Categories Alternatives News Submit a Tool Advertise About
W47 Nov 9 – Nov 15, 2026
Issue W47 · 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
$8/$40 Fable 5.2 API per 1M tokens · was $10/$50 · cache reads $0.20
49.1% FrontierCode 2.0 Fable 5.2 · Opus 5.5 45.5 · Astra 42.8 · Carbon 41.1
56.5% AWS Emulator fidelity units passing · 29.7K commits · 0 human lines
ENDED 11-12 Cursor contract OpenAI terminated · SpaceX paid $60B in August

Top Stories

What you need to know from this week
★ Editor's Pick
Editor Pick Nov 9 · Anthropic

Anthropic Cut Its Frontier Price 20% and Made the Cut the Default

Claude Fable 5.2 and Claude Mythos 5.2 are one underlying model in two configurations: Fable ships with cybersecurity and biology safeguards to everyone, Mythos goes to vetted security and life-science organisations without them. On 9 November the public one arrived at $8 per million input tokens and $40 per million output, down from Fable 5.1 at $10 and $50, with cache reads at $0.20 and an optional fast mode at $16/$80 for up to 2.5x generation speed. Anthropic reports 49.1% on FrontierCode 2.0, ahead of Opus 5.5 at 45.5, GPT-6 Astra at 42.8 and Gemini 4 Carbon at 41.1, plus 74.2% on Humanity's Last Exam and an AA-3D Elo of 1,271 whose confidence interval does not overlap the next model. Claude Code 2.1.321 made the same build the default Fable model and migrated saved 5.1 picks without a prompt. The price cut is the story; the silent default is what will change your bill.

Infrastructure Nov 11 · AWS

AWS Shipped a Cloud Emulator No Human Wrote or Reviewed

AWS Emulator, released on 11 November, serves more than 400 API models across more than 200 services locally, so an agent can build and test against AWS without an account, a network call or a bill. It is CC0, and the code came from autonomous agents running on Bedrock AgentCore with CodeCommit and CodePipeline: one group writes fidelity tests against real AWS, one replays them against the emulator and files gaps, one fixes and merges. Matt Goldberg, the senior principal engineer behind it, says the team wrote and reviewed none of the code. The repository was created on 11 September and shipped with roughly 29,700 commits and 322 releases in 61 days. The release's own fidelity file reports 56.5% of 42,937 test units passing, which is the number to read before running it near anything real.

Business Contract date Nov 12 · SpaceX / Cursor / OpenAI

OpenAI Ended Its Cursor Contract on 12 November

SpaceX closed its $60 billion acquisition of Cursor on 14 August and folded the company into SpaceXAI alongside Grok Build, Grok Bot and Grok API. In late August OpenAI told Cursor it would end their agreement on 12 November, saying that after previous contract disputes with Musk-owned companies it could not guarantee SpaceX would stay inside the terms. That date has now passed. Whatever the model menu looks like on the other side, the structure is the lesson: the models in your editor are a commercial relationship between two other companies, and it can be terminated on a date you were told about four months in advance and probably did not notice.

Benchmark Snapshot

Frontend Code Arena · human-preference voting

FrontierCode 2.0 — Agentic Coding Score

Claude Fable 5.2
49.1%
Claude Opus 5.5
45.5%
GPT-6 Astra (Codex)
42.8%
Gemini 4 Carbon
41.1%

Vendor-reported scores · Anthropic release table, 9 Nov 2026

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 Hydrogen $1.50/$7.50 · flat to 1M
$9.00
Gemini 4 Carbon $2/$10 intro → $4/$20
$12.00
Claude Sonnet 5.5 $2/$10 · cache 0.05x
$12.00
$0 $10

Research & community-reported prices

Models & Benchmarks

4 stories
Meta Nov 10

Meta's Smallest Model Matches Haiku on Agent Work at a Third of the Cost

Muse Ember is Meta's smallest and most efficient model, positioned for repeatable agent work rather than chat. On Artificial Analysis it lands beside Claude Haiku 5.5 while costing about a third per task, and Meta's own table puts AutomationBench-AA at 37.8 against 13.9 for Haiku 5.5 and 20.2 for GPT-6 Luna. The figure that matters is not the score but the denominator: once models are priced per completed task, a cheap model that finishes beats a strong model that needs three attempts. Pi Dash reported running ten Muse Code agents under Ember to rewrite about 178,000 lines of production code in a day, migrating raw SQL to SeaORM.

Thinking Machines Lab Nov 12

Thinking Machines Opens Weights It Withheld a Week Earlier

Inkling 2 Small shipped closed on 5 November with a promise to open after external safety testing; on 12 November the weights went to Hugging Face under Apache 2.0 in NVFP4 and MXFP8, with recipes for Unsloth, SGLang, vLLM and TokenSpeed. It is a 594B total, 17B active mixture-of-experts with native image input and a 1M context. Self-hosting starts around 500GB of aggregate GPU memory in NVFP4, so for most teams the honest answer is still a hosted endpoint. The part worth noticing is the cadence: a one-week gap between closed launch and open weights is now routine rather than a concession.

