Categories Alternatives News Submit a Tool Advertise About
W46 Nov 2 – Nov 8, 2026
Issue W46 · Published every Monday

AI Coding Weekly

Signal, not noise. The models, tools, pricing changes and ecosystem shifts that actually matter for developers — curated once a week.

Read this week ↓ ~5 min read · 22 stories
65.7 Kimi K3.1 Pro Terminal-Bench 4.0 · open weights · 104B active of 2.8T
$32M Ripple-65B US open-weight SOTA at 1/10 the cost · TB2.1 80.5
$6.3B Binance loss Fake interview → poisoned repo → EDR bypass
OFF 11-20 Conductor cloud Read-only 11-13 · data deletion begins at shutdown

Top Stories

What you need to know from this week
★ Editor's Pick
Editor Pick Nov 3 · Moonshot AI

An Open Model Just Tied the Closed Frontier

Kimi K3.1 Pro shipped with its weights open: 2.8 trillion parameters in a mixture-of-experts with 104 billion active per token, 896 experts, 93 layers, a 1,048,576-token context, a MoonViT-V2 vision encoder and MXFP4 quantisation, under a custom licence. It scores 65.7 on Terminal-Bench 4.0 — ahead of Claude Opus 5.5 at 64.8 on the same board — plus 51.7 on SWE Marathon v1.1, 94.2 on GPQA-Diamond and 92.1 on BrowseComp. Last week this newsletter reported the open-weight gap widening from 3 points to 12 on Intelligence Index v5. One week later a downloadable model is at the top of a different frontier board. Both numbers are true; the conclusion available from either of them alone is not.

Benchmark Snapshot

Frontend Code Arena · human-preference voting

TerminalBench 2.1 — Terminal Agent Score

GPT-6 Astra (Codex)
87.4%
Fable 5 (Claude Code)
83.8%
GPT-5.5 (Codex)
83.2%
Ripple-65B (open weights)
80.5%

Research & community-reported scores · TerminalBench 2.1

API Price — per 1M tokens (combined in + out)

Claude Haiku 5.5 $0.10/$0.50 under 100K
$0.60
GPT-6 Luna $0.10/$0.50
$0.60
DeepSeek V4-Flash off-peak
$0.88
Hunyuan Hy4 $0.90/$2.70 · 50% off to year end
$3.60
Mistral Large 4 $1.36/$4.18 · open weights
$5.54
Gemini 4 Carbon $2/$10 intro → $4/$20
$12.00
Claude Sonnet 5.5 $2/$10 · cache 0.05x
$12.00
GPT-6.1 Sol $2/$10
$12.00
$0 $10

Research & community-reported prices

Models & Benchmarks

4 stories
Ricursive Intelligence Nov 4

Ripple-65B Trains to the Front for $32 Million

A dense 65B model, 80 layers, 12 trillion training tokens, context up to 1M, 40% MFU, with no custom kernels and no quantisation. It scores 80.5 on Terminal-Bench 2.1 — the best published number from an American open-weight model, ahead of Reflection Beam 501B at 80.1 — and the entire research effort cost about $32 million, roughly a tenth of the previous state of the art. Weights are on Hugging Face, Together AI and OpenRouter. A 65B dense model beating a 501B MoE says more about training efficiency than about scale.

Tencent Nov 4

Tencent Prices Hunyuan Hy4 at $0.90 In

The general release of Hy4 expands pretraining and post-training on August preview, which was 770B total with 49B active and a 1M context. It adds fine-grained image understanding, chart reading, CLI and GUI computer operation, and targeted training for Blender and CAD. Input is about $0.90 per million tokens and output $2.70, with 50% off on Tencent Cloud TokenHub until the end of the year, and a Flash variant coming. Chinese labs are competing on the price of computer-use capability rather than on benchmark rows.

Anthropic Nov 4

Anthropic Open-Sources Proof That an Inference Ran

FlashProve, Apache 2.0, generates a zkSNARK proving a model produced a given output. Llama 4 Maverick at 400B MoE in 4-bit runs at about 110 tokens per second on a single H100 and emits roughly 19MB of proof per 1,024-token request; verification takes about 17 seconds on 24 CPU cores. The overhead is 70 to 95 times normal inference, against 14.3 hours per 1,000 tokens for the earlier VerInf approach. The limits are real: integer models only, greedy decoding, context of 1,024 tokens. Verifiable inference went from a paper to a library in one step; it is still three orders of magnitude away from routine.

