If I had to build a SaaS from scratch today, the hard part wouldn't be building it — it would be choosing what to build it with. The ecosystem moves so fast that any comparison from six months ago is already wrong. This is my map as of July 2026: what I would use, what I would avoid, and where the pricing traps are.
Fair warning: there is no perfect stack, only stacks that fit your team and your stage. What follows is biased toward what a small startup optimizes for: iteration speed, low maintenance, and predictable bills.
Coding agents: the most important decision
The tool you write code with defines your velocity more than any framework. The current landscape:
| Tool | Base price | What stands out in 2026 |
|---|---|---|
| Claude Code | $20–200/mo | Dynamic Workflows: orchestrates dozens of parallel subagents from one session |
| Cursor | $20–200/mo | Composer 2.5 (in-house model) included in Pro; ~$4B ARR as of June |
| GitHub Copilot | $10–100/mo | Agent mode GA on VS Code and JetBrains; watch out for the new credit billing |
| Devin Desktop | $20/mo + usage | The former Windsurf, now Cognition's agent-management hub |
| Codex (OpenAI) | Bundled with ChatGPT | Open-source CLI + cloud tasks; 5M+ weekly users |
| Antigravity (Google) | Bundled with Gemini plans | Replaced Gemini CLI in June; async background workflows |
My read: the category no longer competes on "better autocomplete" — it competes on orchestration: how many agents you can direct at once, and with how much confidence. Claude Code is my daily driver for exactly that reason; Cursor is still the best editor if you want a traditional IDE with an agent inside.
💡 Trap of the month: GitHub swapped Premium Requests for "AI Credits" in June, and teams are reporting bills 10x higher on agentic workflows. Whatever your tool, set spend alerts before you unleash parallel agents.
Models: July 2026 prices
If you're building product on top of LLMs, this is what a million tokens costs today (input/output):
| Model | Input | Output | Note |
|---|---|---|---|
| Claude Fable 5 | $10 | $50 | Anthropic's top of the line; 1M context |
| Claude Opus 4.8 | $5 | $25 | The workhorse for code |
| Claude Sonnet 5 | $2 | $10 | Intro pricing through August; then $3/$15 |
| GPT-5.5 | $5 | $30 | 1M context |
| GPT-5.6 (Sol/Terra/Luna) | $1–5 | $6–30 | Released literally today |
| Gemini 3.1 Pro | $2 | $12 | Up to 200K context; more above that |
| DeepSeek V4 / Qwen 3.5 | ~open source | — | If you can self-host, the cost math changes leagues |
The sensible strategy for a SaaS: a cheap model (Sonnet 5, Terra, Gemini) for 90% of calls, escalating to the top tier only when the case justifies it. A 30-line model router saves you thousands of dollars a month.
Web frameworks: less drama than it seems
- Next.js 16.2 remains the rational default for SaaS: Turbopack is stable and on by default, React Compiler is integrated, and the ecosystem of examples is unbeatable. Boring and correct.
- Astro — just acquired by Cloudflare in January — is still my pick for anything content-shaped (this blog runs on Astro). Still MIT, still open governance.
- TanStack Start hit stable v1.0 in March and is the serious alternative if you want extreme type-safety without Next's magic.
- SvelteKit and React Router 7 are mature and excellent; choosing them is more a matter of team taste than capability.
The uncomfortable truth: with agents writing most of the code, framework choice matters less than it did three years ago. Agents perform better on frameworks with a bigger public corpus — another point for Next and Astro.
Backend and infra: the stat of the year
The stat that best sums up 2026: Supabase reported that the majority of its new databases are now deployed by AI agents, not humans — database creation growing 600% year over year. Your infrastructure is no longer chosen only by your team; agents "choose" it too, and they go with what they know how to use.
- Postgres wins by a landslide: Supabase (just raised $500M at a $10.5B valuation) as a full backend, or Neon (now inside Databricks) if you only want the database.
- Vercel for deployment if you go with Next: Fluid Compute genuinely eliminated cold starts.
- Cloudflare Workers + D1 + R2 is the cost-effective alternative — and with Astro in the family, increasingly polished.
- Runtimes: Node 24 is still the enterprise default. Bun — acquired by Anthropic in December — completed its Rust rewrite in May and is my pick for new projects: speed plus built-in tooling.
Your product's AI layer
This is where I see the most over-engineering. What a startup actually needs:
// The minimum viable AI layer in 2026
// 1. SDK: Vercel AI SDK 6 (TypeScript), or the Claude Agent SDK
// if the product IS an agent. LangGraph 1.0 if you need
// complex state graphs (that's how Uber and Klarna use it).
// 2. Context: MCP. It's a Linux Foundation standard now,
// 97M monthly downloads. Don't invent your own protocol.
// 3. Vectors: pgvector up to ~10M vectors (~$30/mo).
// Don't pay for a dedicated vector DB before you have the problem.
// 4. Evals from day one: if you don't measure LLM quality,
// every deploy is a gamble.
Point 3 deserves emphasis: pgvector in your regular Postgres covers almost any startup with 8–25ms latencies. Notion cut its search costs ~60% by moving off a dedicated vector DB onto object-storage-backed search; you probably don't even need that yet.
Vibe coding: use it for what it is
Lovable ($400M ARR in February, reportedly in talks to raise at an eleven-figure valuation), Replit ($9B valuation in March), v0, Bolt. They are idea-validation machines: prompt to deployed prototype in an afternoon. What they are not — yet — is the foundation of a product that has to scale with serious security and data requirements. The pattern I see working: prototype in Lovable or v0, validate with real users, and rebuild on your stack once there's signal.
My concrete recipe
If I started a B2B SaaS tomorrow, with no more context than "I want to validate fast and not burn money":
- Code: Claude Code as the primary agent + human review of everything touching money, permissions or data.
- App: Next.js 16 + TypeScript on Vercel. Content/marketing: Astro.
- Data: Supabase (Postgres + auth + storage + pgvector).
- AI: Vercel AI SDK 6, Sonnet 5 by default with escalation to Opus 4.8, MCP for integrations, evals with real cases from week one.
- Runtime: Bun locally and in CI; Node 24 wherever the platform demands it.
Is it the most powerful stack possible? No. It's the one that lets you iterate every single day with two and a half people — and in 2026 that's exactly the game: the edge is no longer in the infrastructure you set up, but in how fast you learn what to build.
