Models
The best AI coding plans per dollar in 2026, and the free models that almost beat them
Every top AI coding plan ranked by the value it gets you per dollar, with the models worth paying for and the free ones that come close. If you are choosing where to spend, start here. Every figure dated and sourced.
The value ranking
Every row is ranked on one number: capability per dollar. It is the model's coding score (Artificial Analysis Coding Index, 0–100) divided by how much a million tokens of it costs. Higher is better: more coding ability for every dollar of tokens. Best route is the cheapest credible place to actually buy the model.
DeepSeek V4 Flash is the best value at about 4,934 capability points per dollar via OpenCode Go, then GLM-5.2, GPT-5.6 Luna, Kimi K3 and Grok 4.5. The Gemini and Claude models, valued from API prices, sit lower.
Bars use a log scale: each gridline above is 10× the last, so equal bar-length gaps mean equal multiples, not equal differences. DeepSeek's real advantage over Claude Fable 5 is about 700×; a linear axis couldn't show both ends at once.
The * rows (Claude, Gemini, GPT-5.5) are quota plans: the vendor publishes no solid data on how much you actually get on the plan, so we fall back to the API price. The real plan value could be better or worse than this estimate. The unstarred rows use published plan economics.
Free models worth running
The best-value model is one that costs nothing. Several routers and plans carry genuinely free frontier and open-weight models, the kind of thing the pricing pages bury. All rotate, so treat these as current, not stable.
Ranked by coding capability, the strongest free coding models you can run today:
Where to get them free
OpenRouter
Free models: Laguna S 2.1, North Mini Code, Nemotron 3 Ultra, GPT-OSS 20B, Laguna XS 2.1, Ling 3.0 Flash, Gemma 4, Nemotron Nano
Limits: ~20 req/min, ~50–200 req/day (no card needed); the :free router picks the best free model automatically
Lineup rotates; several free models were pulled in 2026 (Qwen3 Coder, DeepSeek R1). Good for prototyping, not a stable dependency. Scores are on each model page, from Artificial Analysis.
OpenCode Zen
Free models: DeepSeek V4 Flash, MiMo-V2.5, Laguna S 2.1, North Mini Code, Nemotron 3 Ultra, Big Pickle (stealth)
Limits: Free while in the rotation; no card needed to try them
Rotates monthly and is explicitly not permanent. Some free endpoints log data for model improvement, so don't send client code you care about.
Freebuff
Free models: DeepSeek V4 Flash, MiniMax M3, Kimi, open models
Limits: ~5 free hours of DeepSeek V4 Flash per day; ad-funded, no subscription, no API key
The daily figure is the founder's stated quota, not a contract. Ad-supported.
GitHub Copilot
Free models: Claude, GPT-5.6, Gemini, Grok 4.5 (multi-model)
Limits: Unlimited completions free; agentic coding burns AI credits
New Pro/Pro+/Max signups paused since 1 Jun 2026; still valid for existing subscribers.
Kiro (AWS)
Free models: Auto + model multipliers
Limits: Free 50 credits on signup
Credits run out fast; it is a spec-driven agent, not raw model access.
Cursor
Free models: Claude, GPT-5.6, Gemini, GLM-5.2, Kimi K3
Limits: Hobby tier free
Free tier is usage-quota; you pay for real work.
Flat-rate plans worth checking
These sell a quota rather than a price per token, so they are not on the value chart (we will not invent a number for them). But several are the cheapest route to a strong open model. Each plan is listed individually with its real price.
Method, sources, date
Capability: Artificial Analysis Coding Index (Terminal-Bench v2.1 + SciCode), best variant. Value = capability ÷ effective blended $/M, where blended $/M = 0.85 × (0.34×input + 0.66×cached) + 0.15×output, a coding-agent-typical profile. Unstarred rows use published plan economics; starred rows use the API price because no plan economics are published. All figures re-verified 2026-08-05.
Universe: models reachable through a subscription coding plan, or a free model tier on a router or plan. Where official sources disagree (OpenAI's Codex credit card, the Claude Max token window, Gemini Ultra prices) we show the corroborated figure and flag the conflict rather than average.
This page ranks price. It doesn't rank capability.
Value per dollar is the axis this page can measure honestly, because both halves of it are published. The question of which tool can actually see your project is a different one, and we've written it up separately: the questions worth asking instead of reading a ranking.
Which model you pick matters less than how you review it
The chart tells you what to rent. A fixed-price audit tells you whether the code it wrote is safe to ship.
[email protected] · replies within 1 business day
