How to choose the best AI model for coding
A practical way to pick the best AI model for coding: judge first-attempt quality, speed and cost, match the model to the task, and test it on your own work before committing.
Choosing and comparing models for software work.
A practical way to pick the best AI model for coding: judge first-attempt quality, speed and cost, match the model to the task, and test it on your own work before committing.
AWS's Agent Toolkit for AWS now includes a distributed KV caching skill that lets coding agents size, set up and validate an ElastiCache cache for self-hosted LLM inference.
AWS announced the SageMaker HyperPod Global Inference Gateway, a single managed endpoint that routes inference requests across clusters and Regions with automatic failover.
OpenAI published Agents SDK v1.0.0-alpha.1 on 20 November, adding Codex as an opt-in model provider for the Python SDK, with wheels on GitHub Releases only.
simplexai-labs/linger drew 1,622 stars in ten days for an open-source personal agent that claims a 125B model on a 12 GB GPU, with nightly memory and proactive follow-ups.
Kimi Code CLI v2.7.0 adds a native Amazon Bedrock provider that authenticates through the standard AWS credential chain, so no separate Bedrock API key is needed.
A new MIT-licensed repo turns whiteboard tutoring into a JSON protocol any LLM can drive, then benchmarks six models across 288 runs and 64 blind-reviewed boards.
Exa optimized embedding inference on AWS Inferentia2 and Trainium2 with NKI kernels, reporting roughly 1.5x throughput and a 47% gain in performance and cost per million documents.
Cloudflare says Clairic has left preview and Clairic-flash has arrived on Workers AI, with both extraction models getting an open-weight release.
A TypeScript repo that gathered 301 stars in its first week turns viral short videos into a Viral DNA report, an editable VFML project and 10 costed variants.
opencode v2.3.2 sends the system prompt as a leading developer message on every Responses route so GPT-6.2 caches it, cutting uncached input per step from about 14,000 tokens to 3 on one Copilot model.
AWS added five data preparation capabilities to Amazon SageMaker AI, covering auto-labeling, synthetic data, quality checks, transformation and human review.
GitHub Copilot adds OpenAI's GPT-6.2 Astra and GPT-6.2 Sol for agentic coding and code review, available to Pro+, Max, Business and Enterprise users.
AWS adds Aionic Labs' open source time-series language models to SageMaker JumpStart, with a deployment path and fine-tuning via TimeNet.
GitHub expanded Copilot Fusion on November 17, 2026 with three new models, wider surface support, and a 71.2% Terminal-Bench 4.1 score for Fusion Max.
GitHub added MAI Code 2 and MAI Code 2 Flash to Copilot's model picker on November 17, 2026, with Flash reaching Free and Student plans and both billed by usage.
GitHub Copilot Evals is in public preview. It lets you create repeatable tests for skills, custom agents and MCP tools, then compare runs after a change.
The Vue package release carries [email protected], which validates nested UI message metadata in one pass, plus new gateway access policy errors and a 7 MiB Vertex download default.
OpenAI's GPT-6.2 Astra and Sol are now on Vercel's AI Gateway, with an ultrafast service tier and a CLI setup step for Codex and Cursor.
Gezhi Sheets, an MIT licensed AI spreadsheet assistant, reached 300 GitHub stars in eight days. It runs locally on Windows and connects to five Chinese model providers.
A day-old GitHub repo packages local LLM inference optimization as an agent skill, with benchmarking scripts, ceiling maths and before-and-after numbers on named hardware.
GitHub's November 16 Copilot roundup adds Claude Fable 5.2 and Gemini 4 Carbon for Pro+, Max, Business and Enterprise, plus GHES repositories and agent session tools in VS Code.
GitHub retired Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra from most Copilot experiences on November 16, 2026, with named replacements and a fallback exception for Opus 5.
browser-use 1.0 removes 136,866 lines and lets the agent write Python directly against Chrome DevTools Protocol, with new cloud defaults and a more modest benchmark read.
