Starting Spec Driven Development (SDD)
This guide helps you choose and try AI coding tools based on your budget while maintaining the same Spec-Driven Development (SDD) methodology across all tiers.
Choose the option that fits your budget; the disciplined loop remains the same: Spec → Plan → Prompt → Test/Evaluate → Refactor → Record → Review.
- $0/mo — Qwen Code (CLI) + Gemini CLI: Start free with two terminal tools. Qwen Code offers strong repo exploration, git automation, and optional vision. Gemini CLI provides large‑context prompting, great for long prompts and scripts.
- $20/mo — Cursor Pro: An AI‑native IDE with Agent Mode, multi‑file awareness, parallel agents, and fast tab completion.
- $40+/mo — GPT‑5 Codex + Cursor Pro: Two Autonoumous AI Agents in Cursor AI Native IDE for parallel, repo‑wide work.
- Add‑On (Any budget) — Keep Gemini CLI + Qwen Code alongside anything above: The free tier of both tools can be used in combination with Cursor Pro or GPT‑5 Codex for extra flexibility at no additional cost.
Start free, then upgrade as your needs grow.
Tool Comparison
Below is a comparison of the four SDD tool options, tailored to different budgets, with consistent formatting and focus on their features, strengths, and use cases.
| Aspect | Qwen Code (CLI) | Gemini CLI | Cursor Pro (AI-First IDE) | GPT-5 Codex (Cloud Agent) |
|---|---|---|---|---|
| Core Design | Terminal-based coding agent with strong repo analysis and optional vision support for multimodal prompts. | Terminal-native assistant with large context windows, ideal for long prompts and scripts. | AI-native IDE (VS Code fork) with seamless UI for inline edits, autocomplete, and multi-file awareness. | Cloud-based agent for agentic coding, focusing on parallel task execution and repo-wide automation. |
| User Interaction | CLI prompts for quick tasks, codebase queries, and git automation; runs locally with OAuth/API. | CLI prompts with simple Google auth; optimized for scripting and long, single-flow sessions. | Interactive, local IDE with inline suggestions, previews, and chat integrated with files/projects. | Conversational and autonomous; describe tasks via ChatGPT or CLI, executes with minimal supervision. |
| Strengths | Free with 2K req/day, excels in codebase exploration, vision support, and git utilities. | Free, large context windows, fast setup, ideal for scripting and educational workflows. | Fast daily coding, multi-file edits, Agent Mode, parallel agents, and exclusive tab completion. | Handles complex, parallel, repo-wide tasks; generates PRs with tests/docs; strong automation. |
| Use Cases | Learning, repo audits, architecture reviews, quick refactors via CLI. | Long prompts, scripting, terminal-based workflows, educational tasks. | Individual/team coding, prototyping, iterative development with rapid tab completion. | Large-scale automation, building from scratch, cross-repo tasks with minimal hands-on work. |
| Pricing | Free ($0/mo). | Free ($0/mo). | Pro plan (~$20/mo), limited free tier. | Plus/Enterprise or API (~$40+/mo). |
| Limitations | CLI-only UX; complex edits need editor integration; quotas apply. | CLI-only; fewer repo-wide automation features than Codex/Cursor. | Review needed for refactors; fewer autonomous long-running tasks than Codex. | Slower for simple edits; cloud latency; less interactive in-editor without extensions. |
Getting Started
- Free Tier ($0/mo): Use Qwen Code CLI for codebase exploration and git automation, paired with Gemini CLI for long-context scripting. Ideal for beginners or lightweight projects.
- Pro Tier ($20/mo): Upgrade to Cursor Pro for an AI-native IDE with fast tab completion, Agent Mode, and multi-file editing, perfect for solo developers or small teams.
- Premium Tier ($40+/mo): Combine GPT-5 Codex with Cursor Pro for autonomous, repo-wide automation and PR generation, suited for complex projects or teams needing scalability.
- Start with the free tier and upgrade as your workflow evolves, keeping the SDD method consistent.
Latest SWE-bench Verified
In short: SWE-bench Verified is the most commonly cited, cleaner benchmark for end-to-end, agentic code fixing on real repositories, scored by passing the project’s tests after your patch.
