DEV Community: Rai Ansar

A 200 response and a broken page: what screenshot monitoring taught me

Rai Ansar — Sat, 17 Oct 2026 23:27:52 +0000

An uptime check answers one question: did the server respond? Visitors ask a different one: can I use this page? After a few years of looking after client sites, the gap between those two answers was the thing I hit most often. A CSS deploy left a blank white screen. A WordPress plugin update broke checkout. An SSL certificate expired while the uptime dashboard still showed green.

I am Rai Ansar, a developer and DevOps engineer in Islamabad. I have worked on websites and servers since May 2019, and I built Visual Sentinel to close that gap. Here is what the visual part taught me.

The diff is the easy part

The core of a visual monitor fits in a few lines. This is a minimal sketch with Playwright, not our production code:

import { chromium } from 'playwright';

const browser = await chromium.launch();
const page = await browser.newPage({ viewport: { width: 1366, height: 900 } });
await page.goto('https://example.com', { waitUntil: 'networkidle' });
await page.screenshot({
  path: 'current.png',
  fullPage: true,
  mask: [page.locator('.ad-slot'), page.locator('time')],
});
await browser.close();

Compare current.png with a stored baseline pixel by pixel and you have a change detector. Left like this it will also alert on pages that change on every load, which leads to the lessons below.

Lesson 1: the baseline is the last confirmed good capture

If you compare against the previous screenshot, a broken page becomes the new normal after one run. Compare each capture against the most recent confirmed baseline instead. Keeping the screenshot history then tells you when the break was introduced.

Lesson 2: the threshold decides whether anyone trusts the alert

Ads, timestamps and rotating banners change on every load. Each monitor has a mismatch threshold, 5% by default, calculated over the whole rendered page. A tightly designed landing page or a checkout flow can go lower. A news homepage needs a higher number. For elements that always move, exclude them by CSS selector rather than raising the threshold for the whole page, because a high threshold hides real breakage.

Lesson 3: a change is evidence, not a verdict

On bdifferent.ie, a real customer site (recorded July 2026, shared with permission), 62.23% of the homepage changed while the uptime check still reported that the page loads. That is an easy case. Many diffs are a new hero image or a seasonal banner. The alert should carry a before and after capture so a person can decide in a few seconds. We treat a visual change as evidence to review and never as an automatic outage.

Lesson 4: pixels and text answer different questions

A screenshot tells you the page looks different. It does not tell you the price changed from one number to another. For that we run a separate content check: it extracts text from CSS selectors you define in the fetched source HTML and alerts when it changes. It does not run page JavaScript, so a price rendered client-side needs the visual check instead. Knowing which check covers which case saved me from promising more than either one does.

Lesson 5: capture frequency is a cost decision

Rendering a full page in Chromium costs more than a ping. On our Solo plan visual captures run every 30 minutes, and on Starter and above every 15 minutes, while uptime checks are much cheaper and run more often. A deploy regression is visible on the next capture, so a short interval matters most on sites that ship often.

Lesson 6: group the alerts

When three sites on the same server go down together, three separate 3 a.m. alerts help nobody. Visual Sentinel links sites to a server and sends one grouped incident when they fail together, with the likely shared cause.

What I would build first if I started again

Start with the screenshot, the baseline and the threshold. Add the content check once someone asks about a specific price or sentence. Add server grouping when you run more than one site per machine.

If you want to see the visual side on a live page, Visual Sentinel's visual monitoring shows a captured diff with the last good version beside the current one. Your first check is free when you add a site.

ChatGPT vs Claude vs Gemini (March 2026): The Definitive AI Comparison

Rai Ansar — Tue, 10 Mar 2026 03:39:48 +0000

Three platforms. Three radically different philosophies. And in March 2026, the gap between ChatGPT, Claude, and Gemini has never been more interesting — or more confusing for anyone trying to pick one.

OpenAI just shipped GPT-5.4 with native computer use and a 1M-token context window. Anthropic's Claude Opus 4.6 sits at #1 on the LMSYS Chatbot Arena. Google's Gemini 3.1 Pro quietly posted a 94.3% on GPQA Diamond, the highest score any model has achieved on PhD-level science questions. Meanwhile, the real battleground has shifted to coding agents: Claude Code, GPT Codex, and Gemini CLI are fighting for every developer's terminal.

I've spent the past two weeks stress-testing all three across coding projects, research tasks, creative writing, and daily workflows. Here's what actually matters.

Last Updated: March 2026

Quick Verdict: Best AI for Each Use Case

Use Case	Winner	Why
Coding & Development	Claude (Opus 4.6 + Claude Code)	#1 on SWE-bench (80.8%), Claude Code CLI dominates
Research & Analysis	Gemini 3.1 Pro	1M native context, 94.3% GPQA Diamond
Creative Writing	Claude Opus 4.6	Most natural prose, best voice consistency
Agentic Workflows	ChatGPT (GPT-5.4)	Native computer use, multi-step automation
Best Value	Gemini	Free tier with Flash, $19.99/mo for Pro
Enterprise/Teams	ChatGPT	Most mature ecosystem, Codex for async work

The Latest Models: March 2026

ChatGPT: GPT-5.4 Changes the Game

GPT-5.4, released March 5, 2026, brings native computer use — it can interpret screenshots, operate browsers, and issue keyboard/mouse commands. Key upgrades:

1M token context window (API) — up from 272K
Computer use built-in — first mainline model with native screen interaction
GPT-5.3-Codex capabilities merged — industry-leading code gen baked in
GDPval score of 83% — matches or exceeds professionals across 44 occupations

Claude: Opus 4.6 Takes the Crown

Claude Opus 4.6 holds #1 on LMSYS Chatbot Arena with 1504 Elo — real users preferring Claude over every other model in blind tests.

