Categories Alternatives News Submit a Tool Advertise About
W42 Oct 5 – Oct 11, 2026
Issue W42 · 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
$0.10/$0.50 Claude Haiku 5.5 1M context · 100K threshold reprices the request
0.05x Sonnet 5.5 cache Cache reads cut to 5% of input
Oct 14 GPT-5.5 Leaves ChatGPT, Work and Codex · API unaffected
49B active Mistral Large 4 1.05T total · weights at month end

Top Stories

What you need to know from this week
★ Editor's Pick
Editor Pick Oct 6-7 · Anthropic

Anthropic Cuts Cache Prices and Sets a Threshold That Changes the Bill

Two changes on the same day. Claude Haiku 5.5 shipped with a 1M context window at $0.10 in and $0.50 out per million tokens — but only while a request stays under 100,000 tokens. Cross that line and the entire request is repriced at $0.50 and $2.50, a five-fold jump on work that is only marginally larger. Separately, cache reads on Claude Sonnet 5.5 dropped to 0.05x of the input rate, up to half off. Read together they describe where pricing is heading: not a rate per million, but a rate that depends on the shape of your traffic.

Benchmark Snapshot

Frontend Code Arena · human-preference voting

TerminalBench 2.1 — Terminal Agent Score

SWE-2 (Cognition)
92.8%
DeepSeek V4.1 Flash
90.6%
Qwen3.8 Max
86.6%
Qwen3.8-27B (local)
73.0%

Research & community-reported scores · TerminalBench 2.1

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

Claude Haiku 5.5 new · $0.10/$0.50 under 100K
$0.60
GPT-6 Luna $0.10/$0.50
$0.60
DeepSeek V4-Flash off-peak
$0.88
GPT-6.1 Sol $2/$10
$12.00
Claude Sonnet 5.5 $2/$10 · cache now 0.05x
$12.00
$0 $10

Research & community-reported prices

Models & Benchmarks

5 stories
Anthropic / AWS Oct 7

Claude Haiku 5.5 Ships With a 1M Context Window

The small model in the 5.5 family arrived with a 1M-token context and a 128K maximum output. Pricing is $0.10 per million input tokens and $0.50 output. It went live on Amazon Bedrock and Claude Platform on AWS on the same day, including GovCloud, so it is available inside existing AWS commitments rather than only through Anthropic.

Anthropic Oct 7

The 100,000-Token Line That Quintuples the Bill

The published rate applies while a request stays under 100,000 tokens. Above that, the whole request is billed at $0.50 input and $2.50 output — not just the excess. A 120K-token request costs about five times what the headline price implies. This is the same class of trap as credit pooling and per-seat minimums: correct as written, and invisible unless you read the footnote next to the number.

Anthropic Oct 7

Cache Reads on Sonnet 5.5 Fall to 0.05x

Anthropic cut the cache-read rate on Claude Sonnet 5.5 to 0.05x of the input price — half the previous multiplier, and up to a 50% saving on what a cached workload pays. It only matters if your calls actually reuse a stable prefix, which is exactly the case that agent harnesses were built to create. Caching is now the cheapest lever in the price list and the one least described on pricing pages.

Mistral AI Oct 6

Mistral Large 4 Enters Public Preview

Mistral's new flagship is a mixture-of-experts model: roughly 1.05 trillion parameters in total, about 49 billion activated per token. It was trained on European infrastructure — around 4,000 accelerators over two months — and is positioned for cybersecurity, code, finance and legal workloads that need to run in a controlled environment. The preview API is live; the weights are promised by the end of the month.

OpenAI Oct 6

OpenAI Ships a Decisions API and Collapses Its Usage Tiers

A beta Decisions API landed, running on gpt-6-luna, and the same day OpenAI reduced API usage tiers from five to three — Build, Launch and Grow — with organisations upgrading automatically as cumulative spend crosses each threshold. Fewer tiers is a simplification for buyers, and it also makes the ladder between them more visible than five overlapping steps did.

Tools & Editors

6 stories
Anthropic Oct 7

Claude Code v2.1.293 Picks Up Haiku 5.5

The release adds the new small model to Claude Code, shows the agent type in the subagent status line, and introduces deferred-tool controls aimed at mod authors — the extension surface is already being refined two weeks after mods shipped. Model choice inside a CLI agent is now something that changes week to week, which matters if you set a default and stopped thinking about it.

Anthropic Oct 6

Claude Code v2.1.290: Paging and Survivable Scheduled Tasks

WebFetch gained paging, scheduled tasks now survive a session resume, and web searches have a refill control. The middle one matters more than it reads: a scheduled task that dies when the session restarts is not a scheduler, and fixing persistence is the difference between a demo and something you leave running overnight.

OpenAI Oct 8-9

Codex CLI Gains Worktrees and a /copy Command

v0.162.0 adds a worktree tool — separate working trees so an agent can operate on a branch without disturbing what you have open — along with a /copy command and clickable links in the terminal. v0.162.1 followed a day later with a TUI crash fix and a background-server startup fix.

