/// AI TOOLS EDGE · Saturday, July 18, 2026 ET
GPT-5 wins 54 percent of 300 writing prompts via Claude 3.7 LLM-as-judge eval, but trails Claude 4.5 by 18 points on factual accuracy.
Notion AI cheapest at 0.012 USD per 1k output tokens. Jasper outputs longer-form usable copy 2.3x more often per prompt. Copy AI loses on both axes.
Cursor wins p95 latency at 187ms vs Copilot 312ms, but Copilot completes correctly on first suggestion in 71 percent of cases vs Cursor 64 percent.
Live as of 6:52 am PT · 6 macro tech-finance indicators tracked daily
Data sources : SEC EDGAR, Pitchbook Q1 2026, Carta Equity Summary, AWS+Azure+GCP earnings filings. Updated each quarterly filing release. Not investment advice.
Long-form deep dives, this week · cap table mechanics, SaaS metrics, IPO forensic
Source : AI Tools Edge benchmark suite Q2 2026 n=250
Three personal knowledge AI tools tested with 250 retrieval queries. Mem wins recall at 88 percent. Notion AI synthesis quality 2.1x higher LLM-as-judge.
Source : AI Tools Edge benchmark Q2 2026 n=150
Three AI research tools tested with 150 academic queries requiring citations. Perplexity Pro wins 84 percent citation accuracy. Claude Citations 78 percent. ChatGPT Search 62 percent.
Source : AI Tools Edge code benchmark Q2 2026 n=200 prompts
5 code completion AI tools tested across 8 programming languages with 200 real codebase prompts. Cursor wins JS/TS, Copilot wins Python, Cody wins Rust.
Recent SEC EDGAR + Pitchbook + Crunchbase + Sacra publications cited in AI Tools Edge analyses. Updated weekly.
All citations link to original primary sources. AI Tools Edge does not aggregate secondary commentary.
Primary research papers and model cards tell stories AI marketing decks miss. The ArXiv preprint says one thing, the launch announcement another - and the gap is where reproducible benchmark methodology lives. Every Thursday I publish what the three standardized prompt suites (writing 100, code 200, research 150) actually show, not what the press release needs them to mean.
Latest 12 deep dives · sorted by publication date
GPT-5 wins 54 percent of 300 writing prompts via Claude 3.7 LLM-as-judge, but trails Claude 4.5 by 18 points on factual accuracy.
Notion AI cheapest at 0.012 USD per 1k output tokens. Jasper outputs longer-form usable copy 2.3x more often. Copy AI loses on both axes.
Cursor wins p95 latency at 187ms vs Copilot 312ms. Copilot completes correctly on first suggestion in 71 percent of cases vs Cursor 64 percent.
Mem wins recall at 88 percent. Notion AI synthesis quality scores 2.1x higher LLM-as-judge.
Perplexity Pro wins citation accuracy 84 percent. Claude Citations second at 78 percent. ChatGPT Search trails at 62 percent on 150 academic queries.
Cursor wins JS/TS, Copilot wins Python, Cody wins Rust. No universal winner across 200 real codebase prompts and 8 languages.
Claude 4.5 holds 91 percent factual recall at 200k tokens. GPT-5 drops to 73 percent. Gemini 2.5 Pro middle at 82 percent.
GitHub Copilot Business 19 USD per seat hides true TCO. Cursor Pro 20 USD per seat. Cody Enterprise 9-19 USD depending on volume.
Fathom wins transcription accuracy at 94 percent WER on 50 hours of meeting audio. Otter at 91 percent. Fireflies at 88 percent.
Flux 2 Pro wins portraits at 78 percent rater preference. Recraft V3 wins brand-style consistency. Ideogram 3.0 wins text-in-image accuracy at 95 percent.
Llama 4 70B matches GPT-4 mini on MMLU+HumanEval+GPQA composite. Mistral Large 2 leads code. Qwen 3 leads multilingual.
AI Tools Edge publishes three standardized prompt suites by category. Writing 100 prompts. Code 200 prompts. Research 150 prompts.
Cited by · Outlets we read and reference
Coverage targets - placement pending. p273 placeholder S103 - real coverage M9+ via MIT Tech Review + Stratechery + Hacker News submissions.
How AI Tools Edge sources, writes, and discloses
Every AI tool category tested with 3 fixed prompt suites - writing (100), code (200), research (150). Public, versioned, reproducible.
Quality scored by Claude 3.7 LLM-as-judge with disclosed rubric. Token cost and latency measured production-side concurrent.
Every claim links to ArXiv preprint, model card, official API docs, or first-party benchmark. No press-release recycling.
Refuse 10x productivity rhetoric. Refuse Top 10 AI Tools listicles. Refuse affiliate hype masquerading as review. Headlines are analytical.
Notion AI + Jasper + Copy AI affiliate links disclosed FTC-style at top of money pages + inline at point of mention. No hidden referral codes.
Every review closes with scoring 0-100 transparent verdict box showing token cost / latency / quality / hallucination rate / context window axes.