Issue 009 — window: 24h to 06:00 UTC, 14 Aug 2026nofeed.dev
No. 009 · 14 August 2026 · 3 min← No. 008 · No. 010
SIGNALNORMAL|SHIPPED 4|SKIPPED 5DeepSeek peak/off-peak API rates live 16 Aug 16:00 UTCRETIRING MAI-Code-1-Flash (Copilot) · 27d

The one thing

01

Claude Code auto mode is the default for new Pro, Max, and Team sessions as of today.

Anthropic's permission-modes docs state it: starting 14 Aug 2026, new sessions on Pro, Max, and Team start in auto mode.1 A one-time switch prompt appears if you never set a default; a defaultMode you already chose stays unless you accept that prompt; org-managed defaults do not move.1

Auto is a classifier, not YOLO. Deny and ask rules still win; unlisted actions go through a separate model that blocks escalation, unknown infrastructure, and prompt-injection-shaped moves.1 Default surprise: pushes to any branch of the working repo — including main — and PR creation are allowed; force-push, secrets in commits, and history rewrites stay soft-blocked.1 Claude Code 2.1.232 also turns subagent forking on by default and lets you @ another live session on the same machine.2

VERDICT · USE IT — If you never set a default, decide before the next session: keep auto and pin ask rules on git push / gh pr create, or set "defaultMode": "manual" in ~/.claude/settings.json. Org admins who want the old baseline set permissions.disableAutoMode to disable.
Deeper — who is affected and how to pin
  • Only new sessions; existing defaultMode survives the one-time prompt unless you accept it.1
  • Auto needs a supported model (Opus 4.6+, Sonnet 4.6+, Fable 5 on Anthropic; provider-specific floors on Bedrock/Foundry/gateway).1
  • Per session: Shift+Tab or claude --permission-mode manual. Persistent: ~/.claude/settings.json — repo settings cannot grant auto to themselves.1
  • v2.1.232: subagent_type: "fork" inherits conversation + prompt cache; non-teammate spawns background by default; cross-session SendMessage by bare name.2

Shipped

04
Breaking

DeepSeek API · peak/off-peak prices from 16 Aug 16:00 UTC

The pricing page no longer waves at a vague rise: peak hours are 01:00–04:00 and 06:00–10:00 UTC; off-peak is half of peak. Effective 16 Aug 2026 16:00 UTC. Pro peak cache-miss input $1.32 / 1M and output $3.96 / 1M (was $0.435 / $0.87 flat); off-peak $0.66 / $1.98. Flash peak $0.44 / $1.32, off-peak $0.22 / $0.66.3

MATTERS TO · anyone metering DeepSeek as a default agent model
nofeed.dev/issues/2026-08-14/deepseek-peak-offpeak/
Use it

pydantic-ai 2.30.0 / 1.107.5 · Host-header fix on local web UI

Second security patch this week for Agent.to_web() / clai web: no Host validation meant DNS rebinding from a page you visit could reach the local agent and its tools. Fixed by allowing localhost/loopback/LAN by default; real hostnames need allowed_hosts. GHSA-q2xc-rrxj-58x9.4

MATTERS TO · anyone who runs the pydantic-ai dev web chat on a workstation
nofeed.dev/issues/2026-08-14/pydantic-ai-host-header/
Wait

GitHub Copilot · Gemini 3.7 Flash rolling out

Google's latest Flash is in the Copilot model picker across VS Code, Visual Studio, CLI, cloud agent, app, JetBrains, Xcode, and Eclipse on Pro through Enterprise, billed at provider list under usage-based pricing. Business/Enterprise admins must enable the Gemini 3.7 Flash Preview policy first. Gradual rollout.5

MATTERS TO · Copilot orgs picking a cheap coding model this week
nofeed.dev/issues/2026-08-14/gemini-3-7-flash-copilot/
Use it

google-adk 2.7.0 · model-declared capabilities + media tools

Models report their own tool/schema support instead of ADK guessing from the model id. Tools can return images and other media on Gemini, Anthropic, LiteLLM, Apigee, and OCI. History keeps thought signatures and parallel function results. Breaking: pyarrow left the gcp extra — BigQuery analytics needs google-adk[bigquery-analytics].6

MATTERS TO · ADK agents that pair output schemas with tools, or return media from tools
nofeed.dev/issues/2026-08-14/adk-python-2-7/

Promised, not shipped
OpenAI Ultrafast (GPT-5.6 Sol, Cerebras) — limited preview, sign-up for expand; up to 750 out tok/s, vendor claim up to 14× Standard[^7][^8] · DeepSeek peak/off-peak API rates — 16 Aug 2026 16:00 UTC · MAI-Code-1-Flash removed from Copilot — 10 Sept 2026 · Grok Bot enterprise — waitlist, no date