80.8% on SWE-bench Verified — top-tier for real-world software engineering
200K context window (1M beta) with 128K max output tokens
Adaptive thinking — dynamically decides reasoning depth
Compaction — automatic context summarization for infinite conversations

The sleeper hit is Claude Sonnet 4.6 at 79.6% SWE-bench — one-fifth the cost of Opus and preferred over the previous Opus 4.5 in 59% of comparisons.

Gemini: 3.1 Pro Is a Quiet Beast

94.3% on GPQA Diamond — highest PhD-level science score ever
80.6% on SWE-bench Verified — tied with Claude Opus 4.6
77.1% on ARC-AGI-2 — more than double Gemini 3 Pro's 31.1%
Native 1M token context — no beta flag, no waitlist
Multimodal — text, images, 8.4 hrs audio, 1 hr video, 900-page PDFs

Head-to-Head Comparison

Feature	ChatGPT (GPT-5.4)	Claude (Opus 4.6)	Gemini (3.1 Pro)
Context Window	1M (API) / 272K (Chat)	200K (1M beta)	1M native
Max Output	~32K tokens	128K tokens	65K tokens
LMSYS Rank	Top 10	#1 (1504 Elo)	#2 (1500 Elo)
SWE-bench	77.2%	80.8%	80.6%
GPQA Diamond	92.8%	91.3%	94.3%
ARC-AGI-2	73.3%	75.2%	77.1%
Image Gen	DALL-E 4	None	Nano Banana 2
Computer Use	Native	Via API	Limited
Coding Agent	GPT Codex	Claude Code CLI	Gemini CLI

Coding Showdown: Claude Code vs GPT Codex vs Gemini CLI

The real competition is in the terminal.

Claude Code: The Developer's First Choice

Claude Code hit $2.5 billion ARR — over half of Anthropic's enterprise revenue (more on AI coding tools).

It runs in your terminal, reads your entire project, writes code, runs tests, handles git, and debugs failures:

Parallel subagents — up to 7 simultaneous operations
MCP integration — Google Drive, Jira, Slack, custom tooling
Full terminal access — builds, tests, git, any CLI operation
VS Code and JetBrains extensions

GPT Codex: Async Powerhouse

Codex is a senior engineer you delegate to. It works autonomously in cloud sandboxes:

Runs 1-30 minutes on complex tasks with real-time progress
Cloud sandboxes with test harnesses, linters, type checkers
Interactive mode with GPT-5.4 — steer mid-task
Parallel worktrees — multiple agents on different project parts

The Power Move: Use Both Together

The workflow gaining traction:

Claude Code generates — faster real-time coding, deep local context
GPT Codex reviews — autonomous code review in cloud sandbox
Claude Code iterates — rapid fixes from Codex feedback

Teams report 30-40% more issues caught than either tool alone.

Gemini CLI: Present but Not Ready

Free tier with 1,000 requests/day is generous, but:

Sequential execution only — no parallel tasks
Frequent 429 rate limit errors
Less refined agentic behavior

For professional work, Claude Code and GPT Codex are in a different league.

Pricing

Plan	ChatGPT	Claude	Gemini
Free	GPT-4o	Sonnet 4.6	Flash, 1K req/day
Standard	$20/mo	$20/mo	$19.99/mo
Power	$200/mo	$100-200/mo	$249.99/mo

API (per million tokens)

Model	Input	Output
GPT-5.4	~$2.50	~$10.00
Claude Opus 4.6	$5.00	$25.00
Claude Sonnet 4.6	$3.00	$15.00
Gemini 3.1 Pro	$2.00	$12.00

Which Should You Choose?

ChatGPT → agentic automation, async coding delegation, enterprise teams

Claude → daily coding (Claude Code is unmatched), best writing quality, complex nuanced tasks

Gemini → massive documents (1M context), best free tier, PhD-level reasoning

My Daily Setup

Claude Code (Pro $20/mo) — primary coding tool
ChatGPT Pro ($200/mo) — Codex for async delegation
Gemini AI Pro ($19.99/mo) — research, Google integration

Pick just one? Claude Pro at $20/mo. Best value per dollar.

FAQ

Is ChatGPT still the best AI in 2026?
Most popular, but Claude holds #1 on LMSYS Arena and Gemini leads reasoning benchmarks.

Is Claude better than ChatGPT for coding?
Yes — 80.8% vs 77.2% SWE-bench, and Claude Code CLI has $2.5B ARR.

Can I use Claude Code and GPT Codex together?
Absolutely. Implementation + review. 30-40% more issues caught.

Which has the largest context window?
GPT-5.4 and Gemini: 1M tokens. Gemini's is natively available everywhere.

Originally published on AIToolRanked. More comparisons: ElevenLabs review | Best AI for coding | Grok vs ChatGPT