AI Dev Pulse–2026-09-29

Tuesday is a daily-driver day, because the model most developers reach for changed on its own, while the Codex and runner items wait for you to act: Codex’s new way of handling input mid-turn is opt-in and off by default, and the runner version floor needs a check if you own self-hosted runners. Anthropic says Sonnet 5.5 uses fewer tokens per task and generates 30% or more faster, and its own numbers put it at 70.6% on Terminal-Bench 4.0 against 10.3% for Sonnet 5 and 55.5% on CursorBench 4.0 against 57.8% for Opus 5.5, so the vendor is pitching it as the everyday model.

Treat today as the day you decide on purpose which Claude model each repo uses, try instant_interrupt only at a desk where a human is watching, and check that any self-hosted runner that has fallen behind on updates can still pick up a job.

Top Stories

1. Sonnet 5.5 is Claude Code’s default Sonnet, but only on the Anthropic API

Practical dev impact: After you upgrade to v2.1.284, any session on the Anthropic API that resolves the sonnet alias starts on Sonnet 5.5, which Claude Code’s docs say needs v2.1.284 or later. That is a cost and latency change with no meeting attached, so pin claude-sonnet-5-5 or stay on Opus 5.5 explicitly in the repos where the difference matters. The scope is narrow, because the Claude Code default model on the Anthropic API is still Opus 5.5, so sessions that never touch Sonnet do not move, and on Bedrock, Google Cloud and Foundry sonnet still resolves to an older Sonnet. If your managed settings use availableModelsMatch: "exact", the setting Saturday’s v2.1.283 brief covered, Sonnet 5.5 is blocked only if that list names Sonnet by model ID, and it stays blocked until someone adds claude-sonnet-5-5.

Anthropic launched the model on 28 Sep at $2 in, $10 out and $0.20 per million cache-read tokens, and says it is the first Sonnet to ship with cyber safeguards, so higher-risk cybersecurity tasks visibly fall back to Sonnet 5. Those benchmark figures are Anthropic’s own, and Haiku 5.5 is promised only for “the coming weeks”. Among the other v2.1.284 changes, /mcp reconnect all retries every MCP server that failed to connect, gateway spend limits now show dollar figures in /usage and the status line, and the auto-mode prompt about reads outside your working directories gains a “Yes, but ask again next time” answer.

2. Copilot adds Sonnet 5.5 to the picker your org already governs

Practical dev impact: Sonnet 5.5 shows up in Copilot for Pro, Pro+, Max, Business and Enterprise plans as another billable row at the provider’s list price under usage-based billing, so check the model policy before the next standup that asks why Copilot spend moved. Business and Enterprise admins manage it through the model policy in Copilot settings, and a new model is on by default unless the global default is off or that model is explicitly disabled. Rollout is gradual, so a developer’s IDE may not show the row on the first refresh, and one IDE showing it does not mean the whole team has it.

GitHub says its early testing found Sonnet 5.5 matched Sonnet 5 on coding tasks while using fewer steps, tokens and tool calls and finishing faster, which is an efficiency claim rather than a new quality tier.

3. Codex v0.159.0: opt-in mid-turn steering and a protected .aws

Practical dev impact: OpenAI published Codex v0.159.0 at 4:05 AM ET today, and its notable behavior change is instant_interrupt, an under-development feature flag that is off by default. With it on, new input can preempt a model response, response streaming or retry backoff, and the next request sends the full history, so it suits an interactive desk and should stay off in CI and unattended profiles. Turn it on for one launch with codex --enable instant_interrupt or codex -c features.instant_interrupt=true, because the key has to be features.instant_interrupt, and --config instant_interrupt=true does not enable it. The flag does not appear in OpenAI’s public config docs and can change, so treat it as an experiment.

The other change that can bite is that .aws now joins .git, .agents and .codex as a protected metadata path, including under additional writable roots. A workflow that relied on the agent writing into a .aws directory will need an explicit user rule to override the default protection, and that rule deserves a real justification. Finally, tui.prompt_suggestions was removed, but it only ever existed in 0.158.0 alpha builds and never shipped in a stable release, so you only need to delete it if a dotfile picked it up from a prerelease.

4. Also today: check your GHEC self-hosted runners

Practical dev impact: GitHub says full enforcement of its self-hosted runner version floor on Enterprise Cloud begins today, 29 Sep, after the change shipped on 28 Sep, and runners below v2.329.0 cannot register or reregister. If your runners have not kept up with runner releases, jobs targeting them may remain queued or fail, so check each runner’s version against the deprecations API, update the image or auto-update setting, and remember that v2.329.0 is only the registration floor. GitHub Enterprise Server is not affected.

