Blog

Flux Schnell renders 10,000 images for $9. GPT-image-1 at high quality bills $1,200 for the same volume. Eleven image APIs lined up at identical volume — the spread is 133x, wider than almost anything in text-model pricing.

R2 storage is $0.015/GB vs S3's $0.023 — a small gap. The real gap is egress: S3 charges $0.09/GB out, R2 charges $0. We price one 500GB workload at three traffic levels and the bill splits from a rounding error to a 99% difference.

OpenAI's tts-1 is $15 per 1M characters. ElevenLabs' Creator plan works out to about $182 per 1M characters — a 12x gap for turning text into audio. Here's the real math and what the premium buys.

Amazon SES sends an email for $0.10 per 1,000. SendGrid and Resend work out to ~$0.40 per 1,000 on their entry plans — a 4x premium. Mailgun and Postmark are 15x SES. Here's the real per-email math and what the premium actually buys.

Nova Micro, Command R7B and Ministral 3B all undercut $0.05/1M input tokens. Claude Haiku 4.5 is $1.00. Run the same 10M-call classification job through nine "budget" models and the bill ranges from $59.50 to $1,750 — a 29x spread for an identical task.

o1 costs 27x DeepSeek R1 on the same reasoning workload. Swap in o3 instead — same OpenAI reasoning tier — and the gap drops to 3.7x. The "reasoning tax" collapsed unevenly across OpenAI's own lineup before open weights even entered the picture.

The same 1024×1024 image costs roughly 765 tokens on GPT-4o and roughly 1,400 on Claude — nearly double, for identical pixels, because the two providers use completely different tiling formulas. Detail mode and resolution matter more than which model you pick.

OpenAI, Anthropic and Gemini all cut batch pricing roughly in half for the identical model and output. The catch: the 24-hour window is a target, not an SLA, which makes the real decision rule not "how big is the job" but "is anything waiting on the response right now."

DeepSeek V4 and Llama 4 Maverick are the exact same weights on Groq, Together, Fireworks, DeepInfra, Novita and OpenRouter — yet the bill swings 31% between cheapest and priciest host, in the same order, on every model we checked.

Connect 25 MCP tools and an agent pays 25,000 tokens of schema on every turn, called or not. Uncached that's $2,250/month at 50 sessions a day; caching claws back 85% — until anything about the tool list changes.

Google's $200 monthly credit is one shared pool across every Maps SKU; Mapbox gives each product its own free allowance. Sized on a real store-locator feature — 150,000 map loads, 20,000 geocoding lookups — Google lands at $950/month, Mapbox at $500. Where the 47% gap actually comes from.

"Just use the smaller model" is not one decision, it's six different ones. On a 240M-token month, Opus 4.8 → Sonnet 4.6 saves 40%; Gemini 3.1 Pro → Flash saves 96%. Why Anthropic's tier gap collapsed to 1.7×, why Google's is 24×, and how to tell whether a routing layer is worth building.

o3 and GPT-4.1 both list at $2/$8 per million tokens, yet the same job billed 4.3× more on o3. The reason is hidden reasoning tokens charged at the output rate — a $144 month becomes $624. The real math, plus the one parameter (reasoning_effort) that cut it 62%.

Real 2026 embedding prices per 1M tokens ($0.02 to $0.15) and what a 50M/200M/1B-token corpus costs on each. The twist: the API is a rounding error — the vector database your dimension count feeds is the recurring bill, and switching models means re-embedding everything.

Real 2026 STT prices, all converted to per-hour: Deepgram Nova ~$0.26/hr, Whisper $0.36/hr, AssemblyAI $0.37/hr. What 100/500/2,000 hours a month cost on each — and why add-ons like diarization quietly double the bill.

Real 2026 prices with worked per-call math: GPT-5 flagship is $1.25/$10 per 1M — half GPT-4o's input at the same output, with 3× the context. GPT-5 mini and nano compared, and where the money actually goes.

We priced Cerebras, E2B, Browserbase and LiveKit against one realistic 10-minute agent call. LLM tokens came to 0.89% of the bill; proxy bandwidth came to 74%. Two config habits separate $485 per 1,000 calls from $3,026.

We calculated the cost of 13 AI models across 20 production workloads: customer support, RAG chatbots, content moderation, code review, legal extraction, and more. Gemini Flash-8B wins all 20 on price — 4× cheaper than GPT-4o mini, 145× cheaper than Claude Opus 4.8 for content moderation at 1M calls/month.