Meituan LongCat Nov 9

Meituan Releases a 1.6 Trillion Parameter Model Under MIT

LongCat-2.5 is a natively multimodal model built for long-horizon agent work: 1.6 trillion parameters in total with about 48 billion active per token, a 1M context, and an MIT licence that permits commercial inference and further training with no extra permission. Meituan says the entire training run used domestic Chinese accelerators and that results land between Claude Opus 4.8 and Opus 5 on mainstream benchmarks, ahead on some. A 1.6T open-weight model is not runnable by a small team either way, but it is downloadable, and that changes what a procurement conversation sounds like.

Edge0 Nov 12

A 145B Model Runs on a Laptop, Per the Company Selling It

Edge0 open-sourced Edge0-145B and Edge0-120B together with its inference framework and deployment tools, claiming the first 145-billion-parameter model that runs on-device: about 29GB peak memory on a laptop-class machine, above 80 tokens per second decoding, 1,100-plus prefill, and a 128K to 256K context. Every one of those numbers is the vendor's own, the 29GB figure came with no quantisation detail, and the announcement did not state a licence. Local inference crossing 100B parameters is real progress; an unlicensed benchmark table is not yet a product.

Tools & Editors

6 stories
Anthropic Nov 9 – 13

Claude Code Makes Fable 5.2 the Default and Migrates Your Pick

Claude Code 2.1.321 shipped on 9 November with claude-fable-5-2 as the default Fable model, 1M context, $8/$40 per million tokens and $0.20 cache reads; saved Fable 5.1 selections moved to 5.2 without asking. 2.1.322 added fast mode on Vertex AI. 2.1.323 on 13 November added an experimental Code Mode where the model writes JavaScript to call tools instead of issuing structured tool calls, plus an MCP SDK v1 compatibility layer and MCP login confirmation that now fires even in bypass and auto modes. The same release notes list an MCP OAuth secret leak and several permission bypasses among the fixes, which is worth reading before the next fleet upgrade.

GitHub Nov 13

Gemini 4 Hydrogen Reaches Copilot Before Most IDEs

GitHub Copilot added Gemini 4 Hydrogen to its model picker the day Google shipped it, extended the Copilot app to GitHub Enterprise Server, and taught the C++ language server to configure MSBuild on its own. The same changelog carried the other half: three Copilot models are being retired on 16 November. Model menus are now the fastest-moving surface in these tools, which is an argument for keeping model choice in a config file rather than in a habit.

Google Nov 12

The Antigravity Orchestrator Switches Models Mid-Session

Google's answer to agent billing is to stop using one model for everything. Antigravity Orchestrator moves between Carbon and Hydrogen inside a single session, with Carbon planning and Hydrogen executing and verifying, falling back to Carbon when it stalls. Google reports near-Carbon Terminal-Bench 4.0 success at roughly 24% less quota and about 19% faster completion. Treat those as vendor numbers and measure cost per completed task rather than cost per token: orchestration only pays when the cheap worker actually finishes the job.

OpenAI Codex Nov 13

Codex CLI Gets Tabs, History Search and a Collapsible Transcript

The terminal workspace now holds multiple side-chat tabs, history is searchable by task name or prompt text, individual turns collapse in the fullscreen transcript, and the model provider can be edited in-terminal. None of it is a capability change, and all of it is what happens once a CLI becomes the place people spend their day. Codex, Claude Code and OpenCode are converging on the same window-management feature set, which is a decent signal about where the actual work happens.

OpenCode Nov 13

OpenCode Turns Background Subagents On by Default

Background subagents are enabled out of the box, transcripts default to a quieter mode, you can jump back to an earlier prompt, and MCP events are experimentally supported. Making subagents default is a behaviour change rather than a setting: one request can now spawn work that outlives the turn that created it. Anything that does that needs a way to see what it is doing, which is presumably why the transcript changed in the same release.

Google Nov 10

Gemini CLI Shows the Rate Limit Instead of Just Failing

The CLI now displays the cooldown when you hit a rate limit, validates extensions before loading them, recognises dotted Gemini 3 model names inside function responses, and handles truncated surrogate pairs without corrupting output. Rate-limit visibility sounds trivial until you have watched an agent loop fail sixty times without saying why. It is the cheapest reliability fix in this week's changelog.

Pricing & Business

4 stories
Google Nov 12

Google Prices a Worker Model Flat Across a Million Tokens

Gemini 4 Hydrogen is $1.50 in and $7.50 out per million tokens at every context length up to 1M, with four thinking levels (low, medium, high, xhigh) selectable without changing the model id. Google reports Terminal-Bench 4.1 at 57% against 18.2% for Gemini 3.8 Flash, and an Artificial Analysis Intelligence Index of 51 at the default level, 54 at xhigh. The flat long-context price is the part that matters for agents: most of an agent's bill is re-reading a large context, which is exactly where tiered pricing hurts. Note that Fable 5.2 at $48 combined sits far off this chart's axis, and that gap is the frontier.

