No. 017 · 26 August 2026 · 3 min← No. 016 · No. 018 →

The one thing

01

ChatGPT Plus gets the Codex and Work 5-hour cap back — Pro $100/$200 stay uncapped for now.

OpenAI engineering lead Thibault Sottiaux said Plus accounts regain a five-hour limit across ChatGPT Work and Codex starting 25 Aug, after weeks with only the weekly cap.¹² The stated reasons: smooth compute load so the weekly allotment can stay generous, and stop casual Plus users burning a full week in one sitting.¹ After either the five-hour or weekly limit, wait for reset or buy credits; banked promotional resets still apply.¹ Pro $100 and Pro $200 keep the five-hour limit off for the upcoming months; Enterprise/Edu stay on their credit system.¹ HN carried the report at 112 points / 120 comments inside the window.²

npm @anthropic-ai/claude-code dist-tags this morning: stable still 2.1.231; latest/next → 2.1.246 (25 Aug 22:31 UTC).³⁴ That line is fifteen patches past stable and includes the Bedrock double-bill and gateway cache fixes from #015, plus a large reliability batch in 2.1.246 (MCP empty-schema args, headless interrupt reporting, background session open failures, Write OOM on large overwrite).⁴

VERDICT · BREAKING — Plus Codex/Work: plan around a 5h rolling window again, or move Pro / credits / another harness. Claude Code on Bedrock or a gateway: stay on latest ≥2.1.239 (ideally 2.1.246); stable still lacks those fixes.

Deeper — what 2.1.246 actually fixed

MCP: empty {} schemas keep real arg types; interrupted tools no longer look "completed with no output".⁴
Background sessions: open after sleep/deleted cwd; EACCES during npm reinstall; worktree retention no longer deletes your .claude/worktrees/.⁴
Gateways: third-party tool_use without id; resume 400s from proxy history; telemetry no longer sends gateway keys to Anthropic.⁴
stable remains 2.1.231 — channel decision from #015 still stands.³

Shipped

02

Use it

llama.cpp 0.3.0 · dots3-note multimodal, DeepSeek 4 tensor-split, ggml 0.22

v0.3.0 (25 Aug): dots3-note with DSA-ISWA KV cache and vision/audio in mtmd; DeepSeek 4 gets -sm tensor and multi-sequence rollback fixes; GLM-4.5-Air MTP; ggml → 0.22.0 (meta-backend tensor split, per-op Metal kernels, non-in-place ggml_clamp).⁵ mtmd: WebP via ffmpeg, Pillow-accurate resize. Server: LLAMA_SERVER_SLOTS_N_DIFF. Web UI: tabbed chat. Nightly pointer b10621.⁵

MATTERS TO · local llama.cpp / server / multimodal runners, especially DeepSeek 4 multi-GPU

nofeed.dev/issues/2026-08-26/llama-cpp-0-3-0/

Use it

Claude Code 2.1.246 · MCP honesty, background sessions, gateway edge cases

Latest channel only (stable still 2.1.231).³⁴ Wildcard Bash allow rules now warn at startup. Auto mode rules editable under /permissions. Material fixes: MCP interrupted/empty-schema behaviour, background session open/EACCES, Write OOM on large overwrite, resume 400s from proxy history, telemetry credential host scoping, auto-continue on mid-stream cut in -p/SDK/cloud.⁴

MATTERS TO · anyone on latest/native Claude Code, MCP servers, or ANTHROPIC_BASE_URL gateways

nofeed.dev/issues/2026-08-26/claude-code-2-1-246/

Promised, not shipped
OpenAI Private Safety Processing / frontier ZDR — September white paper (issue 013) · Origin agent-native source-control — early beta ongoing (issue 011) · Claude Code /design skill — research preview · MAI-Code-1-Flash off Copilot — 10 Sept 2026 · Grok Bot enterprise — waitlist · OpenAI Ultrafast — expands with capacity · Ox Alpha free window — ~through 27 Aug · GPT-5.6 Sol promo pricing — at least through 21 Nov 2026 · Plus Codex/Work 5h limit — live from 25 Aug; Pro stays off for upcoming months

The conversation

01

The claim — including ours

Plus Codex is rate-limited again on a five-hour clock; Claude Code latest is a reliability dump, not a stable bump.¹⁴

From the floor

Tibo Sottiaux via 9to5MacVendor lead, 24–25 Aug
5h back for Plus Work+Codex; Pro $100/$200 exempt for upcoming months; weekly cap stays.¹
HN · 5-hour Codex/Work limits112 pts / 120 cmt
Load-smoothing vs surprise stops; team plans that cannot buy Pro feel the Plus ceiling.²
X digests · 25 Aug 05:30Digest ~24.5h old — background
Claude usage-limit complaints; harness cost comparisons (AgentSky, Vetta). Colour only — primaries above carry the facts.

