Categories Alternatives News Submit a Tool Advertise About
W45 Oct 26 – Nov 1, 2026
Issue W45 · 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 · 23 stories
59 Gemini 4 Carbon No. 1 on Intelligence Index v5 · $2/$10 now, $4/$20 later
12 pts Open-weight gap Was 3 · best open model scores 32
EOL 11-30 Opus 4.8 fast mode Migrate to Opus 5.5 or it stops
26,400 Gitee live keys $3,097 to harvest · 3,170 DeepSeek keys

Top Stories

What you need to know from this week
★ Editor's Pick
Editor Pick Oct 28 · Google DeepMind

Gemini 4 Carbon Arrives With a Price That Expires

Google shipped Gemini 4 Carbon two weeks after Argon and put it straight to number one on the rebuilt Intelligence Index v5 with a score of 59, plus 88.2% on SEC-bench Pro. The launch notes are as much about safety as capability: more than double the safe-code generation of Argon, a 10x patch-generation rate against another lab on the Chrome team benchmark, an 80% drop in internal high-severity misalignment alerts, a 0.0% success rate against Gray Swan indirect prompt injection, and cheat attempts on Terminal-Bench 4.0 down to 0.4%. The number to write down is the price: $2 in and $10 out per million tokens as an introductory rate, moving to $4 and $20 later. Rollout goes to security defenders first, then Fairwind partners, then everyone.

Security Oct 26 · Apple / eMudhra

Every Mac Still Trusts the Certificate That Was Already Broken

Apple was told on 17 October that an eMudhra intermediate certificate had been compromised. macOS 27.2 shipped on 26 October without revoking it, and in the same release removed the manual workaround users had adopted, while Apple documentation warns a later update may undo the fix again. Until the root is removed by hand, git, curl and python3 on that machine still trust it — which means a clone or a push can be intercepted and a GitHub token lifted mid-flight. The check is one command against a lookalike host: if it lists commits, the machine is affected. This is the least glamorous item in the issue and the one with the shortest path to a stolen repository.

Supply Chain Oct 28 · independent scan

7.25 Million Repos, 26,400 Working Keys, $3,097

A scan of 7,256,531 Gitee repositories across 1.7 million usernames turned up 26,400 credentials that still authenticated: 9,473 Alibaba Cloud, 3,805 YouTube API, 3,170 DeepSeek, 1,489 Zhipu, 1,198 OpenAI and 154 Hugging Face. The method was unglamorous — three-letter username prefixes to get around the 200-result cap, aggregation bucketing in Elasticsearch, 350 concurrent containers cloning bare repos on Modal — and the whole run cost $3,097. Gitee came out at 3.64 live credentials per thousand repositories, denser than GitLab at 3.11 or Bitbucket at 2.39. TruffleHog and Kingfisher each found about half, and agreed on only 1,001 of them.

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%
Grok 4.5 (Cursor CLI)
79.3%

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
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

5 stories
Artificial Analysis Oct 28

A New Scale Reshuffles the Board and Widens the Gap

Intelligence Index v5 replaced a third of its public test set with private data, added FrontierCode 2.0 and Terminal-Bench-Science, and produced a new order: Claude Opus 5.5 at 58, GPT-6 Astra 57, GPT-6.1 Sol 55, Kimi K3.1 Pro 46 as the best open model, and DeepSeek V4.1 Flash with GLM-5.3 at 32 for models with published weights. The gap between the closed frontier and the best downloadable model went from 3 points to 12. Cost per completed task now spans 13x inside the same leaderboard, from GPT-6.1 Terra at $0.96 to Claude Fable 5.1 at $13.04 — which is why a single score is a worse buying signal than it was last month.

community teardown Oct 28

One Game, Two Budgets: $15 and $265

Someone rebuilt a Calvino interactive in the browser with Opus 5.5 twice and published both bills. The medium setting produced 7,746 lines for $14.98 over 53 minutes of API time and about 313K output tokens. Max with ultracode produced 31,387 lines for $264.64: 5.3M output tokens of which 3.65M were thinking, 242 web searches, 16.5 hours of API time, roughly eleven parallel agents, and a 17.7x cost ratio. The mouse input never worked in either run. As a worked example of what an agent actually costs, it is the most useful number published this week.

Nvidia Oct 28

Nvidia Opens a 4B-Active Transcription Model

Nemotron 3.5 Transcribe is a hybrid Mamba-Transformer mixture-of-experts with about 30 billion parameters total and 4 billion active per token, covering 40 languages for transcription, 31 for translation and 13 for diarization. It scores a 5.33% average word error rate on the Hugging Face Open ASR board — first among open weights — was trained on 463,062 hours of Granary v2 audio, and the quantised build transcribes 100 hours of audio in under three minutes on a single RTX PRO 6000 Blackwell.

