GPT-6 Sol Pro vs DeepSeek V4 Pro: A Flat Rate Against a Timetable

GPT-6 Sol Pro and DeepSeek V4 Pro are the two cheapest credible ways to run a serious reasoning workload through an API right now, and they are priced on completely different principles. GPT-6 Sol — the vendor's model that the gpt-6-sol-pro configuration runs, generally available since September 22, 2026 — costs $2.00 per million input tokens and $10.00 per million output tokens, at every hour of every day. DeepSeek V4 Pro, the open-weights flagship whose 0813 checkpoint reached general availability on August 13, 2026, costs $0.66 and $1.98 for most of the week and exactly double that for two narrow windows on weekday mornings. Any comparison that quotes a single rate is quoting half a price.

The schedule is the story. Everything else about this matchup — the index scores, the context windows, the licence terms — is downstream of the fact that one of these two vendors has decided its list price is a function of the clock.

What each price line actually is

GPT-6 Sol's rate card is one row. Input $2.00, cached input $0.20, cache write $2.50, output $10.00 per million tokens, with a long-context reprice above an input threshold that applies to the whole request once crossed. There is no time dimension and no promotional window. OpenAI cut the predecessor model to these numbers on September 22 and described them as permanent.

DeepSeek's rate card is a table. As read on the vendor's own pricing page:

• Off-peak input — $0.66 per million tokens, cache miss; $0.022 per million tokens, cache hit

• Off-peak output — $1.98 per million tokens

• Peak input — $1.32 per million tokens, cache miss; $0.044 per million tokens, cache hit

• Peak output — $3.96 per million tokens

• Peak windows — 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays

• Everything else — off-peak, including the whole weekend, a rule that took effect at midnight Beijing time on August 23, 2026

Seven hours of peak out of a 168-hour week, five days out of seven, is roughly 21% of the calendar at the higher rate and 79% at the lower one. So the honest single-number summary of DeepSeek V4 Pro is not $0.66 and not $1.32 — it is somewhere between, and where depends on your traffic's shape, not on the vendor's generosity.

The repricing this table came out of

It is worth knowing how recent this structure is, because almost every comparison page still quotes the number it replaced. On May 22, 2026, DeepSeek converted a limited-time discount on V4 Pro into a permanent price — a cut of about 75%, to roughly $0.435 input and $0.87 output per million tokens.

On August 17, 2026 at midnight Beijing time, that was reversed and then some. The peak/off-peak system went live with an average increase across the affected rows of about 3.5 times and a maximum of 12 times — the 12× landing precisely on the cache-hit input rate, which went from ¥0.025 to ¥0.30 per million tokens at peak. Cache-miss input rose about 200% and output about 350%. Six days later the weekend carve-out was added, which clawed some of it back for anyone whose batch jobs could be moved.

Two consequences follow. First, any page telling you DeepSeek V4 Pro is a sub-dollar model is quoting a rate card that has been superseded twice. Second, the shape of the increase is a signal about what DeepSeek is optimising for: the largest proportional rise is on cache hits, which is the rate that matters most to exactly the workload — long agent loops re-sending a stable prefix — that made the model attractive in the first place.

The index number everyone is quoting is retired

Artificial Analysis is the neutral harness for this comparison, and it has re-versioned its Intelligence Index several times through 2026. On the current board, GPT-6 Sol scores 48 at max effort at $1.06 per index task. DeepSeek V4 Pro, on the 0813 checkpoint at max effort, scores 36 at $0.67 per index task.

The figure that circulates for DeepSeek is 53. That number is real and it was published — on an earlier revision of the index, before a re-versioning that moved most models' scores. Comparing a 53 taken from a pre-v4.3 article against a current 48 is comparing a retired measurement with a live one, and it produces the single most common error in this matchup: the claim that the two models are within a point of each other. On the current index they are twelve points apart, and the price gap per completed task is not the 3× the headline rates suggest but closer to 1.6×, because DeepSeek spends more tokens getting to its answer.

Two other independent data points belong beside those, both labelled for what they are. Vals AI, running an isolated harness, puts DeepSeek V4 Pro at 96.40% on SWE-bench Verified at $0.103 per test — second on that board to Claude Opus 5's 97.00%, and cheap per test by a wide margin. Artificial Analysis's coding-agent board shows 14% on Terminal-Bench 4.0 against GPT-6 Sol's 44%. The first number says DeepSeek is competitive on one narrow, well-defined coding task. The second says it is not competitive on agentic terminal work. Both are true and they are measuring different things.