Open Athena Nov 5

There Is No Post-Training Data Commons

A useful accounting of what agent training data actually exists: SWE-bench started at about 90,000 pull requests, only roughly one eighth could be mapped into tasks, and executability filtering left 2,294. SWE-bench Verified was reviewed by 93 developers who flagged problems in more than two thirds of it, and the TaskTrove sets discard over 90% of a 5,000-task Java subset. Every lab is doing this filtering independently on the same scraped repositories, which is why "just train on more public code" has stopped being an answer.

Tools & Editors

6 stories
OpenAI Nov 5

OpenAI Buys Pierre and Folds It Into Codex

Terms were not disclosed and the deal awaits regulatory approval. Pierre brings Code Storage plus the Trees and Diffs libraries, which go into Codex, with a commitment that the open-source projects stay open. Storage and diff primitives are what an agent needs when several sessions touch one repository — this is OpenAI buying the substrate of parallel agent work rather than another model.

Go project Nov 4

Go 1.27 Promotes encoding/json/v2

The v2 API and jsontext graduate from the experiment that began in Go 1.25. Unmarshal is up to 10x faster, Marshal is flat, invalid UTF-8 and duplicate keys are rejected by default, nil slices serialise as [] rather than null, and v1 remains supported indefinitely. The speed is the headline; the stricter defaults are the part that will break code silently in CI rather than in review.

kitty Nov 3

kitty 0.50.0 Makes Windows Dockable

Windows can now dock to another window or tab, and the layout persists across sessions; restore after logout uses the macOS keychain, and an OSC 7501 protocol reports running state. 0.50.0 also fixes an authentication bypass in remote control where a crafted request skipped the password check. Terminal emulators rarely make this list twice in a year; both entries are about configuration as a security boundary.

ISO C++ / WG21 Oct 23, reported Nov 5

C++26 Was Rejected Over Contracts

National bodies failed to reach the two-thirds majority on 23 October, with Contracts cited as not ready, so the DIS did not pass. WG21 meets again on 16 November to resolve the deadlock. A standard that ships with a feature everyone wants and nobody can agree on the semantics of is a good reminder that the language your tools parse is also a moving target.

MindStudio Nov 4

An RTX Spark Laptop Runs Inference Five Times Slower on Linux

The CPU-GPU power table is held by a Windows-only component and encrypted, and the Linux driver has no key for it, so on mains power the GPU locks around 780 MHz against roughly 2.2 GHz on Windows. GPT-OSS-120B runs at 13.9 tokens per second natively versus 65.4 on Windows, and vLLM at eight concurrent requests manages 39.5 against 127 under WSL. Battery mode is unaffected. Until Nvidia ports the power management into the Linux driver, the workaround is to virtualise the operating system you bought the laptop to avoid.

Replit Deadline Nov 8

Replit Forces Arbitration, With a Deadline

New Pro terms effective 9 October impose mandatory arbitration and a class-action waiver; opting out requires an email to arbitration-opt-out@repl.it by 8 November. Nothing about the product changed and almost nothing about the tool page would ever surface this, which is exactly why it belongs in a changelog: for a subscription developer tool, the terms are part of the release.

Pricing & Business

4 stories
Retail / Qualcomm / Apple Nov 5

An RTX 5090 Costs $9,000 at Retail

Street price in the US passed $9,000 against a $1,999 list, a 4.5x premium, as inference demand pulls high-VRAM supply out of the consumer channel. Qualcomm meanwhile shipped a 150W, 128GB, 548GB/s inference card aimed at rack power limits, and Apple added a 512GB M5 Ultra configuration. Local inference has a price now, and it is set by datacentre demand rather than by gamers.

Tencent Cloud Nov 4

The Cheap End of the Market Got Cheaper

Hunyuan Hy4 at $0.90 in and $2.70 out, with 50% off on TokenHub until year end, lands below Gemini 4 Carbon and Sonnet 5.5 by a wide margin and includes computer use. The competitive axis for the sub-$1 tier is no longer tokens of chat: it is whether the model can drive a GUI, which is what makes an agent expensive in the first place.

