DeepSeek V4.1 Flash vs OpenAI Pricing
What exactly is Dario Amodei's slowdown protecting you from when the thing doing it is a 510 GB file that anyone can download for free? The mechanism behind this isn't a court order or a policy shift; it's a public download link that renders the entire proposal to pace the industry completely worthless. DeepSeek released a model named V4.1 Flash just two days before an essay by Amodei was even published, and the weights went straight onto Hugging Face under an MIT license permitting commercial use and modification without royalties owed to anybody. On its own model card, V4.1 Flash outperforms OpenAI's GPT-5.6 Sol and Anthropic's Claude Opus 5 on four of five difficult agentic benchmarks. It hits 74.2 on DeepSWE v1.1, 54.8 on AutomationBench, 31.8 on Agent's Last Exam, and 88.1 on CyberGym, which is identified as the cybersecurity benchmark. So the exact capability Amodei is asking the industry to pace is now a PUBLIC DOWNLOAD LINK that nobody can UNDOWNLOAD. Now look at the price because the story gets brutal. OpenAI priced GPT-6 Astra at $10 per million input tokens and $50 per million output tokens starting September 3. DeepSeek priced V4.1 Flash seven days later at 15 cents and 60 cents off-peak, with cached inputs at a third of a cent. And the reasoning benchmarks tell the same story over a longer window. Achieving an ARC-AGI-1 score of 87.5% cost approximately $4,560 per task in December 2024 but about 30 cents twenty months later. DeepSeek's earlier Flash build achieved an 89.0% score for 2 cents. On OpenDesign's arena on September 9, DeepSeek scored 1.5 points behind OpenAI's newest model while costing $0.023 per task compared to $1.61. That 1.5 point gap is the size of the current lead, and Amodei's proposal limits the industry slowdown to exactly that margin. If they slow down by more than 1.5 points, Chinese projects pull ahead. The distillation crackdown proposed by Amodei does not apply because DeepSeek published weights directly rather than distilling them. There is no theft to prosecute and no copy to trace when you give it away on purpose. DeepSeek's Pro model reached 80.6% on SWE-bench Verified in April at roughly one-thirty-fourth the price of American flagships. This has been happening all year. The frontier still wins the hardest work. On Terminal-Bench 4.0, OpenAI reported a score of 57.9 against DeepSeek's 31.2. Those four benchmark wins cited come from DeepSeek's own model card without replication under a shared protocol. Four benchmark wins cited come from DeepSeek's own model card without replication under a shared protocol. If you look at it that way, American labs still hold the dangerous end of the curve on the hardest work according to the author. But a speed limit only binds companies subject to American court jurisdiction. Everyone else just downloads models without restriction. What exactly is Dario Amodei's slowdown protecting you from? The answer lies in the fact that China is rendering Dario Amodei's AI slowdown proposal ineffective through a simple file available for free download. A 510 GB file available for free download is the mechanism causing this. DeepSeek released a model named V4.1 Flash two days before an essay by Amodei was published. The weights for V4.1 Flash were uploaded to Hugging Face under an MIT license permitting commercial use and modification without royalties. V4.1 Flash outperforms OpenAI's GPT-5.6 Sol and Anthropic's Claude Opus 5 on four of five difficult agentic benchmarks. Specific benchmark results include DeepSWE v1.1 at 74.2, AutomationBench at 54.8, Agent's Last Exam at 31.8, and CyberGym at 88.1. OpenAI priced GPT-6 Astra at $10 per million input tokens and $50 per million output tokens starting September 3. DeepSeek priced V4.1 Flash seven days later at 15 cents and 60 cents off-peak, with cached inputs at a third of a cent. Achieving an ARC-AGI-1 score of 87.5% cost approximately $4,560 per task in December 2024 but about 30 cents twenty months later. DeepSeek's earlier Flash build achieved an 89.0% score for 2 cents. On OpenDesign's arena on September 9, DeepSeek scored 1.5 points behind OpenAI's newest model while costing $0.023 per task compared to $1.61. Amodei's proposal limits the industry slowdown to the size of the current lead, which the author calculates as 1.5 points. V4.1 Flash contains 552 billion parameters but activates only 8 billion per word processed. V4.1 Flash was trained on 45 trillion tokens. The distillation crackdown proposed by Amodei does not apply because DeepSeek published weights directly rather than distilling them. DeepSeek's Pro model reached 80.6% on SWE-bench Verified in April at roughly one-thirty-fourth the price of American flagships. On Terminal-Bench 4.0, OpenAI reported a score of 57.9 against DeepSeek's 31.2. Speed limits proposed in the plan only bind companies subject to American court jurisdiction. Other entities can simply download models without restriction. What exactly is Dario Amodei's slowdown protecting you from?
