DeepSeek V4 Pro beats GPT-5.5 Pro on precision — and on cost, the gap is brutal

DeepSeek V4 Pro has edged out OpenAI’s GPT-5.5 Pro on precision in a comparative assessment reported by Runtime Wire. The headline number is narrow — DeepSeek scored 38.0 to GPT-5.5 Pro’s 33.0 on four text tasks judged by a retired xAI model — but the cost data attached to the story tells a far more consequential one.

The benchmark comes from SwellJoe, a Hacker News user who runs a vulnerability scanning benchmark called “Will It Mythos.” He tested GPT-5.5 Pro against DeepSeek V4 Pro, Anthropic’s Opus 4.8, and MiMo 2.5 Pro. The results, posted to a discussion thread on June 8, are stark. GPT-5.5 Pro blew through a $100 budget halfway through the benchmark, averaging $22 per case. DeepSeek V4 Pro cost about a dollar for the entire benchmark — roughly a dime per case. Opus 4.8 was an order of magnitude cheaper than GPT-5.5 Pro and about 30% cheaper than GPT-5.5 non-Pro. DeepSeek and MiMo were two orders of magnitude cheaper.

“I can’t come up with a use case where I can rationally spend ~31 times what Opus costs to use GPT 5.5 Pro,” SwellJoe wrote. “I won’t be doing any more benchmarking with it.”

The cost disparity is not a bug in the benchmark. It is a structural feature of the market. DeepSeek’s API pricing, available directly at platform.deepseek.com, undercuts OpenAI and Anthropic by factors of 100 to 200 for equivalent token volumes. One commenter, slopinthebag, reported using roughly 16 million input tokens on DeepSeek V4 Pro in a single day — 15 million of them cache hits — and spending $0.47. The same volume on GPT-5.5 Pro would run well over $100.

The precision gap itself is real but narrow. DeepSeek V4 Pro scored 38.0 to GPT-5.5 Pro’s 33.0. The judge was grok-4-1-fast-non-reasoning, a model xAI retired on May 15. Requests to that endpoint now silently route to grok-4.3, a 5x more expensive model. That methodological wrinkle matters, but it does not invalidate the relative ranking. The tasks were fresh — generated on the fly so neither model could prepare — and the judge scored each response blind.

The more interesting finding is what the precision metric measures. The assessment highlights instruction adherence, schema matching, and edge-case resolution. These are exactly the capabilities that matter for institutional use: financial reporting, legal analysis, data processing, structured output generation. GPT-5.5 Pro lost points on “avoidable deviations” — output that was correct in content but wrong in format or constraint compliance. DeepSeek V4 Pro followed the spec more closely.

This is not a new pattern. Several Hacker News commenters report similar experiences. “DeepSeek 4 pro is insanely good for the price,” one wrote. Another, embedding-shape, noted that the judge model itself was retired, but the relative ordering held. A third, andai, flagged the grok migration: “TFA was published today, which implies grok-4.3 was used.”

The implications for the AI economy are straightforward. OpenAI and Anthropic have built their pricing models on the assumption that customers will pay a premium for frontier capability. DeepSeek is testing that assumption with a model that is competitive on quality and dramatically cheaper. If the cost gap persists — and there is no reason to expect it to narrow — the market for API-based coding assistants will bifurcate. High-margin enterprise contracts will stay with the US labs. Price-sensitive developers and startups will migrate to DeepSeek.

One commenter, natrys, argued that frontier labs would rather lose the price-sensitive segment than lower API prices. “They are getting the bag in the enterprise segment,” he wrote. “Those clients aren’t ditching them for DeepSeek.” That is plausible for now. But the enterprise segment is not infinite. As DeepSeek improves its performance on the margin — and as more developers build tooling around its API — the cost argument becomes harder to ignore.

The security concern remains. Several commenters raised the issue of shipping proprietary code to a lab under an adversarial government. That is a real constraint for some use cases. But it is not a universal one. Many developers already use DeepSeek for non-sensitive tasks and are satisfied. One commenter, slopinthebag, described using DeepSeek exclusively: “I have zero usage anxiety unlike when I was using subscription plans.”

The speed advantage is less talked about but may be equally important. DeepSeek V4 Pro is fast — it does not spend minutes reasoning on basic tasks. That matters for developer flow state. “It’s really quick,” slopinthebag wrote. “Doesn’t spend too much time reasoning even on ‘max.’” The flash model is also strong.

The benchmark is small — four tasks, one judge, one user’s harness. It is not a comprehensive evaluation. But it is a real-world test conducted by a developer with no incentive to favor either model. That gives it weight that synthetic benchmarks lack.

The takeaway for AI builders is simple. The cost-quality frontier is shifting. DeepSeek V4 Pro is not just cheaper — it is competitive on the metrics that matter for structured coding tasks. The gap with GPT-5.5 Pro is narrow on precision and enormous on cost. That combination is unsustainable for the incumbent pricing model.

The question is not whether DeepSeek will take market share. It already is. The question is whether OpenAI and Anthropic will respond by cutting API prices, improving their models on precision, or both. If they do not, the migration will accelerate.