Practical Impact Analysis

The through-line today is defaults that move underneath a team, because a model alias, a Codex input path and a runner requirement each changed without anyone editing a repo. The useful response is to make each one a decision, and that comes down to three pins. For Claude Code on the Anthropic API, that means v2.1.284 plus either an explicit claude-sonnet-5-5 or an explicit stay on Opus 5.5, so an alias cannot shift the cost profile unnoticed. For Copilot, it means a deliberate setting for the Sonnet 5.5 row in the model policy instead of whatever the default enablement chose. For Codex, it means v0.159.0 on the PATH and instant_interrupt only where someone is watching the terminal.

The cost story is a bet rather than a price cut, since the list price did not move and the savings depend on fewer tokens and tool rounds for the same ticket. That only shows up if sessions are actually on Sonnet 5.5, and Anthropic’s benchmarks are vendor-reported, so give the model a week of ordinary scoped tickets before moving long agent runs or messy migrations off Opus 5.5.

Tutorial

This takes about ten minutes on a machine you can afford to experiment on. Confirm your Claude Code and Codex versions, launch each with the setting you decided on, and list your runner versions if you own GHEC self-hosted runners.

1. Check that Claude Code is v2.1.284 or newer, then launch it with claude --model sonnet to see what the alias gives you, because no flag starts Claude Code on its default, which is still Opus 5.5.
2. Install Codex v0.159.0 and confirm instant_interrupt reads true when enabled for one launch, before starting an interactive session with it.
3. If you own GHEC self-hosted runners, set ORG, list each runner’s version, and ask the deprecations API about each distinct version. The calls need org admin scope, and the deprecations call comes from GitHub’s REST docs and has not been run against a real org, so confirm what it returns. Skip this on GitHub Enterprise Server.
4. After the first Sonnet 5.5 session, compare one scoped bug against last week’s Sonnet or Opus run for steps, tool calls and reviewable diffs.

#!/usr/bin/env bash
set -euo pipefail

# 1) Claude Code: expect v2.1.284 or newer
claude --version
# Alias check for one session (Anthropic API); run it when you are ready:
#   claude --model sonnet
# Or pin the model explicitly:
#   claude --model claude-sonnet-5-5

# 2) Codex: install v0.159.0, then confirm the flag reads true for one launch
npm install -g @openai/codex@0.159.0
codex --version
codex --enable instant_interrupt features list | grep instant_interrupt
# Interactive desk only, never CI:

... click "Show full code" below to expand

#!/usr/bin/env bash
set -euo pipefail

# 1) Claude Code: expect v2.1.284 or newer
claude --version
# Alias check for one session (Anthropic API); run it when you are ready:
#   claude --model sonnet
# Or pin the model explicitly:
#   claude --model claude-sonnet-5-5

# 2) Codex: install v0.159.0, then confirm the flag reads true for one launch
npm install -g @openai/codex@0.159.0
codex --version
codex --enable instant_interrupt features list | grep instant_interrupt
# Interactive desk only, never CI:
#   codex --enable instant_interrupt

# 3) GHEC self-hosted runners (skip on GHES): versions, then GitHub's deprecation dates
ORG="${ORG:?set ORG to your Enterprise Cloud organization}"
gh api --paginate "orgs/${ORG}/actions/runners?per_page=100" \
  --jq '.runners[] | [.name, .status, .version] | @tsv'
gh api --paginate "orgs/${ORG}/actions/runners?per_page=100" \
  --jq '.runners[].version' | sort -u | while read -r v; do
  echo "== ${v}"
  gh api "orgs/${ORG}/actions/runners/deprecations/${v}" \
    --jq '{registration_deprecates_at, runtime_deprecates_at}'
done

You are done when Claude Code reports v2.1.284 or newer, the Codex check prints instant_interrupt as true, and every runner version is one you have checked against the deprecations API.

Recommended AI prompt

Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever AI you use.

You are a staff engineer helping me decide what to change in one team’s AI coding setup after four updates on 28 and 29 September 2026. I will paste the output of claude --version and codex --version, the model setting each of our repos uses (or the words “family alias: sonnet”), whether our Anthropic access goes through the Anthropic API, Bedrock, Google Cloud or Foundry, whether our managed Claude settings set availableModelsMatch to exact, whether our Copilot model policy has Claude Sonnet 5.5 enabled, whether instant_interrupt is on in any Codex profile, and the version column from our GitHub Enterprise Cloud self-hosted runner list. Using only what I paste, tell me first whether each repo is actually on Sonnet 5.5 given that Claude Code v2.1.284 made it the default Sonnet on the Anthropic API only, and which repos should pin claude-sonnet-5-5 or stay on Opus 5.5. Then tell me whether an exact-match allowlist would block it, whether the Copilot Sonnet 5.5 row should stay on or off for each plan, whether instant_interrupt belongs only on interactive profiles because it is opt-in and under development, whether any workflow that writes into a .aws directory needs an explicit rule under Codex v0.159.0, and which runners to update first, remembering that v2.329.0 is only the registration floor. Keep your first reply to a short checklist, do not invent settings, flags, file paths or version numbers I did not paste, and ask me for anything missing before you answer.