Where the two models are structurally different, not just priced differently

• Licence — GPT-6 Sol is a closed API with no weights; DeepSeek V4 Pro is MIT-licensed with public checkpoints on Hugging Face, including base variants for continued pretraining

• Maximum output — GPT-6 Sol 128,000 tokens vs DeepSeek V4 Pro 384,000 tokens, three times the ceiling at a fifth of the output rate

• Context window — GPT-6 Sol 1,050,000 tokens vs DeepSeek V4 Pro 1,048,576 tokens, effectively identical

• Modalities — both text in and text out; neither accepts images on the API

• Long-context pricing — GPT-6 Sol reprices the whole request above its input threshold; DeepSeek V4 Pro's peak/off-peak schedule is the only price structure it has

• Scale — DeepSeek V4 Pro is a 1.6-trillion-parameter mixture-of-experts model with 49 billion active parameters per token; OpenAI does not publish a comparable figure for GPT-6 Sol

One note on that parameter count, because it is a frequent source of confusion: the 1.7-trillion figure that appears in some listings is the Hugging Face package size with the optional DSpark speculative-decoding module included, not a wider target model. The published architecture is 1.6T total. Subtracting the rounded badges to infer the draft module's size is not a valid calculation, and neither is reading the larger number as evidence that the model grew.

The withdrawal that is the real argument for open weights

On September 9, 2026, DeepSeek announced it was phasing out V4 Pro, and that from 04:00 UTC on September 14 every request to the deepseek-v4-pro model string would be routed to DeepSeek V4.1 Flash and billed at Flash rates, continuing until a V4.1 Pro model shipped. The model string would still answer. The model behind it would not be the one anyone had benchmarked.

Developers objected, and the redirect was withdrawn. DeepSeek's changelog now states that V4 Pro API service continues past September 14 with billing unchanged, and the pricing table still lists deepseek-v4-pro as a live row with its own rates and no retirement tag. The earlier announcement page still contains the routing language, so the two documents disagree — and the live pricing page is the one that describes what actually happens when you call it.

That episode is the strongest argument in this entire matchup, and it is not a price argument. A closed model can be repriced, rerouted or retired on a vendor's schedule, and the only recourse is to migrate. A model with MIT weights on a public repository can be served from your own infrastructure, which means the checkpoint you evaluated is the checkpoint you keep serving regardless of what the vendor's roadmap says six weeks later. DeepSeek demonstrated the risk and supplied the mitigation in the same fortnight.

The honest counterweight is that self-hosting a 1.6-trillion-parameter model is not a decision most teams make casually. It is a real option with a real fixed cost, not a free hedge.

Where a routing layer changes the arithmetic

DeepSeek V4 Pro is on OrcaRouter's catalogue. Both the current checkpoint and the base model string resolve — deepseek/deepseek-v4-pro-0813 and deepseek/deepseek-v4-pro — at the off-peak list rates with the peak multiplier applied on the same UTC schedule, a 1,048,576-token context window, a 384,000-token output cap, and p50 time to first token of 3.56 seconds as measured on our own model page. Provider list pricing passes through with no markup on top, which in this case means the peak/off-peak boundary lands on your invoice at the same moment it lands on DeepSeek's.

GPT-6 Sol is not one of our routes, so nothing here is a claim about its price through us.

The schedule is what makes a single endpoint worth having here specifically. A workload that can be shifted is worth twice as much as one that cannot, and shifting it is a routing decision, not a migration — the same key, the same request shape, a different hour or a different model behind it. Pairing DeepSeek V4 Pro with DeepSeek V4.1 Flash on one endpoint is the sharper version of that: the Flash model is 4.4× cheaper on input and 3.3× cheaper on output than V4 Pro's off-peak rate, on the same context window and the same peak windows, which makes it the natural destination for the traffic that does not need the larger model at all. Automatic failover matters for the same reason it always does — the cheap tier your pipeline is built on should not be a single point of failure.

The decision rule

• Batch and offline work that can be scheduled — DeepSeek V4 Pro, and schedule it. Ninety-five percent of the calendar at $0.66 and $1.98, with weekends entirely off-peak, is the cheapest credible rate on this board by a factor of three on input and five on output.

• Latency-sensitive interactive traffic that lands in the peak windows — GPT-6 Sol. A flat $2.00 and $10.00 is more expensive than DeepSeek's off-peak rate and cheaper than nothing, but it is knowable in advance, and for two hours a day it is the cheaper of the two on output.

• Agent loops with a large stable prefix — run the cache-hit arithmetic before choosing. DeepSeek's cache-hit input is $0.022 off-peak, which is a ninth of GPT-6 Sol's $0.20, but it is also the row DeepSeek raised twelvefold at peak. The advantage is real and it is conditional on the clock.

• Anything regulated, or anything you cannot afford to have rerouted — DeepSeek V4 Pro, self-hosted from the MIT checkpoint. The September withdrawal is the case study, and it is a better reason to accept the fixed cost than any per-token comparison.

• Anything that needs the top of the agentic range — GPT-6 Sol. Twelve index points and thirty points of Terminal-Bench are not a rounding difference, and DeepSeek's own vendor-reported numbers do not close them.

The flat rate and the timetable are both honest pricing designs. They just answer different questions: one tells you what a token costs, the other tells you what a token costs when you need it.