Updated July 2026
Coding with AI
Fast and loose to start, structured once it matters. The workflow for shipping software you would still put your name on.
The workflow, in order
New to this? Work through these four, then come back to the tables below to pick what you run.
- 1 Vibe Coding: From Zero to Hero tuaregs · Guide Start loose. Build by feel and find out fast what these tools are and are not good at.
- 2 Spec-Driven Development tuaregs · Guide Then tighten up: write the spec, let the model build it, keep the final call.
- 3 Git Fundamentals tuaregs · Guide The safety net. You will need it sooner than you think.
- 4 Building effective agents Anthropic · Article The step up: models that work through a task rather than answer a question.
Reference
Route by task
Agent = Model + Harness. Pick a model, a place to run it, and a harness. Re-verify anything older than a month.
Cheap / low-risk
DeepSeek V4 Flash — logging, simple edits
Default workhorse
Kimi K2.6 or Sonnet-class — coding, PR review, refactors, tests
Hard / repeated failures
Frontier — Opus 4.8, GPT-5.5, Fable 5 if available
Private / offline
Qwen local via Ollama
Fallback
Second provider route (e.g. Claude via OpenRouter)
Benchmarks are routing signals, not truth. Set defaults from them, correct with your own failure data.
Measure cost per successful task, not per token. A cheap model that causes rework is expensive.
Models
Price = in / out per 1M tokens, first-party unless noted. Compare SWE-Bench within the same column only.
| Model | Creator · License | $ in / out | SWE-Bench | Also on |
|---|---|---|---|---|
| Claude Fable 5 | Anthropic · Closed | $10 / $50 | 80.3% Pro | Claude Code, claude.ai |
| Claude Opus 4.8 | Anthropic · Closed | $5 / $25 | 69.2% Pro · 88.6% Ver | Claude Code, Cursor, OpenRouter, Bedrock, Vertex |
| Claude Sonnet 5 | Anthropic · Closed | ~$3 / $15 | near Opus 4.8 | Claude Code, Cursor |
| Claude Sonnet 4.6 | Anthropic · Closed | $3 / $15 | strong | Cursor, Windsurf, Cline, OpenRouter |
| Claude Haiku 4.5 | Anthropic · Closed | $1 / $5 | 73.3% Ver | OpenRouter |
| GPT-5.5 | OpenAI · Closed | $5 / $30 | 58.6% Pro · 88.7% Ver | Codex CLI, Copilot, Cursor, Azure |
| GPT-5.3-Codex | OpenAI · Closed | $1.75 / $14 | coding-tuned | Codex CLI |
| Gemini 3.1 Pro | Google · Closed | $2 / $12 ($4/$18 >200K) | 54.2% Pro · 80.6% Ver | Gemini CLI, Vertex, Cursor, OpenRouter |
| Grok 4.x | xAI · Closed | ~$2.60 / $7.80 | Tier B | Cursor, OpenRouter |
| DeepSeek V4-Pro | DeepSeek · MIT | $1.74 / $3.48 | 80.6% Ver | OpenRouter, Together, DeepInfra, Morph, HF |
| DeepSeek V4-Flash | DeepSeek · MIT | $0.14 / $0.28 | 79% Ver | OpenRouter :free, DeepInfra, Morph, Ollama |
| DeepSeek V3.2 | DeepSeek · MIT | ~$0.23 / ~$1 | best value classic | OpenRouter, Ollama |
| GLM 5.2 | Z.ai · Open | $1.40 / $4.40 (~$0.45/$3.31 OR) | 62.1% Pro | OpenRouter, HF, GLM Coding Plan |
| GLM-4.7-Flash | Z.ai · Open | free | decent | OpenRouter, Ollama |
| Kimi K2.6 | Moonshot · Open | $0.95 / $4.00 | 80.2% Ver · 58.6% Pro | OpenRouter, Together, Groq, DeepInfra, HF |
| Kimi K2.7 Code | Moonshot · Apache 2.0 | ~K2.6, -30% thinking | coding-tuned | OpenRouter, HF |
| MiniMax M3 | MiniMax · Open | $0.60 / $2.40 (~$0.10/$1.21 OR) | 80.5% Ver | OpenRouter, Atlas Cloud |
| MiniMax M2.7 | MiniMax · Open | $0.30 / $1.20 | 56.2% Pro | Together, Atlas Cloud |
| Qwen3.7-Max | Alibaba · Closed API | via Alibaba Cloud | 80%+ Ver | OpenRouter |
| Qwen 3.6-27B | Alibaba · Apache 2.0 | free local (22GB) | 77.2% Ver | Ollama, HF, Together, OpenRouter |
| Qwen 3.6 Plus | Alibaba · Open | $0.50 / $3.00 | 61.6% Terminal-Bench | Together, OpenRouter |
| Qwen3 Coder | Alibaba · Apache 2.0 | free on OpenRouter | best free coding | OpenRouter :free, Ollama |
| Qwen 2.5 Coder 32B | Alibaba · Apache 2.0 | free local (18GB+) | best local classic | Ollama, LM Studio, Continue |
| Nemotron 3 Ultra | NVIDIA · OpenMDW | ~$0.42 / $2.61 (OR) | #2 open (AA index) | NVIDIA NIM, OpenRouter :free, HF |
| North Mini Code | Cohere · Apache 2.0 | free local | 33.4 AA Coding | HF, Ollama, vLLM |
| Codestral 22B | Mistral · Open | free local (12GB) | best autocomplete local | Ollama, Mistral API, OpenRouter |
| Devstral Small 24B | Mistral · Open | free local (16GB) | near-frontier local | Ollama, Mistral API |
| Llama 4 Scout | Meta · Open | free local / :free OR | solid · 10M ctx | HF, Ollama, Groq, Together |
| Poolside Laguna M.1 | Poolside · Closed | free tier | #1 Kilo usage | Kilo Code, Poolside |
| openPangu 2.0 | Huawei · Open | free local | competitive | HF, vLLM |
The 10x cliff. Five models hit 80.2–80.6% Verified at $2.40–$12/M out. The next 8 pts (GPT-5.5, Opus 4.8 ~88.6%) cost $25–30/M.