The huawei-csl project MACaQ shipped an Apache-2.0 quantization toolkit that writes GGUF files for llama.cpp and covers text, image, video and audio models.
The mcp-for-blender-pro add-on puts an AI chat panel and an MCP server inside Blender, with 21 scene tools, model choice and multi-agent workflows.
Gemini 4 Hydrogen costs $1.50 per million input tokens, a 25% discount on Carbon and Argon that becomes 75% when those two models double in price on 1 December.
Gemini Decision 1 joins Jev and OpenAI in the decision model category, handling text, images, audio, video and PDFs at 10 cents per million input tokens with free output.
Microsoft added GPT-5.6, GPT-6 and GPT-6.1 to Azure Content Understanding, and published accuracy and cost comparisons that put its cheapest option within a few points of the top.
Gemini 4 Hydrogen targets fast agentic coding at $1.50 per million input tokens, rolling out across AI Studio, the Gemini API, Search and Google Antigravity.
SGLang-Neuron is now publicly available in an AWS Deep Learning Container, bringing the SGLang serving workflow and PD disaggregation to AWS Trainium on native PyTorch.
An MIT-licensed Agent Skill collects your Claude Code transcripts, cleans them, trains a LoRA on Tinker and exports a GGUF for llama.cpp. It had 323 stars in a week.
An OpenAI post says Greptile uses GPT-6 Astra to test user workflows, Daybreak to strengthen security and Codex to reduce rework. Details are thin.
OpenAI Codex 0.168.0 lets you browse model providers, pick a default for new chats and open a side conversation in the terminal, alongside audio attachments and GIF output behind flags.
Anthropic's TypeScript SDK 1.7.0 adds two model IDs, a new organization webhook event and parent organization filters for usage and cost reports.
Ollama v0.40.14 changes how OpenAI-compatible endpoints pick temperature and top_p, fixes cloud web search decoding, and applies model output limits for Claude Code.
Google's walkthrough of the Agents CLI eval loop takes an agent from a 4.0 baseline to a 2.75 regression, then shows how per-case scores explain a model swap.
Vercel's AI Gateway now lists Claude Fable 5.2, with model fallbacks for classifier declines and a data retention condition that rules out ZDR setups.
sand3d, a Python library for writing 3D objects as short programs over Blender's bpy, drew 1,969 GitHub stars in 11 days, with one-call exports to GLB, URDF, USD and IFC.
Google released Gemini 4 Carbon to paying developers at $2 per million input tokens, claiming first place on three agentic benchmarks and previewing multi-agent /swarm in Antigravity.
A Google benchmark finds a cheap classifier can route coding tasks to cheaper Gemini models and thinking levels, while token caps and regex rules fail on unseen work.
Ollama v0.40.16 ships Ripple Preview, a 65B dense model from Ricursive Intelligence built for agentic coding, plus a 27 GiB memory saving on Apple silicon and API fixes.
The gputaskman GPU task manager links per-process VRAM and engine activity to loaded models and KV-cache sizes. It reached 3,125 GitHub stars 11 days after creation.
Vercel added Google's Gemini Decision 1 to AI Gateway. The model returns booleans, choices and scores for routing, policy checks and ranking, without generating output tokens.
browser-use 1.0.4 batches old screenshot pruning so provider prompt caches survive, dropping model cost per task on BU Bench v2.1 from $0.434 to $0.208.
Google AI Edge Gallery now runs speech recognition and text-to-speech fully on device, with an Android ML Kit path and a cross-platform MediaPipe and LiteRT path.
AWS's guide shows three ways to route OpenAI Codex model requests through Amazon Bedrock, from a native IAM pilot to a managed AgentCore Gateway or an LLM gateway.
cinecode, an MIT-licensed Python repo with 2,446 stars after 11 days, has an AI coding agent write the script, build every frame in code and export an MP4.