Nikkei Asia Nov 5

Big Tech Has $3.2 Trillion of AI Debt Off the Books

Five US hyperscalers carry about $3.2 trillion in off-balance-sheet AI liabilities, up 90% in six months and roughly 80% of their combined assets, per Nikkei Asia. In the same week, more than 30 local ballots across nine states voted on datacentre measures, and lenders were reported to be pricing local opposition into credit. The financing structure behind cheap inference is now the least stable part of it.

Y Combinator / DoorDash Nov 5

One Team Documents $100K to $200K a Year in Agent Tokens

Garry Tan reports spending $100,000 to $200,000 annually on agent tokens: 23 Claude Code skills, an internal agent called Stray Light, and a personal agent ingesting email, Dropbox and twenty years of Apple Notes. DoorDash, in the same window, disclosed that external AI agents have produced "hundreds of orders" to date. Both numbers are honest, and the gap between them is the entire state of agent commerce.

Ecosystem & Community

5 stories
Debian Nov 4

Debian apt Could Be MITM-ed Into Root

CVE-2026-109518: when an HTTP redirect assigns the same URI to two different package downloads, apt merges the files, and a man in the middle can substitute a package after verification and get code executed as root. Debian 13 trixie is affected; the fix is apt 3.0.4. Package managers are the one program on a build machine with both network access and root, and this is the second time this year that a compression or parsing detail inside one became remote code execution.

Vino / Synaptics Nov 4

A Patent Claim Took Down an AI-Assisted Driver

Synaptics sent a cease and desist over Vino, a Rust DRM driver for modern DisplayLink hardware built by clean-room reverse engineering with some AI assistance. The GitHub repository is offline and the project is dead. Whether or not the claim holds, the pattern is worth noting: reverse engineering used to be slow enough that legal pressure arrived after the community had a stake, and AI-assisted work changes the timing.

IEEE Spectrum Nov 5

AI Is Being Used to Pay Off Forty Years of Tech Debt

IEEE Spectrum argued this week that AI is forcing teams to settle old debt, using a PHP memory bug that AI tooling escalated into a working RCE as the example; it affected Wikipedia and NASA before being patched. The piece drew 946 points and 576 comments, and the debate was less about the finding than about who pays for the fix. This is the same shape as the PHP disclosure in our W44 issue, with a longer horizon.

community report Nov 4

Gemini Training on Personal Chats, With a Switch

A widely shared thread claimed Google uses personal-account chat history, including paid subscriptions, for training unless the user turns it off manually, while enterprise and education customers were always exempt. The volume of that thread — 563 points and 148 comments — is the signal: most people do not know which of their tools train on their work, and the companies are not in a hurry to make it a checkbox on the pricing page.

AMD Nov 5

AMD Is About to Outship Intel

AMD guided to roughly $17 billion for the fourth quarter and is expected to pass Intel in unit sales for the first time in a single quarter. The laptop you are reading this on is increasingly likely to be AMD, which matters for the accelerator story above: local inference support follows whoever ships the silicon, not whoever writes the best model card.

Our Take

The editorial view
01

Last week this newsletter wrote that the open-weight gap had widened from 3 points to 12. This week an open model is at the top of a different frontier board, and a 65B dense model trained for $32 million beat a 501B mixture-of-experts. Both facts are correct, and the second one should teach the more useful lesson: one leaderboard week is not a trend. We said the same thing about our own traffic data two issues ago. The rule generalises — a single measurement tells you where something is, not where it is going, and the cost of a wrong thesis is highest exactly when it is most satisfying to hold.

02

Conductor gives users nine days to stop relying on it and sixteen before their data starts being deleted, and it is doing this politely, with refunds and unlocked Pro features. Replit gave users a month to opt out of forced arbitration. Neither is a scandal, and both are the same underlying fact: a developer tool is a set of terms that can change while the software stays identical. The only durable hedge is portability — work in formats you can carry, keep the export path tested, and treat "can I leave in a week" as a feature on par with the model it runs.

03

The Binance intrusion is the item to read twice. It began with a recruiter on a real platform, ran through several interview rounds, and ended with a developer executing a take-home exercise — after which a replaced pyenv component bypassed EDR and a container token from the wallet team did the rest. No model was tricked. No prompt was injected. The most expensive attack of the year so far used the hiring pipeline and a package manager, and in the same week Debian shipped a root-level apt flaw. Your laptop and your dependency tree are the perimeter, and neither of them appears in an agent architecture diagram.

Older issues are loaded on demand to keep this page fast.