Insights
Notes on building software.
Guides to pricing an AI feature, hiring a development team, and scoping a product, followed by technical notes on models we evaluate.
Buyer guides
AI Prototype to Production: Launch Checklist
Move your AI prototype to production with practical checks for answer quality, permissions, agent actions, costs, and recovery before launch.
Read it →What It Costs to Run an AI Feature in Production
A worksheet for AI feature cost: published token rates, cache hits, retries, and reviewer time. Send it when you request a project scope.
Read it →How to Hire an AI Development Agency Without Getting Burned
Evaluate an AI development agency with a test-set checklist, sample failure cases and questions about data access, costs and handover.
Read it →What Drives MVP Development Cost? A Scoping Guide
How accounts, payments, AI and integrations affect MVP development cost. Decide what to build first and what a written quote should cover.
Read it →How to Choose a White-Label Development Partner
For agencies handing client work to a dev partner: what to check before you trust someone, the red flags, and fixed-scope vs dedicated developers.
Read it →LangChain vs LangGraph: Which Does Your AI Project Need?
Compare LangChain’s agent interface with direct LangGraph orchestration. When to use each, when a retrieval pipeline is enough, and a 12-step workflow example.
Read it →Model notes
GPT-6.1 Sol: $2/$10 Pricing, Benchmarks, and API Access
GPT-6.1 Sol launched 29 September 2026 at $2/$10. Cached input is $0.10. It scores 52, one point under Astra, at about a quarter of the task cost.
Read it →GPT-6 Sol and Luna: $2/$10 Pricing, Benchmarks, and API Access
OpenAI shipped GPT-6 Sol and Luna on 22 September 2026 at $2/$10 and $0.10/$0.50. Cost per task fell. Peak scores on DeepSWE and OSWorld did not.
Read it →Claude Opus 5.5: $4/$20 Pricing, Benchmarks, and API Changes
Claude Opus 5.5 launched 22 September 2026 at $4/$20. Coding scores rose. Default effort is now medium, and forced tool use returns an error.
Read it →Grok 4.7: Same $2/$6 Price, and Where the Scores Moved
SpaceXAI shipped Grok 4.7 on 21 September 2026 at $2/$6, the same card as Grok 4.6. Intelligence Index 46, up two points. Long tasks improved. Terminal-Bench depends on the harness.
Read it →GPT-6 Astra: Pricing, Benchmarks vs Sol, and Who Can Use It
OpenAI shipped GPT-6 Astra on 3 September 2026. Independent benches tie it with GPT-5.6 Sol at 61. Computer use is the jump. Specs, $10/$50 pricing, Daybreak access, and where we'd use it.
Read it →Ox Alpha Was GLM-5.3-Flash: Specs, Price, and How to Use It
Ox Alpha was the free stealth model on OpenRouter. Z.ai confirmed it as GLM-5.3-Flash on 26 August 2026. Specs, benches, current pricing, and the model IDs that still work.
Read it →Grok 4.6: Tied with GPT-5.6 Sol, Same $2/$6 as 4.5
SpaceXAI shipped Grok 4.6 on 12 August 2026. It scores 61 on Artificial Analysis, tying GPT-5.6 Sol, at $2/$6 per million tokens. The benchmarks, the cache-price catch, and where we'd actually use it.
Read it →DeepSeek V4 Flash 0731: Cheap, Fast, and Hard to Steer
DeepSeek's V4 Flash 0731 costs $0.14 per million input tokens and beats V4 Pro Preview on every agent benchmark they published. The numbers, the instruction-following problem nobody benchmarks, and what it takes to self-host.
Read it →Claude Opus 5 vs Grok 4.5 vs GPT-5.6 Sol: What Changed for Builders
Claude Opus 5 landed on July 24 at $5/$25 per million tokens with a 1M context. How it compares to Grok 4.5 and GPT-5.6 Sol on price, benchmarks and independent testing — and what you'll have to change in your prompts.
Read it →Kimi K3: What an Open-Weight Frontier Model Means for Builders
Moonshot AI's Kimi K3 lands this week — a 2.8-trillion-parameter open model with a 1M-token context and native vision. An honest read on what it changes if you're shipping a product, and where the caveats are.
Read it →