October 1, 2026 — A Frontier Coding Model You Can't Call Yet

Host0:39

One million. That's the output-token limit Google claims for Gemini four Argon, announced Wednesday. The previous limit it describes was sixty-four thousand. Now the catch. You probably can't call it yet. Google is rolling Argon out first to trusted cyber defenders through its Fairwind Program, not to everyone. Broader release is planned for early paid API customers and Google AI Ultra subscribers, and I couldn't find a date. Google's announcement says safely releasing frontier capabilities at this level requires a phased approach. The safeguards it lists are refusal behavior, monitoring model activations, and monitoring chain-of-thought and actions so execution can be stopped. Trusted defenders and Google's internal teams will have a version without the cyber guardrails. How does it manage a million output tokens? Google doesn't say. The post calls Argon built for long-horizon workflows, able to sustain deep reasoning while generating very large outputs. There's no architecture, no API model ID, and no input context window. Pricing is public. The introductory rate is two dollars per million input tokens and ten dollars per million output tokens. Cached input is ninety-five percent off. After the introductory period, it becomes four dollars and twenty dollars. Google gives no predecessor price to compare against. Now the scores, all reported by Google. Seventy-seven point nine percent on DeepSWE version one point one. Fifty-one point three percent on Zapier's AutomationBench, which Google says is first place. Ninety-one point seven percent on LVBench. And a tie for first at sixty-eight percent on CWE-bench version one, for vulnerability remediation. I found no independent verification. Google also gives no numbers for its internal comparison against Gemini three point eight Flash Cyber. The fun detail is a video decoder. Google says Argon agents took an existing Rust port of its libgav1 decoder and replaced thirty-two thousand lines of SIMD code with safe Rust the compiler can vectorize. The result, Google says, runs two point seven times faster than that port, with identical output. Thirty-two thousand lines. Somewhere, a code reviewer just asked for smaller commits. Google adds that large rewrites like this, including internal work on the Fuchsia Zircon kernel, are still going through automated and manual auditing, emulation testing, and review before production. Expect pushback from two sides. Developers outside Fairwind may ask what an announcement they can't use is for them. Security reviewers will want independent validation of the capability claims and the safeguards. My read, and it's a read of the rollout, not a benchmark: eligible Fairwind defenders should try it now. Everyone else should wait for general API availability and independent head-to-head results.

3:58

Argon's gated. This one you can install right now. Magnitude is an open-source local inference engine, and its bet is that precompiled kernels leave speed on the table. Other engines, it says, ship kernels built for broad hardware categories. Magnitude tunes kernel parameters on your actual device before it runs a model. The founders wrote it in Rust, with a custom GPU-kernel runtime and an autotuner. They also describe shared prefix caches across concurrent sessions, because agents hold long conversations, not batched datacenter jobs. The README claims up to two times faster than llama.cpp. That includes ninety-two percent faster decode on Metal, nineteen percent on CUDA, and twenty-seven percent less memory per agent. The founders' specific test is Qwen three point six, thirty-five B A three B, four-bit, sixty-four K context, no speculative decoding. On an M four Pro with forty-eight gigabytes, they report decode at fifty-seven tokens a second versus thirty. On a DGX Spark, they report fifty-eight versus forty-nine. Now the honest part. Those are Magnitude's own numbers, for one model and one setup. On Hacker News, commenters asked why the Apple Silicon baseline was llama.cpp and not MLX, and one shared a case where another engine won. The founder replied that displayed speed estimates can differ from actual runs. To try it, install the desktop app, which includes the CLI. Pick a recommended model in Discover, download it, then connect an agent under Connections. It runs on macOS, Linux, and Windows, and works with Pi, OpenCode, Hermes, Codex, Claude Code, and Cline.

5:54

Magnitude tunes itself to your machine. This next project bends itself to other people's compilers, bugs included. The README for EDG's C and C plus plus front end says it emphasizes parsing compatibility and bug emulation. The goal is to accept source that compiles with Clang, GCC, or MSVC. So code leaning on another compiler's quirk can get through too. A front end parses C and C plus plus source. EDG's went public Wednesday, with The C plus plus Alliance as its nonprofit home. Phoronix reports it's under Apache two point zero. It previously sat inside Intel C plus plus Compiler classic, NVIDIA's CUDA NVCC, and Visual Studio IntelliSense. The interface is simple. You hand it C or C plus plus source, it parses it, and the repo's back ends can hand you C or C plus plus back out. Also in the box: a prelinker for automatic template instantiation, a minimal runtime support library, and intermediate-language utilities. The governance is the interesting part. Contributions get reviewed by experienced EDG developers, and larger features can be community funded, under a Fiscal Sponsorship Committee chaired by John Spicer. It all goes into the same public codebase, with no early access for funders. EDG's transition site calls it a change of steward, not a change of course. Pushback may come from compiler vendors and users of existing licensed EDG products. They may ask about support obligations and integrations. The site says continuity is intended, but it doesn't detail every commercial arrangement. To try it, do a partial clone. Run git clone with the filter option set to blob none, pointed at the edgcpp compiler repository. The README links to separate build instructions.