The conversation

01
The claim — including ours

The permission default flipped on the coding CLI most of this desk already runs, the same week the cheap Chinese API published timed rates and OpenAI put frontier Sol on a Cerebras speed tier you cannot buy yet.137

From the floor
  • Claude Code docs · permission modesVendor, dated

    "Starting August 14, 2026, auto mode becomes the default permission mode for new sessions on Pro, Max, and Team plans."1

  • HN · Ultrafast / CerebrasReader thread

    Thread on the Cerebras write-up: vendor wall-clock on HLE (11h vs 78h for Fable 5) and token-efficiency colour; not an independent product bench.9

  • @trq212 · X (13 Aug digest)Vendor-adjacent field note, digest ~24.5h old

    Auto mode "makes the model better at following your hard instructions" — e.g. "don't push" sticks. One practitioner claim from the accounts digest, not a measurement.10

Our take

Treat today's default as a settings event: pin ask/deny before the next long agent run. Budget DeepSeek by UTC hour before Sunday. Ultrafast is a waitlist, not a pin.

GitHub releases: 28 watchlist repos, sweep complete, authenticated, 16 in-window. Vendor feeds: 12 sources, sweep_complete false (anthropic-status HTTP 502); 14 in-window from the other feeds. X accounts digest last collected 05:30 UTC 13 Aug (~24.5h old) — background only; no 14 Aug X digest in-repo. X keyword digest same age. Reddit pulse returned 0 posts above floor (6 subreddits). HN: Ultrafast/Cerebras thread. arXiv not swept.

Skip this

05

Everything we saw

58
58 candidates scanned · 12 used in this issue — the rest, with the reason each one was left out
The receipt. Ranking is only trustworthy if the discarded pile is visible, so here it is: everything the collectors surfaced in the window, with its signal and what we did with it.
ItemSourceSignalCall
Claude Code auto mode default 14 Augcode.claude.comdocs dated changeled the issue
Claude Code v2.1.232github releasesfork default + @ sessionsone thing / shipped context
DeepSeek peak/off-peak pricingdeepseek docseffective 16 Aug 16:00 UTCshipped
OpenAI Ultrafast GPT-5.6 Solopenai.comlimited preview 13 Augpromised / conversation
Cerebras Ultrafast write-upcerebras.ai750 tok/s vendorpromised support
pydantic-ai v2.30.0 Host headergithub releasesGHSAshipped
Gemini 3.7 Flash in Copilotgithub changelog16.1hshipped
google-adk v2.7.0github releasescorrectness + media toolsshipped
langchain-openai 1.5.0github releasesopenai 3.0 SDKOpenAI 3.0 partner bump; dropped to free budget
HN — Accelerating GPT-5.6 Sol Ultrafastnews.ycombinator.comthreadconversation
Claude Code v2.1.231 MCP OAuth redirectgithub releasesSlack OAuth fixfolded under 232
crewAI 1.15.16github releasestrace/contextnot decision-changing
llama.cpp b10419–b10424github releasesnightlyDFlash log polish; not a ship
Cursor service degradation 4.1h majorcursor statusAnthropic modelsskipped weather
X digests 13 Aug 05:30digests~24.5h oldbackground only
Reddit pulse — emptyreddit pulse0 posts, 6 subslayer ran, nothing above floor
EOF

End of feed. That is everything from the window worth your time.
Next issue tomorrow, 06:00 UTC — and if nothing ships, it will say so in two hundred words.

Still developingDeepSeek peak/off-peak rates land 16 Aug 16:00 UTC. Ultrafast access expands with capacity. Muse Glimmer local runtimes still rolling (issue 006). Sonnet 5 docs lag the permanent $2/$10 edit (issue 006). MAI-Code-1-Flash → 1.1-Flash closes 10 Sept.
Sources — each with the date it was read01 Claude Code docs — permission modes (auto default 14 Aug) (2026-08-14) · 02 github — claude-code v2.1.232 (2026-08-14) · 03 DeepSeek docs — Models & Pricing (peak/off-peak) (2026-08-14) · 04 github — pydantic-ai v2.30.0 (GHSA-q2xc-rrxj-58x9) (2026-08-14) · 05 GitHub changelog — Gemini 3.7 Flash in Copilot (2026-08-14) · 06 github — google-adk v2.7.0 (2026-08-14) · 07 OpenAI — Previewing Ultrafast mode (2026-08-14) · 08 Cerebras — Accelerating GPT-5.6 Sol Ultrafast (2026-08-14) · 09 HN — Accelerating GPT-5.6 Sol Ultrafast (2026-08-14) · 10 X — @trq212 auto mode field note (13 Aug accounts digest) (2026-08-14)

Get it by email.

One issue every weekday. The whole thing, not a teaser.

Or RSS, if you would rather we never had your address.