Open AI’s New Race: More Useful Work for Less

James Turner

Here's a number to hold onto: about one point seven cents. That's the effective price of a million input tokens on DeepSeek's official endpoint, once you account for how often requests hit the cache. And that number comes attached to a model that just got noticeably better. This is 48-Hour AI, I'm James Turner, and we're covering October 16 to 19, 2026.

James Turner

Last time, we talked about the messy side of agents. Tencent's fleet poking at Amap, a hidden recording device in an OpenAI conference room, Google's Harness Squared. This weekend the story flipped. Instead of agents misbehaving, open weight releases are building the machinery agents actually need, and doing it cheaply.

James Turner

Start with DeepSeek V4.1 Flash 1017, which DeepSeek is calling the Pareto update. It landed October 17 as an in place snapshot on the existing deepseek flash API endpoint, so if you already built on it, nothing breaks. You just get a different model underneath. The MIT licensed weights went up on Hugging Face and ModelScope the same day.

James Turner

The architecture is a causal encoder decoder with a 552 billion parameter backbone. But only 8 billion parameters activate per input token during prefill, and 16 billion per output token during decode. The context window is 1,048,576 tokens, and it was pretrained jointly on text and images.

James Turner

So what changed? DeepSeek says five more weeks of reinforcement learning with self feedback. They call the approach wish driven programming. Instead of rewarding the model for answering one prompt well, they reward it for taking a high level goal and finishing a multi step task from start to end. The showcase projects were a 3D reconstruction of the Forbidden City, a Godot based recreation of Portal 2, and a browser racing game. Those are DeepSeek's demos, so treat them as a statement of intent, not an independent test.

James Turner

Now, the Pareto claim. A Pareto improvement means you got better on one axis without getting worse on another. Here, benchmarks went up while outputs got shorter. That matters more than it sounds, because in an agent loop, output tokens are what dominate runtime. The model writes, calls a tool, reads the result, writes again, over and over. If each turn is shorter, you save money and time on every single lap.

James Turner

The numbers, from DeepSeek's own table. Terminal Bench 4.0, run in the Claude Code harness, jumped 11.1 points, from 31.2 to 42.3. That's the biggest gain in the table. DeepSWE version 1.1 hit 77.0, ahead of Claude Opus 5 at 74.0, GPT 5.6 Sol at 73.0, K3 at 67.5, and GLM 5.3 at 66.9. ProgramBench went from 20.3 to 29.1. Apex Bench 1.1 went from 64.6 to 72.1. And Humanity's Last Exam without tools moved from 36.8 to 42.6.

James Turner

Back to that one point seven cents. Off peak list pricing is 15 cents per million input tokens, 60 cents per million output, and cache reads cost three tenths of a cent. Peak hours double the first two, to 30 cents and a dollar twenty. But DeepSeek's official endpoint reports a 91.5 percent cache hit rate, and when you blend that in, effective input cost lands around one point seven cents per million. Agents resend huge amounts of the same context every turn, so caching is where the savings really live.

James Turner

A few caveats if you want to run it. There's no Jinja chat template, so you need DeepSeek's official encoding dot py or the deepseek recipe bindings, in Rust or Python. They recommend temperature 1.0, top p 1.0, and a max output of at least 256K. And local serving of the full 552 billion model takes hundreds of gigabytes of memory. Roughly 320 gigabytes for a Q4 K M quantization. That's open, but it's not laptop open.

James Turner

DeepSeek also opened its internal Search and Fetch APIs, at 75 cents per thousand calls per endpoint. They're free through October 31 at 16:00 UTC, if you call them through DeepSeek Harness, Claude Code, or Codex.

James Turner

Now, DeepSeek isn't the only open release worth your attention. On October 16, Meta shipped Muse Spark Open, an 816 billion parameter open weight model built for coding agents and long tasks.

James Turner

The trick is sparsity. It's a mixture of experts with 256 expert networks, and only 8 of them fire per request. That's 42 billion active parameters. Picture 256 specialists in a building, where only eight show up for each job. You get big model quality without paying for all 816 billion every time.

James Turner

It has a 1 million token context window and takes text, code, images, audio, and video natively. And here's the detail I like: it can run locally on two Apple M5 Ultra Mac Studios using Apple's MLX framework. No cloud. If you'd rather hit an API, it's on Groq, Together AI, and OpenRouter, and you can fine tune with PyTorch. One caution. The license is the Muse Model License, so check it for commercial limits before you build a business on it.

James Turner

Same day, Tencent released Hy Memory 2.0, and I think this one is the most interesting design idea of the weekend. The problem is the memory write path. Every time an agent finishes a conversation or a coding session, something has to decide what to extract, how to index it, and what to keep. Usually you hand that job to your big, expensive generalist model.

James Turner

Tencent's answer is a small specialist whose only job is remembering. It's called Hy Memory 30B A3B, 30 billion parameters total, about 3 billion active. It handles chat, work, and code memories, and a management layer called MemHub sits on top. It's trained with a method described in arXiv paper 2610.22196, called attributed reinforcement learning. And Tencent claims, and this is Tencent's claim, that it edges Claude Opus 5 and DeepSeek V4.1 Flash on long term memory retrieval benchmarks. The big model thinks. The small model takes notes.

James Turner

Also on October 16, NVIDIA released Nemotron 3.5 Super VL, a multimodal reasoning model that scored 26.5 on the Artificial Analysis Intelligence Index while using substantially fewer output tokens than its predecessor. There's a smaller sibling, Nemotron 3.5 Bolt 74B A9B, plus near lossless NVFP4 quantized weights.

James Turner

Notice the pattern. DeepSeek trims output length. NVIDIA trims output tokens. Meta activates 42 billion of 816 billion. Tencent hands memory to a 3 billion active specialist. Nobody here is chasing raw size. They're all chasing cost per useful task.

James Turner

One more item. Kaggle launched a competition on October 16 called Name That LLM. You get unlabeled passages, and you identify which of 18 frontier models wrote each one. It's a game, but it points at provenance and benchmark auditing, and at how recognizable a model's fingerprint is.

James Turner

Two dates before you go. OpenRouter's 50 percent discount on ElevenLabs models expires October 19. And DeepSeek's free Search and Fetch window closes October 31 at 16:00 UTC. If you're going to test that new endpoint on real agent work, do it before the meter starts. I'm James Turner, and I'll see you in 48 hours.