Fable 5 caveat. US export-control suspended access June 12, 2026; back for US users ~July 1. Verify outside the US.
Where to run them
| Provider | Type | What it does | Go |
|---|---|---|---|
| Anthropic API | First-party | Claude direct · cache reads at 10% input | open |
| OpenAI API | First-party | GPT direct | open |
| Google AI Studio | First-party | Gemini direct · free tier | open |
| DeepSeek API | First-party | Cheapest frontier-class · off-peak discounts | open |
| Moonshot API | First-party | Kimi K2 direct | open |
| Z.ai | First-party | GLM direct · coding subscription | open |
| Alibaba Cloud | First-party | Qwen direct | open |
| Mistral La Plateforme | First-party | Mistral, Codestral · EU-hosted | open |
| xAI API | First-party | Grok direct | open |
| OpenRouter | Router | 400+ models, one key, fallback · +5.5% credit fee | open |
| ZenMux | Router | Auto-routing (zenmux/auto), failover, output insurance | open |
| HF Inference Providers | Router | Routes to 18 hosts · zero markup | open |
| LiteLLM | Router (self-host) | OSS proxy, keys on-prem · SOC 2 / HIPAA pick | open |
| DeepInfra | Open host | Cheapest per token · widest open catalog | open |
| Together AI | Open host | Broad catalog + fine-tuning · batch 50% off | open |
| Fireworks | Open host | Fastest structured output · day-0 releases | open |
| Groq | Speed host | Lowest first-token latency · free tier 30 RPM | open |
| Cerebras | Speed host | Highest throughput (~3,000 tok/s) · tiny catalog | open |
| Novita AI | Hybrid host | Model APIs + raw GPUs (H100/H200) · spot | open |
| Morph | Open host | bf16 serving, no fp8 quant · codegen speculative decode | open |
| Hyperbolic | Open host | Budget batch inference | open |
| fal.ai | Media gen | 1,000+ image/video/audio models · dev API | open |
| Replicate | Media + LLM | 50,000+ community Cog models | open |
| Higgsfield | Media (consumer) | 50+ video models · $15–99/mo credits | open |
| Modal | Deploy your own | Code-first GPU containers · <1s cold start | open |
| Baseten | Deploy your own | UI-first deploys · sub-400ms voice | open |
| RunPod | Deploy your own | Raw GPU rental · cheap spot | open |
| HF Inference Endpoints | Deploy your own | Any Hub model · dedicated autoscaling | open |
| Nebius | Deploy your own | EU-hosted · GDPR-friendly | open |
| Ollama | Local | Free unlimited · ~0.6GB/B params at Q4 | open |
| Ollama Cloud | Managed local | Same CLI/API, hosted GPUs · Pro ~$20/mo | open |
| LM Studio | Local | GUI for local models | open |
| vLLM | Self-host engine | Production-grade serving | open |
Agents (harnesses)
| Agent | Type | Models | Go |
|---|---|---|---|
| Claude Code | First-party CLI/app | Claude family | open |
| Codex CLI | First-party CLI | GPT family | open |
| Gemini CLI | First-party CLI | Gemini | open |
| Copilot Agent Mode | IDE-native | GPT family | open |
| Cursor | AI IDE | BYOM (Claude, GPT, Gemini…) | open |
| Windsurf | AI IDE | Multiple | open |
| Cline | VS Code ext · BYOM | Anything via OpenRouter/Ollama | open |
| RooCode | VS Code ext · BYOM | Anything · strong multi-file | open |
| Aider | Terminal · BYOM | Anything · git-native | open |
| Kilo Code | BYOM | Anything · free tiers · leaderboard | open |
| Continue.dev | VS Code ext · local | Ollama models | open |
Cost levers
-
Prompt caching
Cache reads ~10% of input (Anthropic). Fireworks/Groq 50% off cached. Biggest lever for agent loops.
-
Batch APIs
50% off async work.
-
Long-context surcharge
Claude 1M ctx: no premium. Gemini 3.1 Pro: 2x input above 200K.
-
Same model, cheaper host
DeepSeek V4 Pro: $1.30 DeepInfra vs $1.74 direct vs $2.10 Together.
-
Subscriptions vs API
Claude Pro/Max, Codex Plus, GLM Coding Plan, OpenCode Go ($10/mo).
-
1/36th rule
V4 Flash ≈ 82% of Opus at ~1/36th input price. Let Opus stop doing Flash's job.
This page ages fast
One email a week, including when the numbers above change. Which they do, often enough to be annoying.
You're in. Check your inbox for the confirmation link.