Advertised: ~$3,100. Real year-1 bill: $6,170. We modelled a 500-user SaaS and added every hidden fee — Stripe Radar, Billing add-on, Twilio A2P 10DLC registration, Pinecone storage at scale. The combined stack costs 2× what the pricing pages suggest.

Caching changes the ranking. Claude charges a write fee but gives 90% off reads. GPT-4o caches automatically with 50% off. Gemini's context cache costs per hour stored. For a 10k-token system prompt at 500 req/day, Claude Sonnet becomes cheapest after caching.

Stripe went from 2.9% to 3.1%. Google Maps doubled some routes. AWS started charging for public IPv4 that used to be free. A running log of provider price increases with old vs new numbers and bill impact.

Real token math for 100k AI support chats/month. Gemini Flash: $432. GPT-4o: $10,125. The 23× difference, how conversation history grows your bill, and three levers to cut costs without changing models.

We priced 20 production workloads — email classification, code review, RAG, legal extraction and more — across four pricing tiers at a fixed 1,000 req/day. Tier-by-tier token math: cheapest vs most expensive differs by 100×.

We tracked every major LLM price event since GPT-4's $30 launch: frontier-class input is down 97% in three years, both discontinuous shocks came from DeepSeek, and output prices quietly deflate slower than input. What the ~10×/year curve does to annual contracts, unit economics and build-vs-wait calls.

Every API call re-sends the full conversation history as input. By turn 30, you're paying to process 11,300 tokens instead of 1,150. A 30-turn conversation totals over $1 — 4× what it would cost stateless. The math behind the accumulation and four ways to manage it.

The per-call rate you compare between vendors is usually under a fifth of the real cost. We added up developer build time, monthly maintenance and usage for a typical integration — and the "cheaper" API on paper is often the pricier one in your ledger. Where volume flips the story, and how to actually compare two APIs.

A free tier isn't a plan, it's a runway. At 20% monthly growth, jumping from 40% to 80% utilisation doesn't halve your runway — it cuts it four-fold. We ran the logarithmic math on real 2026 free tiers (Gemini, Resend, Supabase, Cloudflare) and the reset-vs-hard-stop trap that breaks apps in production.

Same workload, every major model: $450 on Grok 4.1 Fast versus $17,500 on Claude Opus 4.6 — a 39× spread. xAI took the cheap end of the frontier in 2026, OpenAI and Google split into budget and premium tiers, and Claude stayed top of the ladder. The real per-million numbers.

Your tier allows 500 requests a minute, so why the 429s at eight? The tokens-per-minute limit binds first — at 4,500 tokens a request, OpenAI's entry tier caps you under 7 RPM, about three concurrent users. The real math on every tier, and the four levers that fix it.

We size a real RAG workload — 500k vectors, 1536 dimensions, 1M queries/month — and the bill lands in single-to-low-double-digit dollars on every provider. The storage layer is the cheap part of RAG; here's the math and which model to pick.

Plivo wins on platform rate, but the US carrier pass-through fee adds ~$300 the same for everyone — and emoji-laden templates quietly bill as 2–3 segments. The all-in math on 100k messages, and when to skip SMS entirely.

Reusing the same system prompt or RAG context? You're paying full price to re-read it. Caching cuts input cost 50–90% — with the 2026 discounts for OpenAI, Anthropic and Gemini, plus the write-premium gotcha that can make your bill go up.

Free tiers compared on calls/min and data freshness, where paid plans start, and the per-asset-class billing quirk that turns Polygon's $29 into $58. Which to pick for dashboards, live quotes or tick data.

A worked breakdown on $10,000/month: why the spread is $3,000/yr, why small orders pay nearly double, and when a Merchant of Record's higher fee is actually the cheaper choice.

The headline price is rarely what you pay. SMS carrier fees, email IPs, maps SKUs, payout charges and AI output tokens — the line items that add 30–50% after launch, with real numbers.

Same workload — 1M requests, 1,000 in / 500 out tokens — priced on every major model. The gap between cheapest and priciest is 35×.

A numbers-first comparison of OpenAI, Anthropic and Google API pricing — and the 175× spread between cheapest and priciest for the same workload.

Genuinely useful free API tiers across AI, email, maps, crypto data and search — and how far each really gets you.

Payment API fees compared — and when a merchant-of-record like Paddle is worth the higher cut.

Why the carrier fee, not the platform, often decides your SMS bill.

SendGrid vs Resend vs Amazon SES vs Postmark — pricing, free tiers and which is cheapest at your volume.

Pricing, models, strengths — and which to pick for your use case.