Tools & models
LangChain Core 1.7.0 changes streamed tool-call arguments
LangChain applications that parse raw streamed tool-call events.
NO FEED / Topics
Tools & models
Choosing and evaluating coding assistants, models and execution options.
32 results
Tools & models
OpenAI Python 3.37.0 expands Decisions types
Python applications using Decisions or Live response models.
Tools & models
OpenAI’s Decisions API is GA. Use it for bounded choices, not open-ended work.
Pilot one high-volume bounded decision against labeled production examples.
Tools & models
Claude Code vs Codex vs Cursor: choose the workflow you can finish
Compare Claude Code, Codex and Cursor by execution environment, permissions and review handoff. Use a small task-based trial instead of a universal ranking.
6 min readTools & models
Evaluate coding agents on your repository with a reusable task card
Build a fair coding-agent comparison with frozen tasks, trusted graders, repeated trials and complete cost records. Includes a reusable task-card checklist.
6 min readTools & models
SGLang 0.5.22 repairs a hybrid-model state path
SGLang operators combining hybrid models, speculation and constrained output.
Tools & models
When a local coding model makes sense
Decide whether a local coding model fits your work. Check real context, CPU/GPU placement, tool access and privacy before buying hardware or expanding scope.
6 min readTools & models
Copilot retires three model offerings on 16 November. Check the annual-plan exception.
Move affected Copilot workflows before 16 November; check annual-plan eligibility separately.
Tools & models
Gemini 4 Argon reaches Copilot
Copilot users choosing a model for complex coding tasks.
Tools & models
OpenAI’s Decisions API returns typed classifications
Harness builders routing requests or screening candidate tool calls.
Tools & models
Sol Ultrafast is available. Measure the whole coding loop before changing the default.
Compare accepted-task time and cost on a fixed workload; change the default only after that comparison.
Tools & models
Sonnet 4.5 has a retirement date. Put the migration on the calendar.
Find the pinned model calls now; finish and verify the migration before 30 November.
Tools & models
Claude Sonnet 5.5 is out at Sonnet 5 prices. Code written for Sonnet 5 can break on it in five documented ways, so change the model string only after checking them.
Nothing fails until you switch models. Before switching, search your code for "disabled" thinking and forced toolchoice, and test any flow that moves a conversation between accounts.
Tools & models
The loudest thing on the internet today is a v0.3.0 alpha that needs a second alpha underneath it.
go install the CLI and read the concepts doc if the density argument interests you. Nothing here survives a migration plan yet.
Tools & models
Gemini 3.8 Live and 3.8 Live Extended Thinking, in the API today
If you have been stitching a cascaded stack out of a recogniser, a model and a voice, this collapses three hops into one, with Live API integrations through LiveKit, Pipecat and Vercel. Enterprise access is still private preview.
Tools & models
llama.cpp v0.4.1 · a signature changes, and GDN normalisation is fixed
Anyone compiling against libllama, and anyone serving Qwen, Kimi or GLM locally
Tools & models
Codex 0.154 puts GPT-6 Astra in the picker; Claude Code pins to 2.1.267; MAI-Code-1-Flash leaves Copilot today.
Codex users who want Astra: ≥0.154.0. Claude Code: pin ≥2.1.267. Expect MAI-Code-1-Flash gone from Copilot. Copilot Business/Enterprise admins: review agent permission policies.
Tools & models
Quiet window: zero watchlist releases to 06:00 UTC 7 Sep; hold Friday’s pins.
No new decision pin inside the window. If you track npm latest, 2.1.263 is the tag; stable remains 2.1.236. MAI-Code-1-Flash leaves Copilot 10 Sep.
Tools & models
Copilot · Gemini 3.8 Flash in; four models out 2 Oct
Copilot Pro through Enterprise pinning Gemini Flash, Kimi, or Opus 4.7
Tools & models
Claude Fable 5.1 · default Fable, Copilot GA
Long-running agent work on Claude Code, API, or Copilot
Tools & models
Claude Fable 5.1 is the default Fable model — in Claude Code, on the API, and in Copilot.
Pin Claude Code ≥2.1.258. Try Fable 5.1 for long agent runs; on Copilot, read the retention policy before enabling org-wide.
Tools & models
Cline 4.1.17 · Hub memory balloon + Fable default
Cline extension/desktop on long tasks or unpinned Anthropic models
Tools & models
Quiet Monday. No watchlist release that changes a pin.
No pin from today. Plan Cursor OpenAI traffic through 12 Nov; re-budget Claude Code weekly after 14 Sept.
Tools & models
llama.cpp 0.3.0 · dots3-note multimodal, DeepSeek 4 tensor-split, ggml 0.22
local llama.cpp / server / multimodal runners, especially DeepSeek 4 multi-GPU
Tools & models
GitHub Copilot · Gemini 3.7 Flash rolling out
Copilot orgs picking a cheap coding model this week
Tools & models
DeepSeek V4 Pro leaves preview as 0813; Grok 4.6 ships the same day for long-running agents in Cursor and Grok Build.
Pin Pro 0813 and run one trusted agent loop; budget for a price change with no date. Try Grok 4.6 on one long job this week under the 2× promo; keep Composer for cheap daily edits.
Tools & models
Copilot · MAI-Code-1.1-Flash in, 1-Flash out 10 Sept
Copilot teams on MAI-Code-1-Flash or picking a cheap coding model this month
Tools & models
Ollama 0.32.9 · Nemotron 3.5 Lightning
local-agent hosts already on Ollama who want a small-active MoE next to Glimmer
Tools & models
Meta ships Muse Glimmer: 30B, Apache 2.0, built to run local agents on one GPU.
weights and HF load path are real; treat consumer-GPU speed claims as vendor until you time a tool loop on your box. Pin a digest, run one agent scaffold you already trust, measure tokens/sec and tool-call success before you move anything off the cloud default.
Tools & models
Open weights inside a closed harness
A lower token rate is useful only if the agent can still reach its tools. An August Codex/DeepSeek integration shows where that comparison can go wrong.
4 min readTools & models
Kimi K3 in GitHub Copilot
anyone about to move a team's default model on the strength of a benchmark table
Tools & models
DeepSeek shipped an MIT-licensed model that speaks Codex natively, and it runs a task for three cents.
the verbosity objection resolves in its favour. It emits 210M tokens on the index run against a median of 100M, and still costs $0.03 per task where GPT-5.6 Sol costs $1.86. Nine index points, sixty times the money.