What AI coding assistants actually do
AI coding assistants do three jobs: autocomplete the line you're typing, answer questions about a codebase, and — increasingly — act as agents that edit files and run commands themselves. The productivity evidence is real: in the controlled GitHub Copilot study, developers completed a task 55.8% faster with the assistant than without [1]. The same study's nuance matters more than the headline: gains concentrate on routine implementation work, and review discipline is what keeps agent-made edits safe.
GitHub Copilot, decoded — including the "Copilot laptop" trap
Copilot's timeline: technical preview (June 2021), general availability at $10/month (June 2022), a free tier (December 2024), Pro+ at $39 (May 2025), a Max plan (May 12, 2026), and the switch to AI Credits usage-based billing on June 1, 2026 — fixed premium-request allotments were replaced by metered credits, and overages are billed [2][3]. Current pricing (September 2026): Free $0; Pro $10; Pro+ $39; Max $100; Business $19/user; Enterprise $39/user [4]. Microsoft reported Copilot passed 20 million users in August 2025 [5]. And the naming trap: searches for "review copilot laptop" or "lightest copilot laptop" are about Copilot+ PCs — Microsoft's branding for Windows laptops with NPU AI hardware — which have nothing to do with the GitHub Copilot coding tool. If a review mentions battery life and Snapdragon chips, it's a laptop; if it mentions code completion, it's the coding assistant.
Using Copilot in the terminal
"How to use copilot cli" and "how to use gh copilot cli" both point to GitHub's CLI integration: authenticate with gh auth login, install the Copilot CLI extension, and you get command suggestions and explanations in your shell — the "zsh ai copilot" workflow. The free tier includes limited CLI usage; heavier terminal use is what Pro+ is for.
The alternatives: Cursor, Claude Code, Windsurf, Amazon Q
- Cursor (Anysphere) — launched 2023, Cursor 2.0 (Aug 2025). Pricing: Hobby free, Pro $20, Ultra $200, Teams $40/user [6]. Business context: $500M annualized revenue by June 2025 [7], and reported $2B ARR with talks for a raise at a $50B valuation in 2026 [8]. It's the strongest dedicated AI editor — a fork of VS Code rebuilt around the model.
- Claude Code (Anthropic) — research preview Feb 24, 2025, GA May 22, 2025, web version Oct 20, 2025 [9]. Included with Claude Pro $20; Max at $100/$200 for 5x/20x usage [10][11]. Anthropic calls it its fastest-growing revenue product ever [12]. It is the most autonomous of the mainstream agents — it edits files and runs commands, so Anthropic's own docs advise reviewing changes [13].
- Windsurf — Codeium's editor, rebranded Nov 2024; Google hired the founders and Cognition acquired the product in July 2025 [14]. Pricing: Free, Pro $15, Teams $35/user [15].
- Amazon Q Developer — GA April 30, 2024 [16]. Free (50 agentic requests/month) or Pro $19/user with IP indemnity and AWS SSO [17]. The obvious choice if you already live inside AWS [18].
Run AI locally: Ollama, LM Studio and the open models
"How to host ai locally" starts with two free runners. Ollama (first released July 2023, MIT-licensed) is the terminal-native runner with a huge model library — it hit nearly 9 million users and raised $65M in July 2026 [19][20]. LM Studio is the graphical alternative (built on llama.cpp/MLX): Free for local models, with paid tiers Bionic+ $20 and Pro $100 for US-hosted cloud inference with zero data retention [21]. Then pick an open-weights model:
- Llama (Meta) — Llama 4 Scout/Maverick (April 5, 2025, Apache 2.0) [22]; Llama 4.5 Scout (July 16, 2026) with a 10M-token context window and 70B active parameters [23] — a GPU-class model, not a laptop one.
- Gemma (Google) — Gemma 3 (Mar 2025); Gemma 4 (April 2, 2026, Apache 2.0) [24][25]. The family passed 400M downloads [26].
- Phi (Microsoft) — Phi-4 (Dec 12, 2024, MIT license, 14B) benchmarks like models 5–10x larger on reasoning [27][28]. The best-known "best lightweight llm" candidate for reasoning-heavy local work.
- Qwen (Alibaba) — Qwen3 (April 29, 2025) [29]; Qwen3.8-Flash-Next (Aug 27, 2026) [30]. The larger builds need datacenter GPUs; smaller Qwen3 variants run on consumer hardware.
The privacy payoff is the point: everything runs offline, nothing leaves your machine. The cost is capability — a local 8B model will not match ChatGPT or Claude on hard problems.
