AI API Pricing, Verified — What DeepSeek & the Others Actually Charge
AI moves faster than the posts written about it. Everything here is built the same way: go to the source, save a dated snapshot, then publish — whether that is a rate, a spec, or a fix for an error you just hit. The short answer comes first; the reasoning follows for anyone who wants to check it. Where a partner is not the cheapest or the best fit, the page says so out loud — affiliate links and ads are both disclosed on every page they appear on.
48 articles · 6 topics · Every figure traceable to its source · Last verified 12 Oct 2026
Browse by topic
Each topic is a self-contained index — every page lives under one, so nothing here depends on being linked from this front page.
- DeepSeekOfficial rates, the off-peak discount, and what the same model costs on OpenRouter and Novita.8ComparisonsModel and tool head-to-heads: what differs, what we measured, and who each one suits.5Guides & answersWhat things mean, how to start in minutes, and the errors everybody hits — the short answer first.13AI coding toolsField reports from building real projects with AI coding tools: what budget models actually ship.11Self-hostingRunning models on your own hardware: VRAM math, quantization trade-offs, and when local actually beats the API.5AI newsReleases, pricing changes and deprecation deadlines, reported the day they land, with sources and a dated caveat.6
Start here
One from each kind of answer we publish — a number you can check, a thing you can build, and a first-hand account.
- Cheapest DeepSeek V4 API — 4 Options ComparedDeepSeek V4 Flash costs $9–$77 a month depending on where you run it. We compared all four options — and why most price quotes you'll find are wrong.DeepSeek V4.1 Flash vs V4 Pro: We Checked the “Comprehensive Win” ClaimWe checked DeepSeek’s claim that V4.1 Flash beats V4 Pro across the board: 14 of 16 benchmarks go to Flash, but GPQA Diamond and HLE don’t. USD pricing inside.DeepSeek V4 API Cost CalculatorCompare what DeepSeek V4 costs per month across OpenRouter, Novita AI and the official DeepSeek API, using your own token volume.DeepSeek Long-Context Flakiness, Measured: Phase Sensitivity in Chunked KV-Cache CompressionA ByteDance Seed paper pinpoints why DeepSeek long-context retrieval flakes: chunked KV-cache compression makes accuracy swing up to 40 percentage points with token position. Numbers and fixes.
Recently updated
- 80B Models on 4GB RAM: The Real Numbers Swiftlet runs Qwen3-Next-80B in 4.3GB of RAM at 4.5-5 tok/s on an M5 Mac. An iPhone 17 manages 1 tok/s. AirLLM fits 70B in 4GB VRAM with no quantization and publishes no speed at all. The numbers that decide whether any of this is usable.
- Cursor vs GitHub Copilot (2026): Full Price Comparison Cursor vs GitHub Copilot 2026 pricing compared tier by tier from official pages: $10 vs $20 entry, $39 vs $60, Max $100 vs Ultra $200, plus Copilot's new AI Credits base-and-flex model.
- Microsoft-Decision-1: A $0.042 Decision Model for Agent Loops Microsoft-Decision-1 (launched Oct 9, 2026) scores fixed option sets with calibrated probabilities — post-trained from Alibaba's Qwen3.5-9B, 32,768-token context, hosted API only. Pricing: $0.042 per million input tokens, output free — matching Jev's September playbook. 83.5% average accuracy across 36 vendor-run benchmarks, 85 ms p50, 35× faster than GPT-6 Sol. What it's for, the OpenRouter listing gap, and what to verify first.
- TRAE Merges Code and Work: What It Costs Now TraeCode and TraeWork merged into one TRAE app on October 9, 2026. Current plans: Free $0, Pro $20, Pro+ $60, Ultra $200 a month — yearly billing saves 17%. The catch: free-tier usage is a one-time allowance that does not reset monthly. Full price table, migration steps for existing users, and the fine print before you commit.
- DeepSeek Long-Context Flakiness, Measured: Phase Sensitivity in Chunked KV-Cache Compression A ByteDance Seed paper pinpoints why DeepSeek long-context retrieval flakes: chunked KV-cache compression makes accuracy swing up to 40 percentage points with token position. Numbers and fixes.
- GLM-5.3 on Amazon Bedrock GLM-5.3 (753B MoE, 1M context, CyberGym 84.5) is now on Amazon Bedrock, gated to eligible enterprise customers via cross-Region inference only. Bedrock charges $1.68/$5.28 per million tokens on the Global profile — a flat 20% above Z.ai's direct $1.40/$4.40 — but the Flex service tier halves Standard to $0.84/$2.64, undercutting direct pricing for queue-tolerant batch workloads. Priority runs +75%. Full price table, the enterprise-gating fine print, and which channel fits which team.
Looking for the machine-readable version? Every page is listed in LLMs.txt.