Our take

If Plus Codex is your day job, budget the 5h window or leave Plus. If Claude Code is your day job on Bedrock/gateway, you are still not on stable.

GitHub releases: 28 watchlist repos, sweep_complete true, authenticated, 7 in-window. Vendor feeds: 12 sources, sweep_complete true, 12 in-window (Cursor Grok degradations 0.9h and 2.7h minor resolved; Copilot Customize tab GA). X accounts (~95 roster) and X keyword digests collected 05:30 UTC 25 Aug (~24.5h old at window close) — background only. Reddit: fresh pulse (6 subs, 13 posts) plus 25 Aug digest. HN: Codex 5h thread 112/120 in-window. arXiv not swept. Jalapeño chip post (OpenAI news) unreachable (crawl timeout) — not used.

Skip this

05

01

OpenAI Jalapeño inference-chip first results. Vendor bench post; primary page timed out on fetch. Not a ship decision without numbers we could open.

02

Cursor Grok 4.6 degraded 2.7h minor; Grok models 0.9h minor. Resolved weather with duration — not a pin.

03

Cline Desktop v0.0.17 Customize hub redesign. Desktop-only UI reshuffle; VS Code/CLI path unchanged for most readers.

04

gemini-cli v0.57.0; pydantic-ai v2.35.0; opencode v1.18.23; MCP Python 2.1.1. Capacity retries, capability rename deprecation, Cloudflare AI Gateway slugs, FastMCP import pointer — no default-stack pin today.

05

GitHub Copilot app Customize tab GA; Headlong / AgentSky harness launches on X. Copilot surface discovery only. Harness launches are demos without a release we verified for a weekly decision.

Everything we saw

52

52 candidates scanned · 10 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.

Item	Source	Signal	Call
9to5Mac — OpenAI restores 5-hour Codex and Work limits for Plus	9to5mac	Tibo announcement	led the issue
HN — OpenAI restores 5-hour Codex and Work limits	news.ycombinator.com	112 pts / 120 cmt	one thing
npm @anthropic-ai/claude-code dist-tags	npm registry	stable 2.1.231 / latest 2.1.246	one thing
Claude Code v2.1.246	github releases	7.6h	shipped + lede
llama.cpp v0.3.0	github releases	19.7h	shipped
pydantic-ai v2.35.0	github releases	4.6h deprecations	skipped
gemini-cli v0.57.0	github releases	11.5h	skipped
Cline desktop-v0.0.17	github releases	21.0h	skipped desktop UI
opencode v1.18.23	github releases	23.6h CF gateway	skipped
MCP Python SDK v2.1.1	github releases	14.1h	skipped docs pointer
Cursor status — Grok 4.6 degraded 2.7h	cursor status	resolved minor	skipped weather
GitHub Copilot Customize tab GA	github changelog	10.0h	skipped surface
OpenAI — Jalapeño first results	openai.com	crawl timeout	skipped unverified
X digests 25 Aug 05:30	digests	~24.5h old	background only
Reddit pulse — 6 subs, 13 posts	reddit pulse	layer ran	layer ran
HN — Headlong microharness	news.ycombinator.com	120 pts	skipped launch

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 developingClaude Code stable lag vs latest (015/017). Origin early beta (011). DeepSeek peak/off-peak live (010). Ultrafast capacity (009). MAI-Code-1-Flash off Copilot 10 Sept. DeepSeek Harness preview. OpenAI ZDR white paper September. Files/Skills GA (013). anthropic Python 1.0 (014). Ox Alpha free ~through 27 Aug. Sol promo ≥21 Nov 2026. Plus Codex/Work 5h from 25 Aug.

Sources — each with the date it was read01 9to5Mac — OpenAI restores 5-hour Codex and Work limits for Plus (2026-08-26) · 02 HN — OpenAI restores 5-hour Codex and Work limits (2026-08-26) · 03 npm — @anthropic-ai/claude-code dist-tags (2026-08-26) · 04 github — claude-code v2.1.246 (2026-08-26) · 05 github — llama.cpp v0.3.0 (2026-08-26)