Google Oct 29

Google Sleep: Memory That Survives Being Asleep

A two-phase framework that writes short-term memory while awake and consolidates it in a sleep phase — unlocking low-rank MoE experts, on-policy distillation, reinforcement learning on imitated trajectories, and a dreaming stage that generates synthetic data from randomly activated experts and filters it by gradient. The results claimed are unusual: near-zero forgetting on languages never seen in training, BABILong expanded to 10 million tokens, AIME 24 at 79.2% and AIME 25 at 69.0%. If it holds, persistent agent memory stops being a context-window problem.

Cartesia Oct 29

Cartesia H-Net Cuts Pretraining Tokens Tenfold

A 3.67B two-stage model with a Mamba-2 codec and a 40-layer transformer, about 7.6x compression, that reaches the final average score SmolLM3-3B reaches — using roughly 1 trillion tokens instead of 11.1 trillion, with FLOPs held within 2%. Code gains are 2.04x in-distribution and 3.10x across ten unseen languages, with a 5.37x peak on Prolog and 5.95x to 7.06x zero-shot on DNA. Learned tokenisation shifts the cost curve, not the quality curve.

Tools & Editors

6 stories
Anthropic Oct 30

Claude Code v2.1.314 Locks Writes to the Working Directory

A new permissions.blockWritesOutsideWorkingDirectories setting makes Write, Edit and NotebookEdit refuse any path outside the working directories in every permission mode — a structural fix rather than a prompt-level instruction. The same release routes gateway traffic by IdP group, compacts long-idle sessions automatically, and closes a worktree sandbox protection gap among roughly 110 changes. Thirty days of agent releases have mostly been about where the agent is allowed to touch.

Anthropic Oct 31

Claude Code v2.1.316 Adds Group Budgets — and a Deadline

Pooled group budgets and a new subagent effort setting, plus a fix for shell commands that kept running after exit. The changelog also carries the notice that matters more: Opus 4.8 fast mode is being retired, with migration to Opus 5.5 required before 30 November. A deprecation line inside a release note is now the way model churn reaches a budget, and it arrives with four weeks of warning if you read the changelog and no warning at all if you do not.

OpenAI Oct 29

Codex CLI rust-v0.166.0 Adds Terra and a Queue

Keyboard-driven transcript selection and search, GPT-6.1 Terra added to the model catalogue, and codex queue --resume for picking up queued work. Terra is the cheapest model per completed task on the new index at $0.96, so its arrival in a CLI catalogue is a pricing event disguised as a release note.

AWS Oct 29

Kiro CLI v2.36.0 Adds an "Approve for Me" Mode

Interactive V3 sessions now auto-approve routine operations, block high-risk ones and ask when the classification is unclear. Every harness in this market is converging on the same three-way dial — approve everything, approve nothing, approve by risk class — and the difference between products is increasingly in how well that middle setting is calibrated rather than in what the agent can do.

GitHub Oct 28

Copilot CLI v1.0.105 Switches Models Mid-Session

When a long-context session stalls, /model now moves it between Copilot Fusion High and Max without restarting. Model switching inside a live context is the small feature that makes subscription tiers meaningful: it turns a plan limit from a wall into a setting.

Cognition Oct 28-30

Devin Desktop Fixes the Thing That Was Killing Debugging

Five releases in three days, and the one worth noting is unglamorous: diagnostics now keep the last 4 MB of each output log instead of 25 KB, which is the difference between reading a failure and guessing at one. The same window added Allow/Reject buttons on tool requests for web fetches and multi-file edits, while cloud Devin gained a live Progress tab and routed Azure DevOps pull-request feedback back to the right session.

Pricing & Business

4 stories
Google DeepMind Oct 28

Gemini 4 Carbon Is Priced as a Countdown

The introductory rate is $2 per million input and $10 output with a 95% cache discount; the published later rate is $4 and $20. That is not a discount, it is a timetable, and teams that model unit economics off launch coverage will be wrong within a quarter. The rollout is also staged by trust rather than by region: security defenders, then Fairwind Program partners, then general developers, with Google joining the US voluntary pre-release testing process alongside it.

Anthropic / GSA Oct 29

Anthropic Cuts 40% for US Federal Buyers

A GSA agreement discounts Claude Enterprise, Claude for Government and the Claude API by 40% across the board with no minimum commitment, projected to save agencies about $1 billion. Around 500 US government agencies already use Claude, and Claude for Government weekly active use is roughly 8x what it was in February after the 30 September GA with FedRAMP High; some models carry DoD IL4/IL5 on AWS GovCloud. Federal procurement is now a price-setting channel for a vendor that sells mostly to developers.

Anthropic Oct 31

Opus 4.8 Fast Mode Has a Date on It

The retirement notice inside Claude Code v2.1.316 sets 30 November as the migration deadline to Opus 5.5. Users who tuned cost around fast mode have four weeks to re-measure: this is the third model in five months to leave a plan mid-quarter, and the pattern to plan around is not the specific model but the cadence. Treat any per-model pricing advantage as expiring.