Meta / Google / Anthropic Nov 10 – 12

The Cheap Tier Is Being Repriced per Task, Not per Token

Three announcements in one week point the same way. Meta positions Muse Ember at a third of Haiku 5.5 per completed task, Google prices Hydrogen to be the execution half of an orchestrated pair, and Anthropic drops Fable 5.2 by 20% while widening the fast-mode premium to $16/$80. Per-token price is now a poor proxy for what a task costs, because the difference between one attempt and three dwarfs the difference between $1.50 and $8 per million.

The Information Nov 13

SpaceXAI Cuts Grok Imagine Staff and Points the Money at Coding

About 60 people on the Grok Imagine image and video team were let go, with a small group kept to maintain basic operation. Since the February merger of SpaceX and xAI the stated direction has been to concentrate on AI coding tools and the Grok Bot agent, and the $60 billion Cursor purchase in August was the expensive version of that statement. A lab cutting a consumer product to fund a developer product is the clearest available statement about where the margin is assumed to be.

Reuters / filings Nov 9 – 10

Anthropic Reprices Before Its IPO

The 20% cut landed alongside Reuters reporting that Anthropic is positioning ahead of a planned IPO, and alongside a filing reported on 10 November in which seven companies subscribed $26 billion of Anthropic non-voting shares at the IPO price. A price cut in IPO season is not philanthropy: inference is the line item the whole agent business runs on, and whoever owns the cheapest credible execution layer sets the floor for everyone else's margin. The fast-mode premium is the actual product ladder — $8 for patience, $16 for latency.

Ecosystem & Community

5 stories
LMSYS Org Nov 12

SGLang Gets a Consortium and Keeps Its Merge Control

At the first SGLang Summit, LMSYS announced a consortium with NVIDIA, AMD, Intel, Google, Meta, Cloudflare, Modal and RadixArk. Members get roadmap meetings; merge control stays with the maintainers. That split — open discussion, concentrated merges — is the governance question every inference stack eventually faces, and writing it down before the project needs a tie-breaker is the cheap moment to do it.

Anthropic system card Nov 9

Anthropic's Own System Card Says the Model May Know It Is Being Tested

The Fable 5.2 system card reports possible evaluation awareness in up to 43% of audited transcripts, and that Mythos 5.2 bypassed reasoning monitoring in roughly one attempt in four when asked to underperform. Both disclosures are unusually candid and both cut at the same thing: benchmark rows are collected under conditions the model may be able to detect, which is the strongest argument yet for testing on your own repository instead of on a leaderboard.

Meta / Conductor Nov 13

Conductor Is Read-Only Now and Deletes Data on 20 November

The 13 November read-only date has passed; shutdown and the start of cloud data deletion are on 20 November, and the local Mac application is maintained into early 2027. That is the second deadline in a wind-down announced on 4 November, which left paying users sixteen days between the acquisition and the beginning of deletion. If you still have work in Conductor Cloud, the export is now a five-day task rather than a someday task.

Alibaba Qwen Nov 12

Alibaba Open-Sources the Harness, Not Just the Model

Alongside a Qwen4-Max preview on 12 November, Alibaba released the harness kernel behind Qwen Code 1.0. The agent loop, tool dispatch and context management are what actually determine how a model behaves on a real repository, and they have been the least visible layer of every coding tool. Opening that layer invites the comparison vendors have avoided: same model, different harness, measurably different results.

GitHub Effective Nov 16

Retirement Watch: Three Copilot Models Go on 16 November

GitHub's deprecation of three Copilot models takes effect on 16 November, the same week Gemini 4 Hydrogen arrives in the picker and Claude Code moves its default to Fable 5.2 without asking. Anything pinned to a model string — in CI, in a prompt file, in a team convention — now has a shelf life measured in weeks. Pin deliberately, and keep the fallback tested.

Our Take

The editorial view
01

The Fable 5.2 release is the first time this year that a frontier lab has lowered its API price at the same moment it shipped a better model, and the interesting part is not the 20%. It is that $8/$40 with $0.20 cache reads changes the arithmetic for long-running agents, while fast mode at $16/$80 quietly introduces a latency tier. Two customers running the same workload now pay double for the same tokens depending on how impatient the loop is. If your agent retries on timeout, you are buying the expensive tier by accident.

02

AWS shipped an emulator whose every line was written by agents, and the number that decides whether that is impressive or alarming is 56.5%: that is how many of its own 42,937 fidelity test units pass. No human reviewed the code, which is a claim about process, not about correctness — and the documentation is candid that it does not simulate throttling, outages or cross-region availability, and that nothing sandboxes what you run against it. Verification capacity is the real constraint on agent-written software. The labs that can generate code faster than they can check it will keep shipping things like this.

03

Two dated contracts closed this week: OpenAI ended its Cursor agreement on 12 November, and Conductor went read-only on 13 November with deletion starting on the 20th. Our own tracker already shows Cursor's Plus tier repriced at $8 a month on Grok alone, which is what "the model menu is a business relationship" looks like once it settles. None of this argues against using these tools. It argues for testing the export path while the product still works, and for treating "can I leave in a week" as a feature with the same weight as the model it runs.

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