October 19, 2026 — DeepSeek's Cheap Web Search Has No Published Reliability Measurements in the Reviewed Docs

Host0:45

Ask DeepSeek's new Search API for one result, or for fifty. Same price. That's because the billing unit is a successful request. Not a token. Not a result. The price is seventy five cents per thousand successful requests. There are two APIs. Search returns structured web results. Fetch takes one URL and returns the page's title and its content as Markdown. DeepSeek's docs say both give agents and RAG applications access to live web content. You post to the search endpoint with a bearer token, the same auth as the rest of DeepSeek's API, and a JSON body with a query. An optional count asks for one to fifty results. An optional `need_content` flag asks for page text too. Both calls are read-only. A request only bills when Search returns results or Fetch returns content. Empty and failed responses cost nothing under the documented rules. Now the catch. Each API caps you at five requests per second per account, and Fetch adds a per-domain limit. Go over, and you get an HTTP four twenty-nine. Server-side calls through DeepSeek Harness, Claude Code, or Codex are free until billing starts at midnight Beijing time on November first. Activepieces did the independent math. Ten thousand successful searches cost seven dollars fifty. But the article found no published latency or reliability measurements in the docs it reviewed. Ingrid Kovarová wrote that you will have to test reliability yourself, because neither page says anything about it. DeepSeek also says Search runs on its own in-house engine. That's the company's claim, and the docs don't verify it. Hacker News commenters raised the trade-off: hosted APIs can spare you scraping blocks and operational hassle, while also raising privacy and data-handling concerns. My read: prototype now, while the free window is open. Keep it off production paths until you've measured latency, search quality, and fetch success under your own load. To try it, put your key in an environment variable. Then curl the search endpoint with the query DeepSeek V four release, count five, and need content true.

3:12

DeepSeek priced one agent call. Vercel is now showing you what a session cost, after the fact. AI Gateway logs have a new Sessions view. It groups requests into sessions and shows how session affinity routed them. Affinity is the mechanism. Vercel says it sends each turn to the provider that served the one before, so that provider can reuse more of the prompt cache. It's on by default for text and chat requests. The new part is visibility. Each session row can show the agent or application, cost, models, providers, and how much input was cached across turns. Open a session and you get cache hit rate, duration, and a turn-by-turn path. To group requests yourself, reuse a session ID in the gateway provider options, or send an x session id header. Skip both, and the gateway groups by a fingerprint of how the conversation opens. Single-turn sessions are hidden by default until you clear the minimum-turns filter. A conversation of one turn is just a request with ambitions. Now the limits. The savings story is Vercel's own explanation, and the changelog offers no independent measurement. Teams that value failover or load balancing over cache locality may not love staying on the previous provider. It's also unclear how well the view attributes cost across retries, multiple providers, or incomplete turns. To try it, send the same session ID on every turn, then watch the cache hit rate.

4:53

Linux seven point three has a speed claim too, and it also arrives without a measurement. Linus Torvalds, as quoted by LWN, wrote that it was Sunday afternoon and he'd somehow kept his wits together long enough to track the day of the week. Phoronix describes the release as on time, and Phoronix and LWN both report it landed on Sunday, October eighteenth. Phoronix highlights faster IO on modern systems, more AMD Zen 6 preparation, a driver for the Steam Controller twenty twenty-six, older AMD GPU improvements, and VRAM management work. But the report gives no benchmark. So faster IO is a description, not a number. The number we do have comes from LWN. Eighteen thousand seven hundred sixty five non-merge commits, from three thousand twenty one developers. Eight hundred seven of them were first-time kernel contributors. Under the hood, LWN lists binfmt misc extensions, group-scheduling work for large systems, fchroot, global per-CPU data for BPF programs, and sched ext work. It also lists failfs, a new filesystem whose purpose is to fail all operations. At last, a filesystem that does exactly what it says. Kernel operators will want benchmarks on their own hardware, and a clear upgrade path from their distro, before believing any of it. One oddity. The kernel.org page read for this episode still listed seven point three release candidate seven, dated October twelfth, as mainline. It showed seven point two point eleven as the latest stable. Whether the archive is lagging, or the timing was different, is unclear.

