October 27, 2026 — A Cheaper Computer-Use Model, With Asterisks

Host0:33

One dollar per million input tokens. Five dollars per million output. That's what OpenAI charges for GPT six point one Terra, which shipped Monday. OpenAI says that's fifty percent below both GPT five point six Terra and its own GPT six point one Sol. Terra is the mid-tier model, pitched on visual reasoning, computer use, coding and writing. Here's the interface. You send it text and images, and it hands back text, tool calls or structured outputs. The API name is gpt dash six point one dash terra. It's in the OpenAI API, ChatGPT Work, Codex, GitHub Copilot and Amazon Bedrock. Open weights and license? Unknown so far. OpenRouter lists a one million fifty thousand token context window and up to one hundred twenty-eight thousand output tokens. AWS says up to one million input tokens. Confirm against OpenAI's docs. Terra pairs visual reasoning with computer-use actions. An agent reads the screen, then carries out multi-step desktop tasks. Now the numbers, and these are OpenAI's own. On Agents' Last Exam at medium effort, Terra beats GPT five point six Terra by fifteen points, at about a third of the cost. On Terminal-Bench four point zero at max effort, it reports fifty point three percent, against twenty-one point five. Artificial Analysis, listed on OpenRouter, shows fifty-five point six. Nobody explains the gap, though different setups are possible. The launch post's example task is reviewing an animated film shot by shot and writing a time-coded restoration log. That's either a benchmark or somebody's weekend. The catch. OpenAI says Sol and Astra stay better for the hardest coding jobs. Ted Sanders of OpenAI, quoted by The Register, said Terra is probably better at technical communication, similar on vision and computer use but much faster, and a bit worse on most other things. Verdict. Try it on high-volume, visually grounded work and bounded computer-use tasks, where cost and speed matter. Then measure cost per task, retries and failure recovery included, because that's where a cheap token price can stop being cheap. For complex coding or anything high-stakes, test against your current model before you switch.

3:08

Terra makes hosted agents cheaper. Hob asks what if the model never leaves your laptop. Hob is an MIT-licensed desktop assistant for Mac, Windows and Linux. It runs open-weight models locally through llama.cpp, with Metal on Mac and CUDA or Vulkan elsewhere, and picks a model to fit your hardware. The README lists Gemma four variants by memory tier, gpt-oss one hundred twenty B for workstations, and a smaller Qwen three VL for images. The agent gets tools for files, PDFs, images and Python. Python runs sandboxed to the chat's folder, and the app asks approval before changing files, offers undo, and sends no telemetry. The limit is model quality. The maker says the hardest part was getting small models to use tools reliably. It also says machines with sixteen gigabytes of RAM or more did well, by its own evaluations, not an independent benchmark. Downloads and updates still need the network, and one launch-thread user found eighteen gigabytes of models left behind. Cleanup is promised. To try it, you need Node twenty-four and pnpm ten. Clone the repo, then run pnpm dev from the project root.

4:31

Small local models are one end of the range. At the other, OpenAI says it just broke a famous conjecture. The preprint is titled The Exponential Time Hypothesis is False. Bold thing to put on a PDF. ETH is a conjecture about how hard Boolean satisfiability is. Hundreds of papers lean on it for conditional hardness results, and if it falls, those lose their footing. OpenAI says the result was found in natural-language form on October sixteenth and formalized in Lean the next day, with both released. The Verge describes a deterministic algorithm that exactly counts satisfying assignments of fixed-width CNF formulas, in subexponential time and polynomial space. For three-SAT, that would refute ETH. It's theory, not a faster solver, so nothing in your build speeds up. And Lean checks a proof against the formal statements as written, not whether they match the ETH everyone means. An external group is examining it, led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego and Ryan Williams of MIT. OpenAI says the first two were among ETH's original proposers. Until they report, this is OpenAI's claim, not a settled result. The announcement had three results. For the physics one, OpenAI asks everyone to send all communications to a Gmail address named after the problem. Enormous claim, free email. The Verge notes an AGMAI panel advised against giving chosen researchers early access, which cuts against OpenAI's approach. The thing to watch is whether the formal statement matches the standard ETH. That's what the group has to check.

6:24

Another document built to be picked apart. Anthropic's S one runs six hundred fourteen pages, ninety-four of them risk factors. Anthropic filed it Monday for a Nasdaq listing under the ticker ANT. No share count or price yet, and nothing can be sold until the statement is effective. Fortune reports first-half revenue of sixteen point three billion dollars and a net loss of eighty point seven billion. But seventy-two point four billion of that loss is a non-cash charge tied to convertible notes. First-half operating loss was six hundred ninety-three million, down from two point one billion, and operating cash flow was positive three point nine billion. The compute line matters most to builders. In the second quarter, Anthropic spent five point six billion dollars on compute. It booked eleven point six billion in revenue. The company framed that as two dollars and eight cents of revenue per compute dollar. That ratio alone doesn't prove durable unit economics. Fortune says the five largest customers made about twenty-five percent of revenue over nine months, and nearly half of revenue flowed through Amazon, Google and Microsoft marketplaces. The Register adds that the founders would retain fifty point one percent of voting power under the proposed structure. The founders' letter invokes an age of miracles, citing curing cancer and shorter lines at the DMV. Notice which one sounds like the stretch goal. These are filing figures as reported by Fortune, not audited here. Anthropic itself says it expects costs and losses to keep rising. The number to watch is that two dollars and eight cents, alongside the compute bill.