Here’s an apples-to-apples snapshot of SWE-bench Verified results for the four we care about. We list the most recent, citable scores, and call out when the number depends on the evaluation harness (single/multi-attempt, custom agent, or 500-turn runs).
| Model / Tool | SWE-bench Verified score | Notes (harness / caveats) |
|---|---|---|
| Anthropic – Claude Opus 4.1 | 74.5% | Anthropic’s own announcement for Opus 4.1; update over Opus 4.0’s 72.5%. |
| OpenAI – GPT-5 Codex | 74.5–74.9% | 74.9% stated in GPT-5 launch (initially on 477/500 tasks); follow-up post clarifies reporting; tech press widely quotes 74.5% for Codex variant. Treat as ~74.5–74.9 depending on harness. |
| Google – Gemini 2.5 Pro | 63.8% | Google/DeepMind blog cites 63.8% using a custom agent; product page also shows 31.6% (single attempt) / 42.6% (multi-attempt) without the custom agent. |
| Qwen – Qwen3Max-Coder / Qwen3-Max | ≈69.6% | Reported in multiple roundups for 500-turn runs; Qwen’s blog claims SOTA among open-source but doesn’t pin a single number. Use ~69.6% as the current ballpark. |
| Cursor Pro (IDE) | — | Not a model—no SWE-bench score; performance depends on which model/agent you run inside Cursor. |
Practical performance (what you’ll feel day-to-day)
-
GPT-5 Codex – shows the best agentic coding results on public leaderboards; Best at large, multi-file changes and autonomous bug-fix PRs; strongest SWE-bench Verified 74.5%, showing among public numbers right now. Expect better planning + code review abilities.
-
Gemini CLI (2.5 Pro) – Very capable agentic runs with generous free limits; excels in terminal-centric workflows, huge context, and easy pairing with Google’s Code Assist/MCP. SWE-bench Verified ~63.8%
-
Qwen Code (Qwen3Max-Coder) – Excellent free/open models for local/offline and repo-friendly code gen & repair; top open-source results on classic code benchmarks; for agentic repo-wide tasks you’ll likely pair it with a framework and careful tooling.
-
Cursor Pro – Big productivity win (indexing, apply-diffs, background agents). Quality maps to the model you choose (GPT-5 Codex, Gemini, Claude, etc.). Pricing is transparent and reasonable for daily use.
Free-Tier Ranking
Here’s a free-tier ranking (most to least generous), across the five you’re tracking:
| Rank | Tool / Model | What you get for free | |
|---|---|---|---|
| 1 | Qwen Code (open-weights) | Run locally = no API fees (you only pay your own compute). If you use Alibaba’s hosted API, new users get free token quotas (e.g., 1M tokens on select Qwen tiers). 2,000 requests/day at no charge for Code | |
| 2 | Gemini CLI (Gemini 2.5 Pro) | Official blog: up to 60 requests/min and 1,000 requests/day at no charge for CLI usage. | |
| 3 | Claude Code | Claude Free tier includes coding features, but with daily usage limits (no precise numbers published; resets daily). | |
| 4 | Cursor (Hobby) | Free IDE plan with limited Agent requests and limited tab completions. | |
| 5 | GPT-5 Codex | No meaningful free tier: access is via ChatGPT paid plans or paid API; free ChatGPT doesn’t include GPT-5/Codex. |
Notes
- If you can run models locally, Qwen open-weights is the most “free” (no per-token bill). Hosted Qwen still starts with a promotional free quota for new Alibaba Cloud accounts.
- Gemini CLI’s free allowance is the most concrete/quantified among hosted options (1,000 req/day).
- Claude Free clearly exists, but Anthropic doesn’t publish strict daily numbers; multiple roundups confirm the tier and that heavy users hit caps.
- Cursor is an IDE: its Hobby plan is free, but actual LLM usage limits are tight until you subscribe—and model tokens are billed by the provider when you connect paid APIs.
Pricing Ranking
Here’s a price-ranked comparison (cheapest → most expensive) using current list prices for coding-capable models. We’ve split API token pricing (per 1M tokens) from subscription seats like Cursor.
API pricing (per 1M tokens)
| Rank | Model / Tool | Input | Output | Notes |
|---|---|---|---|---|
| 1 | Qwen3-Coder-Flash (Alibaba Cloud) | $0.144 | $0.574 | Tiered by prompt size (0–32k bucket shown). |
| 2 | Qwen3-Coder-Plus (Alibaba Cloud) | $1.00 | $5.00 | Also tiered by prompt size (0–32k bucket). |
| 3 (tie) | OpenAI GPT-5 / GPT-5-Codex | $1.25 | $10.00 | OpenAI states Codex is same price as GPT-5. |
| 3 (tie) | Google Gemini 2.5 Pro | $1.25 | $10.00 | Higher tier for >200k-token prompts. |
| 5 | Anthropic Claude Sonnet 4 | $3.00 | $15.00 | Standard (short-context) rates. |
| 6 | Anthropic Claude Opus 4.1 | $15.00 | $75.00 | Flagship Claude tier. |
Note: Qwen open-weights can be run locally (no API fee; you still pay infra). The Alibaba Cloud prices above apply when you use their hosted API.
Subscription / seat pricing (not per token)
| Product | Price | What it covers |
|---|---|---|
| Cursor Pro (IDE) | $20 / user / month (individual). Teams from $40 / user / month. | Editor features; LLM usage billed at provider’s API rates you choose. |