VRAM & model sizes: the honest numbers
As an editorial rule of thumb for "best lightweight coding llm" shopping: at 4-bit quantization you need roughly 0.5–1GB of VRAM per billion parameters. That means 8GB VRAM → 7–8B models, 16GB → ~14B, and 24GB+ → 32B comfortably. Llama 4.5 Scout's 70B-active MoE and Qwen3.8's 125B main model are workstation-class. When a laptop won't do, use a quantized model one size down rather than expecting a large model to run acceptably on undersized hardware.
What "temperature" actually means
"Best temperature llm" searches usually mean the model-sampling setting, not a model tier: temperature scales randomness in output, typically 0–2. Low values (0–0.3) give consistent, deterministic answers — right for coding, math and factual Q&A; higher values (0.7–1.0+) give varied, creative phrasing. It does not make a model "smarter" — a weak model at low temperature is just consistently weak. Our default recommendation: 0.2–0.4 for code, 0.7–0.9 for drafts.
Comparison table (verified 2026)
| Tool | Free tier | Entry paid | Best for | Watch out |
|---|---|---|---|---|
| GitHub Copilot | Yes | Pro $10/mo | IDE integration, price | AI Credits metering since Jun 2026 |
| Cursor | Hobby | Pro $20/mo | Dedicated AI editor | Credit caps push power users to Ultra $200 |
| Claude Code | With Claude Free | Claude Pro $20/mo | Autonomous agents | Review its edits; usage limits |
| Windsurf | Yes | Pro $15/mo | Agentic editor value | Limited free Cascade quota |
| Amazon Q Developer | Yes, 50 req/mo | Pro $19/user/mo | AWS environments | AWS-centric SSO/management |
| Ollama | Yes (MIT) | — | Terminal model runner | Bound by your hardware |
| LM Studio | Yes | Bionic+ $20/mo | GUI local + cloud tiers | Free tier web-search limits |
Prices verified from vendor pages and reputable trackers, 2026. Sources: [4][6][10][11][15][17][21].
Frequently asked questions
GitHub Copilot vs Cursor vs Claude Code — which is best?+
Is GitHub Copilot free?+
How do I run AI locally on my PC?+
How much VRAM do I need to run Llama locally?+
Ollama vs LM Studio — which is better?+
Is a "Copilot laptop" the same as GitHub Copilot?+
What does LLM "temperature" mean?+
Sources
- Peng et al., "The Impact of AI on Developer Productivity" — arxiv.org/abs/2302.06590
- GitHub Blog, "GitHub Copilot is moving to usage-based billing" — github.blog
- GitHub Blog, "Copilot individual plans: flex allotments and a new Max plan" — github.blog
- CloudZero, "GitHub Copilot cost" — cloudzero.com
- App Developer Magazine, "GitHub Copilot surpasses 20M users" — appdevelopermagazine.com
- Cursor, pricing — cursor.com
- KrAsia, "Cursor's $500M annualized revenue", Jun 2025 — kr-asia.com
- The Next Web, "Cursor (Anysphere): $2B funding, $50B valuation" — thenextweb.com
- TechCrunch, "Anthropic brings Claude Code to the web", Oct 2025 — techcrunch.com
- Anthropic support, "What is the Max plan" — support.claude.com
- Claude pricing — claude.com
- Le Monde, "Anthropic valued at $380 billion after a new funding round", Feb 2026 — lemonde.fr
- Anthropic docs, Claude Code overview — docs.claude.com
- The New York Times, "Cognition AI acquires Windsurf", Jul 2025 — nytimes.com
- DEV, "Windsurf AI pricing 2026" — dev.to
- AWS, "Amazon Q Developer generally available" — aws.amazon.com
- AWS, Amazon Q Developer pricing — aws.amazon.com
- Superblocks, "Amazon Q Developer pricing" — superblocks.com
- Wikipedia, Ollama — en.wikipedia.org
- TechCrunch, "Ollama raises $65M, grows to nearly 9M users", Jul 2026 — techcrunch.com
- LM Studio, pricing — lmstudio.ai
- Meta AI, "Llama 4: multimodal intelligence" — ai.meta.com
- MintedBrain, "Llama 4.5 Scout" — mintedbrain.com
- Google AI for Developers, Gemma releases — ai.google.dev
- Android Developers Blog, "Gemma 4: new standard for local agentic intelligence", Apr 2026 — android-developers.googleblog.com
- Geo.tv, "Google launches Gemma 4 open models: 400M downloads" — geo.tv
- Hugging Face, microsoft/phi-4 — huggingface.co
- Abdin et al., "Phi-4 Technical Report" — arxiv.org/abs/2412.08905
- Alibaba Cloud, "Alibaba introduces Qwen3" — alibabacloud.com
- Alibaba Cloud blog, "Qwen3.8-Flash-Next" — alibabacloud.com