OpenAI Oct 7

Codex CLI v0.161.0 Adds /mcp login

MCP servers can now be authenticated from inside the session rather than requiring a separate setup step, and Daybreak gating moved behind an opt-in flag. The direction across every CLI this month is the same: the agent is absorbing the configuration that used to live in a separate tool.

Google Oct 6

Gemini CLI v0.63.0

Another weekly increment for Google's terminal agent, and another entry in the pattern this issue keeps showing: five of the six tracked CLI agents shipped at least one version between 5 and 10 October. Release cadence has itself become a selection criterion — a harness that ships weekly will absorb the model churn described elsewhere in this issue.

OpenCode Oct 8

OpenCode 2 v2.0.26

The open-source terminal agent continues its steady release line, with v1.18.34 and then v2.0.26 across this window. For teams that cannot send code to a vendor, the open harnesses are the ones whose cadence they can actually track and pin — which is a different kind of guarantee than a benchmark score.

Pricing & Business

4 stories
OpenAI Sep 16 (retiring Oct 14)

GPT-5.5 Leaves Every ChatGPT Plan on 14 October

OpenAI's notice is unusually explicit on both halves: GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on all plans on 14 October, and the retirement does not apply to the OpenAI API. If your usage goes through Codex on a subscription, the model changes under you; if it goes through the API, nothing changes. Two customers of the same model now get different answers, and the difference is which door they came in through

Google Oct 6-7

Six Gemini Previews Get Shutdown Dates in One Day

gemini-omni-flash-preview is dated 22 October. Three TTS previews, the live preview and the native-audio preview are all dated 17 November. Google also listed a shutdown for gemini-3.1-flash-image and then withdrew it, which is its own reminder: a published retirement date is a statement about the vendor, and it can be revised in the other direction too.

Mistral AI Oct 6

Mistral's Python Client Breaks Seven Stable Symbols

client-python v3.1.0 removed seven symbols from the stable API surface. A breaking change in an SDK is a different kind of event from a model retirement: it fails at import, which is honest, rather than silently switching the model behind a name you already pinned.

DigitalOcean Oct 7

DigitalOcean Sells a Drop for Agents

Agent Droplets are virtual machines sized and billed for agent workloads, marketed as one price covering everything an agent needs on one bill. It is the infrastructure layer repricing itself for machine clients rather than human sessions — the same shift visible in Supabase describing itself as agentic infrastructure last week.

Ecosystem & Community

4 stories
OutSystems Oct 7

OutSystems Makes MCP Generally Available for Every Tenant

The low-code platform put its MCP server into general availability across all ODC tenants, explicitly positioned as the layer that connects whichever coding agent you already use — Claude Code, Codex or Cursor. This is what MCP adoption looks like from the enterprise side: not a new client, but an old platform becoming reachable from the clients everyone already runs.

Google Oct 6

Gemini Nano Banana 2.1 Replaces the Flash Image Model

Google shipped an updated image model and pointed it at the place of gemini-3.1-flash-image, the model it had dated for shutdown and then un-dated the same week. Image generation inside a coding workflow is mostly about diagrams, screenshots and UI drafts, which is why the model churn here is now on a weekly clock too.

Amazon Bedrock / Z.ai Oct 5

GLM 5.3 Reaches General Availability on Bedrock

Z.ai's GLM 5.3 became generally available on Amazon Bedrock, joining the pattern of Chinese open-weight models being consumed through Western cloud marketplaces rather than self-hosted. For an enterprise buyer the practical effect is an option that appears inside an existing contract instead of a new procurement exercise.

OpenAI Oct 8

gpt-6.1-sol Picks Up Ultrafast With EU Data Residency

The premium speed tier extended to gpt-6.1-sol, now offered with European data residency. Speed tiers arrived as a product two weeks ago priced with no public numbers; they are now regionalised. Two things that used to be properties of a deployment — how fast, and where — have become line items.

Our Take

The editorial view
01

Three of this week's five most consequential items were about money, not capability. Anthropic shipped a cheap model and repriced cache reads at the same time, while writing a 100,000-token line into the Haiku 5.5 rate card that quintuples the bill for a slightly longer request. Google dated six preview models for shutdown within a single day. GPT-5.5 left every consumer plan while remaining available on the API. None of this shows up in a benchmark table, and all of it lands on anyone who pinned a model inside CI, an agent or a batch job. That gap — between what is announced and what is billed — is the one this directory exists to close.

02

Mistral's trillion-parameter preview was the most-discussed model of the week and you cannot download it. The weights are promised for the end of the month and Hugging Face holds a placeholder page in the meantime. That is not a scandal; it is the current meaning of the phrase open weights, which now describes a commitment plus a date rather than an artefact you can fetch. The practical rule is unchanged and this week restated it: check the repository for a file before you plan around a launch.

03

Pricing is turning into a function of request shape. Haiku 5.5 charges differently above 100,000 tokens; Sonnet 5.5 charges differently when your prefix is cached; Ultrafast charges differently again, and is now regionalised. Three levers, all set by the vendor, none of them visible in the headline rate that gets compared across labs. The number that wins a pricing table is increasingly the number a specific workload will never pay.

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