The whole field on one printable page. Pick the right model, the right generator, the right prompt structure — and the right cost lever — without the trial and error.
From the team that built Neuron's 89-lesson curriculum. Last updated: May 2026 · v2.0
Six frontier models, six personalities. Prices are USD per 1M tokens, May 2026. Output is typically ~5× input. Context = max tokens in/out per call.
| Model | Best for | Context | $/1M in | $/1M out | The catch |
|---|---|---|---|---|---|
| Claude Sonnet 4.7 | Code, agents, careful long-context, instruction-following — the default workhorse | 200K | $3 | $15 | Not for image generation |
| Claude Opus 4.7 | Hardest reasoning · architecture · math · strategic writing | 200K | $15 | $75 | 5× the cost of Sonnet — use only when Sonnet fails |
| Claude Haiku 4.7 | Classification · routing · cheap bulk work · simple extraction | 200K | $0.80 | $4 | Doesn't reason deeply — pair with caching for the cheapest stack |
| GPT-5 | General chat · multimodal · familiar UX for non-technical users | 256K | $10 | $30 | No durable quality edge — strong everywhere, dominant nowhere |
| Gemini 2.5 Pro | Massive context (2M) · native video · cheap per token | 2M | $1.25 | $10 | Pick it when you need to dump entire codebases or video |
| Gemini 2.5 Flash | Cheap high-volume routing · 1M context · fast | 1M | $0.075 | $0.30 | 40× cheaper than GPT-5 input. Quality dip on hard reasoning. |
| Llama 4 (self-host) | Data that can't leave your servers · fine-tuning · regulated industries | 128K | $0 | +GPU | You run the infrastructure. Real cost ≈ $0.30–$1.20/M @ scale |
| Mistral Large 2 | EU data residency · GDPR-clean · strong open weights | 128K | $2 | $6 | Geography is the moat — benchmark-wise it's mid-pack |
| Grok 4 | Real-time X data · fewer guardrails · breaking news | 128K | $5 | $15 | Optional for everything that isn't "what's happening right now" |
Five generators. The right one depends on whether you need beauty, readable text, photo realism, prompt fidelity, or full control.
| Tool | Best for | Signature |
|---|---|---|
| Midjourney v7 | Beautiful by default — mood, art direction, concept art | --sref style refs |
| Ideogram 3.0 | Anything with readable text — logos, posters, ads, signs | text fidelity |
| Imagen 4 Ultra | Photorealism + clean text — product shots, ads | photo realism |
| DALL·E 3 | Complex multi-element prompts where every detail must land | prompt fidelity |
| Flux 1.1 Pro | Open weights — run locally, fine-tune, no platform rules | self-host |
All six can make 5–10 second clips. Pick by what you need: cinematic shot · synced audio · iteration control · wild motion · cheap b-roll.
| Tool | Best for | ~ Cost / 10s |
|---|---|---|
| Sora 2 | Cinematic, physics-rich, long hero shots | $0.80 |
| Veo 3 | The only one with native synced audio — dialogue, ads | $1.20 |
| Runway Gen-4 | Control + iteration — motion brush, image-to-video, camera nodes | $0.50 |
| Kling 2.0 | Wild dynamic motion and physics | $0.40 |
| Pika 2 | UGC vibe + lipsync | $0.30 |
| Hailuo | Cheap, fast b-roll volume | $0.10 |
Four lanes: voice generation, music generation, transcription, and sound design. Pick one per lane — they don't overlap.
| Tool | Lane | Best for |
|---|---|---|
| ElevenLabs v3 | voice | Voice cloning, 30+ languages, inline emotion tags |
| OpenAI Whisper v3 | transcription | The gold standard — multilingual, punctuation-aware, ~$0.006/min |
| AssemblyAI | transcription | Speaker diarization + sentiment + topic detection out of the box |
| Suno v4 | music | Full songs with vocals from a prompt; stems for remixing |
| Udio | music | Instrumental quality edge over Suno; tighter genre adherence |
| ElevenLabs SFX | sound design | Generated sound effects (footsteps, ambient, foley) from text |
Replace "write something good" with five concrete asks. Bake this into every prompt — it's the single biggest delta between novice and expert AI users.
Beyond the 5-part formula. Stack these for compound quality gains.
A naive AI integration runs 5–10× more expensive than necessary. Pull these in order — the savings stack multiplicatively.
| Lever | Mechanic | Typical savings |
|---|---|---|
| Cap max_tokens | Don't let a 50-word answer generate 500. Set max_tokens=300 for emails, 100 for classifications. | 3–8× cheaper |
| Route by task | Use Haiku/Flash for easy work (~80% of calls), reserve frontier models for hard reasoning only. | ~95% cheaper on routed traffic |
| Prompt caching | Reused system prompts and context become "cached" tokens. Cached input is ~90% cheaper + faster. | ~10× cheaper on hits |
| Batch API | Non-real-time jobs (overnight summaries, bulk classifications) get half-off if you can wait ≤24h. | 50% off list price |
Five common roles, five battle-tested tool combinations. Start here, then customize.
Every one of these is something we see weekly in audits. Most cost you 3–10× more than they should — and produce worse output.
The five shifts already in beta. Each one breaks an assumption you have today. Skating to the puck = ~10× advantage by Christmas.
This cheat sheet is the surface. Neuron is 89 lessons that make it stick — spaced repetition, a real Claude tutor, AI-graded projects, and certificates that actually verify what you learned.
Start free → 18 lessons, no card