8:17

Anthropic's pitch is agents doing the work. Somebody has to review it. Thomas Dohmke, the former GitHub CEO, says the pull request has become an audit log that only the agent reads. His new company, Entire, launched Monday with a Git-compatible platform and a workflow called Trails. Here's how Trails works. You push code to a Trail, and independent review agents run in cloud sandboxes. They hand back scores, rationales and severity-ranked findings. Then you can gate merges on cleared findings, human approval and passing CI. Entire says it pushes up to eighty-nine times faster than GitHub, and its own comparison lists four hundred eighteen pushes per second against four point seven. That's Entire's benchmark, not an independent one. Its Continuous Memory tools index code structure and past agent sessions, and Entire claims up to seventy-one percent token savings in its own testing. Limits. At launch it's native hosting for private repositories only. Public repos and self-hosted nodes are on the roadmap. And GitHub already offers agentic review through Copilot. The New Stack also called out the timing, two days before GitHub Universe. Skeptics will ask whether green agent scores buy confidence or just replace understanding. It's free to start, with a one hundred dollar starter credit for review-agent tokens. Create a repo in the web interface or the CLI, or mirror one from GitHub. GitHub had its own Monday announcement, and it's about agents too.

10:01

GitHub Copilot assisted permissions are now generally available in the CLI, the app and VS Code. A language model judges each request to run a command, access a path or open a URL, and clearly safe ones skip your approval. In the CLI, run slash permissions and pick assisted. The judge is itself a language model, and the docs warn it can be wrong, so sandboxing stays separate. Transformers five point twenty-one. Python three point eleven is now the minimum, so pinned CI jobs on older Pythons will need updating to use this version. Custom tensor-parallel plans need adapting too, since input embeddings are now sharded by default. New is MiniCPM-OCR, for scans, receipts and tables, with no accuracy numbers in the notes. OpenTofu one point thirteen. New convert and assume functions let module authors tell the planner about values unknown until apply, which can help count and for each resolve during planning. A wrong promise fails at apply, like promising a vpc prefix when the value starts with subnet. And base sixty-four gzip may output different bytes, so a resource can look changed over equivalent compression. iOS twenty-seven point two. 9to5Mac counts three hundred twenty-nine unique vulnerabilities across Apple's platforms, and twelve hundred sixty-five CVE listings, since one flaw can hit several systems. MacRumors counts more than one hundred sixty for iOS and iPadOS alone, and says Apple didn't state that the iOS 27.2 flaws were actively exploited. Staying on iOS twenty-six? Version twenty-six point seven point two carries many of the same fixes.

11:53

A sixty-three billion dollar financing, with Bank of America, Blackstone, Citi and Morgan Stanley, is meant to fund compute for Anthropic built on Broadcom technology. Yahoo reports a new special-purpose vehicle will buy Broadcom accelerators and networking gear, then lease it to Anthropic, with Broadcom giving residual-value support. Deliveries are expected to begin in twenty twenty-seven. So this is financing for capacity planned for future delivery, not a report that Anthropic has already received sixty-three billion dollars of chips. That's a lot of trust being financed. Now for some trust that was misplaced.

12:38

Fast Boot, starring three files nobody should have trusted. Brotli, renamed. A malformed Brotli stream overflows an eight-byte buffer, and self-contained .NET apps and containers must be rebuilt and republished with the updated runtime, because updating the host won't reach them. The nickname HTTP2Shell is misleading too, since the flaw isn't specific to HTTP two, so the name needs a patch. Poisoned instructions. Forever Security reported finding a malicious AGENTS.md in Downloads trying to make a coding agent install an npm package; its product stopped the observed attempt before code ran. Their Codex reproduction worked over ninety percent of the time, and the fix is delightfully low-tech, namely, don't start agents in Downloads. Worm with a plan. A Miasma-family npm worm armed forty-two GitHub repositories, hiding behind payload files up to thirteen point seven megabytes that evade GitHub code-search and API paths commonly used to find this kind of compromise. It also quits silently when it spots a debugger, the only software today with an exit strategy. Patch the decoder, keep your agent out of Downloads, and note the worm got an exit strategy without filing an S one.

14:02

That's the show. Today: OpenAI's cheaper GPT six point one Terra, its claim against the Exponential Time Hypothesis, Anthropic's IPO filing, and Entire's rethink of the pull request. Your try-it for today: pick one bounded, screen-reading task and run it on Terra next to your current model. Count retries, not just tokens. Links for every story are in the show notes. Follow The Daily Reboot in your podcast app, and tomorrow's news will find you. Thanks for listening. Go patch something.