6:44

Peter Hawkins, the JAX tech lead at Google, says he hasn't waited for a Pytype action to complete since his project switched to Pyrefly. That's the story. Google says Pyrefly, an open source type checker from Meta, is now the Python type checker at Google. It replaces Pytype. Why did Pytype lose? It analyzed compiled bytecode, not source. Bytecode changes across Python releases, making it harder to keep pace with typing changes. Google had already decided Python three point twelve would be Pytype's last supported version. Pyrefly is written in Rust and uses lazy, parallel evaluation. Google says its approach is intended to follow Python's evolving syntax and typing rules more directly. The numbers are all Google's. An order of magnitude faster than Pytype in its internal benchmarks. Roughly ninety seven percent on the official typing-conformance suite. Up to ninety eight percent lower incremental rebuild latency, and more than eighty percent lower peak compute occupancy for Python type checking. Those are internal measurements, not independently reproduced. InfoQ confirms the adoption and repeats the claims, which isn't the same thing. Google says Pyrefly is stricter: a value typed int or None isn't accepted as an int argument unless your code narrows away the None. It also ships a language server, with navigation, autocomplete, hover, and inlay hints. The license is MIT, and the README calls it stable. Teams happy with Mypy or Pyright will want local numbers before moving. Anyone with Pytype-specific configuration should check compatibility. To try it, pip install pyrefly. Then run a cold check and a warm check on your own codebase.

8:44

vLLM zero point thirty two. If you're on Python three point ten, this upgrade drops support, so check compatibility before upgrading. Default CORS origins are now restricted, too. In return you get a new systemone route for structured choice and score decisions, and a parser that reads a checkpoint's response template. The notes claim forty three to eighty five percent more requests per second on short requests. That's a pull request benchmark on one small Qwen three model, so treat it as workload-specific. Low-precision KV-cache support also widens across GPUs. And schemas nested deeper than sixty four levels are now rejected. For the documented CUDA thirteen route, install with `uv pip install vllm --torch-backend=auto`. FastAPI. If you pass root path to the FastAPI constructor, the release notes say that's deprecated. The replacement is the root-path option on fastapi run. The old behavior could apply when you used the app directly, not only behind a proxy. The notes also list removal of the old openapi prefix parameter. Move it, then check your generated docs URLs. Next.js sixteen point four point two. The patch fixes Turbopack dependency edges left partially torn down by two garbage-collection paths, which sounds like a demolition report. It also backports task-snapshot consistency, Pages Router dev module-graph chaining, and dynamic-import chunk collisions to the active LTS line. There's no benchmark. Turbopack users should run `npm install next@16.4.2`. Webpack users can mostly watch.

10:28

Everything in that log runs on memory, and here's a report about who builds the next generation of it. CNA, citing Seoul Economic Daily, reports that Samsung will use TSMC's three nanometre process for base dies in Nvidia's custom HBM. The report says that would be a first for Samsung at the HBM manufacturing stage with TSMC. Nvidia reportedly asked Samsung, SK Hynix, and Micron to use that process for NVHBM base dies. The memory dies above them would use sixth-generation DRAM. The base die would carry Nvidia's custom memory controller, to standardize implementation across suppliers. Samsung would keep in-house production for standard HBM and other customers. It may use its own foundry if TSMC capacity falls short. CNA attributes the report to Seoul Economic Daily and unnamed sources; there's no company confirmation. Nothing here is callable today. But if it holds, one vendor's controller design shows up in every supplier's stack.

11:41

Fast Boot, where today's theme is things nobody was supposed to open. SiteGuard flaw. Attackers are exploiting a command injection in SiteGuard's statistics feature, enabled by default, to get root or SYSTEM if the feature is active. Cyber Kendra puts the CVSS four score at eight point five. Finally, someone found a reason to open the statistics page. Namecheap dispute. A developer says Namecheap disabled all their domains after an email giving seventy two hours' notice, and Namecheap's side isn't in the coverage yet. The customer also said they couldn't find an appeal process in the dashboard, which is the one menu you'd build first. Free Pascal packages. The GetLazarus project says its fork has dynamic packages that add IDE components without a rebuild. More than twenty years in the making, according to the project author, and a package still can't share its unit's name. Statistics page, appeal button, merge request. Somebody should open all three.

12:48

Today: DeepSeek's cheap search, Vercel's session logs, Linux seven point three, and Google's switch from Pytype to Pyrefly. Try this today. Run pip install pyrefly, and time a cold and a warm check on your own code. Show notes with every source are on the website, next to this episode. Follow The Daily Reboot in your podcast app, so the next one shows up on its own. Thanks for listening. Go build something, and read the changelog first.