DeepSeek Made Its 75% Price Cut Permanent. For LatAm Founders, the AI Cost Floor Just Collapsed.

Six months ago, the economics of building a high-volume AI product in Brazil looked like this: $15 per million input tokens to Claude Opus, $10 to GPT-4 equivalent, and margin pressure that made any consumer-facing AI product with high inference volume essentially impossible to price for the Brazilian market. On May 22, DeepSeek changed those numbers permanently.

DeepSeek V4 Pro's permanent pricing — $0.435 per million input tokens, $0.87 per million output — isn't a promotional rate. It's the new floor. The model runs 1.6 trillion parameters, 49 billion activated per token, on a 1 million token context window. It scores 80.6% on SWE-bench Verified, within striking distance of Claude Opus 4.7 and GPT-5.5 on real-world coding tasks. MIT-licensed and openly available on Hugging Face. NIST's independent CAISI evaluation in May 2026 passed it on every enterprise safety benchmark.

The LatAm calculus changes here specifically. The $15-per-million Claude Opus pricing was a real barrier for startups building in markets where B2B software prices are compressed relative to North America. A Brazilian legaltech automating high-volume contract review, a healthtech processing clinical notes at scale, a fintech running credit underwriting inference across millions of applications — all of these products required either pricing structures that didn't work for the Brazilian market or model quality compromises that didn't work for regulated industries. At $0.435, both constraints collapse simultaneously.

The same workloads that were economically marginal at $15 are economically comfortable at a 97 percent lower cost. This is not an incremental efficiency improvement. It's a threshold crossing — the difference between a product category that can exist in a market and one that can't.

The defensibility question shifts as a result. When the best open-weight frontier model costs one thirty-fourth of the best closed model and performs within 3 to 5 percentage points on benchmark tasks, model selection stops being a strategic decision and becomes a cost optimization decision. The moat was never in the model access. The permanent price collapse just makes that fact impossible to rationalize around.

For LatAm founders, the signal is precise: the intelligence commodity is now priced for your market. The barrier isn't compute cost anymore. The question is what proprietary data, what domain context, and what distribution advantages you bring to a layer where the intelligence itself is approaching free — and whether you're building for that world or still pricing for the one that existed six months ago.