7:59

Seven days of live transactions, joined against five years of historical Parquet, in one query. That's AWS's own walkthrough for a new Aurora PostgreSQL feature. EDG keeps one shared codebase for everyone. Aurora's new trick is the data version of that: it queries Iceberg and Parquet data in your data lake directly, without an ETL copy step. The engine is DuckDB, embedded in Aurora to process analytical scans. AWS says queries can combine lake records with live operational data, uncommitted writes included. Until now, per AWS and SiliconANGLE, that usually meant reverse-ETL pipelines, duplicated data, and synchronization work. Setup has four steps. Attach an IAM role with the AuroraAnalytics feature. Enable the aurora analytics extension. Create foreign tables pointing at your Iceberg or Parquet data. Then query in ordinary PostgreSQL syntax. In AWS's example, the foreign table has an empty column list, because Aurora infers the Parquet schema from file metadata. Declare zero columns, get all of them. SiliconANGLE reports the role covers S three and Glue. It also says catalogs compatible with the Iceberg REST Catalog work through Glue federation. You need Aurora PostgreSQL seventeen point eleven or later, or eighteen point six or later. It's in all commercial regions and GovCloud US. There's no added feature charge, but you pay incremental Aurora compute and S three request costs. The limit comes from AWS itself. For single-digit-millisecond latency, materialize the hot lake data into native Aurora tables with SQL. Reads can run on a writer or replica, but materialization runs on the writer. And the complexity savings are AWS's claim, with no independent benchmark. What nobody's shown yet is how it behaves at real customer scale. Latency, concurrency, and cost, against a purpose-built analytics engine or a materialized copy, are still unknown.

10:21

Whatever you decide about Aurora, there's something to patch today. Next.js sixteen point three point eight fixes a high-severity server-side request forgery in Image Optimization. It also covers App Router metadata-image disclosure, multiple cache-poisoning issues, Draft Mode leakage through pending use-cache fills, a nested use-cache root-parameter leak, and a disclosure in the dev server's MCP endpoint, because of course there's an MCP endpoint. The fifteen line gets a subset in fifteen point five point twenty-seven, so use the fixed version for your line. The notes give no CVE and no exploit conditions, so upgrade first and read later. Vercel Agent can now install private npm packages and custom-registry dependencies. Store credentials as team-shared variables, not project-scoped ones, on Development or Preview. Use NPM_TOKEN for private packages on the npm registry, and NPM_RC for custom or multiple registries. Vercel says the values stay outside the sandbox, so the agent can't read them. That's Vercel's own description, and security teams will want rotation details. npm trusted publishing can now manage dist-tags with short-lived OIDC credentials. So moving latest, next, or beta no longer needs a long-lived token. Enable Allow npm dist-tag in the settings. It's off by default, even for existing configurations, and token-based management keeps working. If you stripped tokens everywhere except tags and rollbacks, flip it. S three Vectors now evaluates metadata filters before similarity search. AWS claims up to five times more matching vectors on highly selective filters versus CLASSIC mode, and that's AWS's number alone. Its example is one customer's four hundred tickets out of eight million. Existing indexes stay CLASSIC until you run the s3vectors update-index-mode command in the AWS CLI, after your IAM policy grants the new actions. No re-ingestion, no query changes, no extra charge.

12:41

Fast Boot, connecting now. Please, nobody pick up the extension. Zimbra burglars. Shadowserver found two hundred seventy-four compromised Zimbra instances the week before Ars's report. The fix, ten point one point twenty, shipped on July twentieth. Attackers briefly made a public directory writable for a web shell, then restored permissions, like burglars who relock the door. RedFlick shortcut. Microsoft says Star Blizzard ran at least thirteen phishing campaigns since January, affecting over one hundred organizations, with a technique needing one user interaction: opening a shortcut. The shortcut wears a PDF costume while a hidden command works, which is a lot of tailoring for malware. Dial-up returns. The site fifty-six k dot rip rebuilds nineteen ninety-six in your browser, with a rate-limited byte budget at the speed you pick. It even simulates someone picking up the extension and dropping your call, a failure mode no status page admits to. So the lesson is extensions. The one on your phone line can drop your call, and a shortcut disguised as a PDF can start an infection chain.

13:56

Today we covered Google's Gemini four Argon and its gated million-token release, Magnitude's self-tuning local inference engine, EDG's C and C plus plus front end going open source, and Aurora PostgreSQL querying Iceberg and Parquet directly. Your try-it for today: do a partial clone of the EDG front end, or download Magnitude and race it against your own model on your own hardware. Links to everything are in the show notes. Follow The Daily Reboot wherever you get your podcasts, so tomorrow's news finds you on its own. I'm your host. Go patch Next.js. See you tomorrow.