624 / 1.6667 = 2.02 minutes. That is the two minute figure: each extra failure has to be diagnosed and fixed in about two minutes before the saving is gone. The $100 per hour rate is a stated assumption, not a measured figure; the conclusion scales linearly with whatever rate you substitute.

The 15.4 extra failures are not a random sample of the backlog. The SWE-bench Pro figures above show the same vendors separating by 24.6 points rather than 13.6 once tasks get longer and need multi step reasoning, so the failures concentrate in the harder tail.

Real token volumes

A coding task in the SWE-bench class consumes between 1 million and 3.5 million tokens per task in measured agentic runs, including retries and self correction. The same task run twice can differ by up to 30 times in total tokens, and higher token consumption does not correlate with higher accuracy.

In a production agent workflow, 82 to 83 per cent of input tokens were cache reads: context the agent has already paid to read, being re read as the conversation grows.

The 62,000 token attempt used in the fixed model is therefore optimistic by between 16 and 56 times against that measured range.

Cost per resolved task at real volumes

Modelling one task as 1.5 million input tokens with 83 per cent of them served as cache reads, plus 100,000 output tokens:

cache reads  1.5M x 0.83 = 1.245M
new input    1.5M x 0.17 = 0.255M
output                   = 0.100M
total                    = 1.600M

Those are the token volumes shown on screen: cache reads 1.25M, new input 0.26M, output 0.10M. As a share of every token processed, cache reads are 1.245 / 1.600 = 77.8 per cent, which is the sense in which cache reads are the largest single line on the bill.

DeepSeek V4 Pro:

1.245M x $0.003625 = $0.004513
0.255M x $0.435    = $0.110925
0.100M x $0.870    = $0.087000
total              = $0.202438
per resolved task  = 0.202438 / 0.806 = $0.251164  -> $0.25

Claude Opus 5:

1.245M x $0.50  = $0.622500
0.255M x $5.00  = $1.275000
0.100M x $25.00 = $2.500000
total           = $4.397500
per resolved task = 4.3975 / 0.96 = $4.580729      -> $4.58

4.580729 / 0.251164 = 18.24, quoted as 18.2 times. The multiple grows from 14.3 to 18.2 once cache reads are priced in, because DeepSeek’s cache read discount is far deeper than Anthropic’s: 99.2 per cent off its own input rate against 90 per cent off.

Redoing the 100 ticket comparison at these volumes:

DeepSeek V4 Pro  100 x $0.202438 = $ 20.24   resolves 80.6
Claude Opus 5    100 x $4.397500 = $439.75   resolves 96.0
saving                           = $419.51
saving per extra failure  419.51 / 15.4 = $27.24

At $100 per hour, 27.24 / 1.6667 = 16.3 minutes, quoted as 16 minutes. The headroom per failure grows from two minutes to sixteen purely because the task got bigger, which is the reversal the video describes.

Running the weights yourself

DeepSeek V4 Pro’s BF16 weights are about 3.2 TB on disk. Quantised to four bits they are about 800 GB.

An eight GPU H200 node provides 8 x 141 GB = 1,128 GB of VRAM, so the full precision weights do not fit on a single node while the four bit version does.

H200 cloud pricing varies widely by provider. On demand rates run from about $2.30 to $5.99 per GPU hour, with a tracked median around $4.22. An eight GPU H200 node is quoted at $34.32 per hour.

$34.32 x 24 x 30 = $24,710.40 per month, running continuously

Against DeepSeek’s own API price for a task at the real volumes above:

24,710.40 / 0.202438 = 122,065 tasks per month   -> about 122,000
122,065 / 30         = 4,069 tasks per day       -> about 4,000

Below roughly 4,000 tasks a day, paying DeepSeek’s API is cheaper than renting the hardware to serve the same model, before any staff time to operate it.

Caveats worth knowing