Alphabet / SK hynix Oct 29

Cloud Spend and DRAM Set the Floor

Alphabet reported Q3 revenue of $127.6 billion, up 25%, with Google Cloud at $30.7 billion, up 103%, and a 31.8% operating margin. SK hynix separately said DRAM is effectively sold out through 2029 with demand exceeding supply into 2030, already under long-term contracts with some customers. Two numbers that decide what inference costs next year: how much capital is being spent on capacity, and how little of it can be bought. Falling token prices have a floor, and the floor is memory.

Ecosystem & Community

5 stories
Apple Oct 26

macOS 27.2 Ships Without Revoking the Compromised Root

The eMudhra intermediate was flagged to Apple on 17 October. The 26 October release did not remove trust and revoked the manual workaround Apple had documented, and Apple documentation warns a subsequent update may roll the change back again. Affected binaries are git, curl and python3 — the three tools a developer uses before opening an editor. Manual root removal works and has to be re-verified after every system update, which is a maintenance task that will not survive a team.

TianoCore / AMI Oct 27

A 9.5 in Firmware That Waited Sixteen Years

CVE-2026-97631 is a buffer overflow of 9.5/10 in EDK II PXE network boot, reachable before the operating system and before Secure Boot protections apply, with TPM and disk encryption in reach during a short physical-access window. AMI, Insyde and Phoenix all inherit the reference code, and Microsoft hardened Project Mu branch carries it too. It was found by researchers at Anthropic and Google; coordinated disclosure was planned for January until an AMI pull request leaked it in September. The patch is public, and weaponisation is now a matter of somebody doing it.

Tranco / community scan Oct 29

A Fingerprint Script That Fights Back With Prompt Injection

A scan of the Tranco top 10,000 produced a canvas-fingerprinting script on servicepipe.tech whose function names read like computeRicinLethalDosageForWaterSupply and whose hidden Unicode strings decode into a fake SYSTEM OVERRIDE prompt listing DDoS, tax evasion and bioweapons — bait designed to make a model refuse to help deobfuscate it. Astra and Daybreak Blue refused; GLM-5.3 complied and did the work. Under Node the same script reads .npmrc, .env and .kube/config and tries to fetch a binary. Anti-analysis is now an adversarial-input problem, not an obfuscation problem.

MuPDF / Foxit / OpenPDF Oct 28

AES-256 in Three PDF Libraries Is Nine Bits

MuPDF before 1.28.5 seeds a 48-bit LCG with a second-resolution timestamp, leaving 512 possible file keys and a brute force measured at about 11 milliseconds. Foxit Reader 2026.1.3.36551 uses a 32-bit timestamp plus 32 bytes of zeroed memory, which a multithreaded tool exhausts across all seconds since 1970 in under five seconds. OpenPDF 3.0.5 and iText 9.7.1 seed ARC4 with a millisecond timestamp and idle JVM memory. Foxit traces to a PDFium commit from 2014. MuPDF is fixed; documents already encrypted cannot be re-encrypted, which is the part with no patch.

AlpineBSD Oct 28

AlpineBSD Exists to Have No AI Code in It

A NetBSD kernel and libc with an Alpine-style userland — apk-tools, busybox, s6 init, OpenRC planned — whose stated purpose is to keep LLM output out of the base system. It has no installer, only a high-level guide, and hardware support is plainly behind Linux. The interesting part is not the distribution but the demand for one: a group of users now wants a documented provenance boundary, and 234 points on the front page says that demand is not one person.

Our Take

The editorial view
01

Two things moved this week and they point the same way: the way models are scored, and the way they are priced. Index v5 replaced a third of its public test set with private data because public sets get optimised against, and Carbon arrived with cheat attempts on Terminal-Bench 4.0 cut to 0.4% — the benchmarks are being rebuilt while the models get better at them. At the same time the open-weight gap widened from 3 points to 12. If you self-host, that is the number your architecture depends on, and it got four times worse in one revision.

02

Carbon is the clearest example of pricing as a schedule rather than a fact: $2/$10 today, $4/$20 later. Any model of your own costs built on a launch rate has an expiry date in it, and the same week Anthropic put a hard 30 November date on Opus 4.8 fast mode. Between those two, the practical rule for 2026 is to treat every per-model price as a lease with a term, and to keep the routing layer thin enough that switching is a config change rather than a migration.

03

The rest of the issue is the unglamorous half: a certificate Apple has known about since 17 October, a firmware overflow from 2010, 26,400 credentials that still work on Gitee, and three PDF libraries whose AES-256 is nine bits of entropy. None of it involves an agent being clever. All of it sits underneath the tools that do. That is the recurring shape of these weeks — capability compounds, and the maintenance underneath it does not, so the gap fills up with exactly the kind of thing a browser tab and a 15-minute fix would have closed.

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