A quiet-news day for new dev tools — reaching into the watchlist for one that's earning a fresh look today.
curl -L https://dl.dagger.io/dagger/install.sh | sh
cd your-project
dagger init --sdk=python --name=ci
dagger call test
No breaking changes this period.
(Note: Starting September 14, 2026 — two days from now — all deepseek-v4-pro API requests will route to DeepSeek V4.1-Flash at V4.1-Flash rates. If you pin to deepseek-v4-pro and have budget expectations based on V4 Pro pricing, update your billing estimates and test against deepseek-flash before the switchover. DeepSeek has confirmed they will continue serving V4 Pro requests after Sept 14 with unchanged billing, so the routing change may not apply universally — verify your account tier.)
deepseek-flash on the DeepSeek API — it supports thinking and non-thinking modes, tool calls, structured outputs, the Responses API format, and the Anthropic API format. If you're testing against Anthropic-format code, V4.1 Flash is plug-compatible with minimal changes to your API base URL and key. (2) Pricing: off-peak $0.003/M cache-hit input, $0.15/M uncached input, $0.60/M output; peak is 2x these rates. For context: this undercuts V4 Pro pricing substantially. (3) Open weights: MIT-licensed weights on Hugging Face; community projects are already running it on NVIDIA DGX Spark hardware (see 0xBakeer/deepseek-v41-flash-spark on GitHub). This means you can self-host for zero variable cost if you have the hardware. (4) Migration note: from September 14, 2026, deepseek-v4-pro API calls are scheduled to route to V4.1-Flash at V4.1-Flash rates — update your billing models now. DeepSeek has since said they'll continue V4 Pro access on request with existing billing; the situation is evolving — check your account tier. (5) Benchmark caveats: results are self-reported; DeepSWE v1.1 is a separate benchmark from SWE-bench Verified and tests a different task distribution (closed-form agentic software engineering). The improvement is large enough to be structurally meaningful, but third-party reproduction is still in progress.console.sakana.ai — drop-in replacement for any OpenAI-format client code. Fugu Ultra v2 supports configurable reasoning effort (high, xhigh, max), function calling, structured outputs, image and PDF input, and built-in web search; standard API parameters carry over. (2) Fugu Max as a routing play: if you're building multi-step tasks where some sub-tasks need frontier reasoning and others don't, Fugu Max's learned routing is worth benchmarking against your own orchestration — at $6/M output it's significantly cheaper than a dedicated frontier model for every call, and it can still escalate to more powerful sub-models when needed. (3) Benchmark caveats: Chartography and DeepSWE v1.1 are specific benchmarks that may favor the orchestration-and-synthesis approach; SWE-bench Verified (where Claude models score 95-96%) is a different task distribution. Don't extrapolate Fugu's Chartography advantage to all coding tasks. The "no proprietary models" claim is worth verifying if you have data residency concerns — Sakana's model pool for Ultra v2 should be confirmed in their documentation.No breaking API changes this period. See Quick Hits for the OpenAI project API key expiration feature.
arXiv (export.arxiv.org) and Hugging Face Papers (huggingface.co/papers) were both egress-blocked this scan. No research items reached the quality bar from accessible sources. See Near-misses below.
git log, git status, git show, git ls-files, and similar — were unexpectedly prompting for permission approval after a session had been running for a while, instead of executing without interruption.git log, git status, git show) triggering permission prompts after your session had been running for some time — this is the fix. The regression was in the session-state tracking for Bash permission evaluation. Upgrade to v2.1.270 with npm install -g @anthropic-ai/claude-code@2.1.270. No configuration changes needed. If you're running Claude Code in CI or automated sessions, this regression would have caused those pipelines to stall on git introspection commands — v2.1.270 resolves it.npm install -g @anthropic-ai/claude-code@2.1.270.npx skills add <owner>/<repo>. This includes repos from individual developers, companies like HeyGen, and projects from prominent practitioners like Matt Pocock and Corey Haines. The volume (at least 12 skills-focused repos in the weekly top 25) is more than noise — it's a pattern.news.ycombinator.com was egress-blocked. No HN threads with score >200 and confirmed technical depth were verifiable in the 24h window. Nothing else cleared the quality bar.
clai sessions; fixed text part boundaries in OpenAIChatModel after tool calls (prevents content leaking across message boundaries); temporal tool opt-out checks now based on operation kind, not MCP import presence. [github.com/pydantic/pydantic-ai/releases]dagger ws migrate for legacy dagger.toml SDK fields). On steady march to v1.0.0. [github.com/dagger/dagger/releases]Starting 04:00 UTC September 14, deepseek-v4-pro requests are scheduled to route to V4.1 Flash at V4.1 Flash pricing. DeepSeek has since indicated they will continue V4 Pro availability in response to user demand — the exact rollout terms are in flux. If you use deepseek-v4-pro, verify your account tier and billing expectations before this date. 2 days remaining.
(Carried forward from September 11 digest) The gemini-omni-flash-preview endpoint retires September 30. GA replacement: gemini-omni-1.1-flash. Migration: update model string; verify resolution defaults (720p, not 360p); test 4K output if in use. 18 days remaining.
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.