AI
The Labs Wrote Their Own Rulebook This Week While Open Models Rewrote the Price Tag
DeepSeek released V4.1 Flash 1017 on October 17 with MIT licensed weights and a brand new Search and Fetch API for agents priced at 75 cents per 1,000 calls. The model posts a 77.0 on the DeepSWE v1.1 coding benchmark, ahead of Claude Opus 5 and GPT 5.6 Sol on that test. Anthropic quietly pitched Capitol Hill a draft AI safety bill that would let Washington halt a frontier model only in narrow catastrophic cases, with independent evaluators registered at the Commerce Department and a federal judge as the only off switch. GitHub set November 16 as the retirement date for three Copilot models, while Cursor shipped version 0.45 with dependency mapping across as many as 50 linked repositories. The thread running through the week is easy to spot. Capability keeps racing forward, the builders are writing the rulebook themselves, and the cost of intelligence keeps falling.
The DeepSeek release is the week's clearest example of the open model strategy working. V4.1 Flash 1017 is described as a minor update to the Flash model from September, but the numbers DeepSeek published show gains on all 11 benchmarks it reports, led by an 11.1 point jump on Terminal-Bench 4.0 to 42.3. Task execution gets faster, outputs get shorter, and API pricing holds at 15 cents per million input tokens and 60 cents per million output tokens off peak. The weights are on Hugging Face and ModelScope under the MIT license, and the model carries a 1 million token context window. DeepSeek opened its Search and Fetch APIs to every developer, powered by an in house search engine the company says serves hundreds of millions of requests a day. Each endpoint costs 75 cents per 1,000 calls, and calls routed through server side tools in DeepSeek Harness, Claude Code or Codex stay free until per request billing starts on November 1. One caution stays attached. Every benchmark figure here is DeepSeek's own, produced on harnesses it chose, and independent reproduction is still pending, so treat the leaderboard claims as the company's story until third parties confirm them.
While DeepSeek pushed capability outward, Anthropic pushed the rulebook forward. Politico reported on October 16 that Anthropic representatives had quietly pitched Capitol Hill staff on a draft bill to regulate frontier AI safety, citing three people familiar with the effort, and Reuters confirmed the details a day later. Washington could halt a model's release or use only when it poses a foreseeable imminent danger of more than 1 billion dollars in damages, or of killing or seriously injuring more than 50 people. The danger has to come from one of three sources, the creation or release of a biological weapon, a model evading its developer's controls, or a cyberattack. Independent evaluators would register with the Commerce Department and test frontier models on a schedule, and evaluators would have 72 hours to report an imminent enumerated danger to Commerce. From there only Commerce could ask a federal judge in Washington to order a suspension, and the draft explicitly keeps every other agency and officer out of that decision. An Anthropic spokesperson told Reuters the draft reflects proposals in the company's Advanced AI Framework from earlier this year. The industry is preempting heavier regulation by writing a narrow, court supervised version itself. A 1 billion dollar tripwire keeps everyday model releases and open weights well clear of government review, while the judicial backstop gives policymakers a real answer when someone asks what happens in a true emergency. Whether Congress takes the draft up is the open question.
The coding tool wars delivered the week's most concrete shakeup. GitHub's changelog announced that three models leave Copilot on November 16, Claude Opus 4.8, Claude Opus 5 and GPT 5.6 Terra, with Claude Opus 5.5 and GPT 6.1 Sol or GPT 6 Luna as the suggested replacements. Individual subscribers on annual plans keep Opus 4.8, but Opus 5 and Terra leave every tier. Business and Enterprise customers get the replacements switched on automatically. The pattern behind the cut is consolidation. GitHub retired ten models since the start of October, and the menu inside the world's most used coding assistant is shrinking toward newer generations while the open weight world expands in the other direction. Cursor, meanwhile, is moving the other way. Version 0.45, released October 17, adds dependency mapping across as many as 50 linked repositories in one workspace, tracing function calls and type definitions for big engineering teams. The closed platforms are pruning their lineups while the open ecosystem keeps adding options.
Step back and the week's shape is clear. The frontier keeps getting cheaper to run, the people building it are drafting the laws that will govern it, and the tools developers use every day are being reshuffled toward whatever proves most useful. For the reader the practical notes are simple. If you build with agents, this week's releases cut your bill and widen your options. If you follow the policy debate, the most important document of the week came from a lab. And if you write code for a living, audit your pinned model names before November 16, because the menu is about to change. Intelligence is getting cheaper, more open, and more governed all at once, and that combination favors the people who actually ship.
Quick answers
What is this story about?
DeepSeek released V4.1 Flash 1017 on October 17 with MIT licensed weights and a brand new Search and Fetch API for agents priced at 75 cents per 1,000 calls. The model posts a 77.0 on the DeepSWE v1.1 coding benchmark, ahead of Claude Opus 5 and GPT 5.6 Sol on that test. Anthropic quietly pitched Capitol Hill a draft AI safety bill that would let Washington halt a frontier model only in narrow catastrophic cases, with independent evaluators registered at the Commerce Department and a federal judge as the only off switch. GitHub set November 16 as the retirement date for three Copilot models, while Cursor shipped version 0.45 with dependency mapping across as many as 50 linked repositories. The thread running through the week is easy to spot. Capability keeps racing forward, the builders are writing the rulebook themselves, and the cost of intelligence keeps falling.
Why does this story matter?
Step back and the week's shape is clear. The frontier keeps getting cheaper to run, the people building it are drafting the laws that will govern it, and the tools developers use every day are being reshuffled toward whatever proves most useful. For the reader the practical notes are simple. If you build with agents, this week's releases cut your bill and widen your options. If you follow the policy debate, the most important document of the week came from a lab. And if you write code for a living, audit your pinned model names before November 16, because the menu is about to change. Intelligence is getting cheaper, more open, and more governed all at once, and that combination favors the people who actually ship.
Sources
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.