Model profile & tracker
Best LLM for Coding: Tiers, Prices and Local Picks (2026)
The best LLM for coding by tier, with a 200K-token coding task priced on seven models. Frontier, budget, open-source and local picks. Checked 22 November 2026.
There is no single best LLM for coding, and anyone who tells you otherwise is selling one. What the numbers can tell you — and what this page does — is which models sit in each tier of the coding market as of 22 November 2026, what a typical coding task costs on each, and where the honest uncertainty is. The prices are measured; the coding ability is what each provider positions the model for, cross-checked against what the model is actually built for.
The tiers, in one table
| Tier | Model | Input $/1M | Output $/1M | Context | Weights |
|---|---|---|---|---|---|
| Maximum quality | Claude Fable 5.2 | 8.00 | 40.00 | 1M (128K max out) | Closed |
| Maximum quality | GPT-6.2 Astra | 8.00 | 40.00 | 1.05M | Closed |
| Anthropic's default | Claude Opus 5.5 | 4.00 | 20.00 | 1M | Closed |
| Balanced | Gemini 4 Neon | 3.00 | 15.00 | 1.05M | Closed |
| Budget hosted | Qwen4-Max | 1.50 | 4.50 | 1M | Closed |
| Budget hosted, image input | Mistral Large 4 | 0.70 | 2.10 | 1M | Open (preview) |
| Self-host, multi-step coding | DaisyRose 27B | 0.15 | 1.875 | 262K | Open (Apache 2.0) |
| Self-host or cheap API, volume | Inkling 2 Mini | 0.15 | 0.50 | 1.05M | Open |
Notes: Fable 5.2's fast mode is $16 / $80 and its batch rate $4 / $20; Astra's batch rate is $4 / $20; Mistral Large 4's $0.70 / $2.10 is a preview promotion (list $1.36 / $4.18).
What a coding task costs
A representative refactor or debugging session: 200K tokens of repository context in, 20K tokens of diff and explanation out. Priced end to end:
| Model | Input cost | Output cost | Total |
|---|---|---|---|
| Inkling 2 Mini | $0.03 | $0.01 | $0.04 |
| Mistral Large 4 | $0.14 | $0.042 | $0.18 |
| Qwen4-Max | $0.30 | $0.09 | $0.39 |
| Gemini 4 Neon | $0.60 | $0.30 | $0.90 |
| Claude Opus 5.5 | $0.80 | $0.40 | $1.20 |
| Claude Fable 5.2 | $1.60 | $0.80 | $2.40 |
| GPT-6.2 Astra | $1.60 | $0.80 | $2.40 |
The spread is 60× between Inkling 2 Mini and the frontier pair — on the same task. Multiply by your daily session count before you pick a default.
Which tier you actually need
If the code has to work the first time
Claude Fable 5.2 or GPT-6.2 Astra, both $8 / $40. This is the tier for architecture, gnarly debugging and code that ships to production without a reviewer. Anthropic positions Fable 5.2 (released 9 November 2026, always-on thinking at high effort) for exactly this: hardest long-horizon work, and on per-task cost it has beaten far pricier models on coding benchmarks. Astra is the same price with a 1.05M window; the choice is which lab's failure modes your codebase punishes less.
If you want one model and a lower bill
Claude Opus 5.5 at $4 / $20 — the model Anthropic itself recommends for most tasks — or Gemini 4 Neon at $3 / $15 if you also feed it screenshots or design files. Both carry a full ~1M context window, which matters more for coding than for chat: the whole module has to fit, not just the function.
If the task is boring and the volume is high
Qwen4-Max at $1.50 / $4.50 for boilerplate, tests and small diffs, or Mistral Large 4 at $0.70 / $2.10 — the cheapest model that also reads images, so it can work straight from screenshots and error screenshots. Both are a fraction of the frontier bill for the 80% of coding work that is repetitive.
The cheap-first pattern
The cheapest coding setup that still ships: run the cheap model first and escalate only the failures. A $0.39 Qwen4-Max attempt plus a $2.40 Fable 5.2 rescue costs less than two frontier attempts. Budget retries into any comparison — a cheap model that needs two attempts has doubled its price.
Best open-source LLM for coding
Three candidates, and they are not interchangeable:
Mistral Large 4 (1.05T total / 52B active, 1M context, text and image input) is the strongest open-weight model in the index and the only one with multimodal input. The weights shipped at the end of October 2026; it is still a preview release, and the $0.70 / $2.10 hosted price is a promotion against a $1.36 / $4.18 list rate.
DaisyRose 27B (Apache 2.0, 262K context) is the better fit for multi-step coding and reasoning on modest hardware — the licence is the cleanest in the table and the model is built for tool use.
Inkling 2 Mini (26.4B total / 3.4B active, 1.05M context) is the better fit for volume: $0.50 output, $0.015 cached input, and a window long enough to hold a whole repository. It will not architect your system; it will review diffs, write tests and draft the boring parts all day for cents.
The honest boundary: open weights mean you can self-host, not that self-hosting is cheaper. At low or spiky usage, the hosted API (even the same model) usually wins; at high steady volume, owning the GPUs wins.
Best local LLM for coding
If "local" means Ollama on your own machine: Inkling 2 Mini is the practical pick — 3.4B active parameters puts it in reach of a 32GB machine, with the full 1.05M window intact. DaisyRose 27B runs multi-step coding on modest hardware but caps at 262K context. For anything heavier, local stops being cheaper than a frontier API and you are better off paying per token.
For a local llm comparison that includes throughput, Inkling 2 Mini decodes at 210 tok/s on a 96GB M5 Ultra Mac Studio — the LLM speed comparison has the local numbers alongside the API models.
What local actually buys you: the code never leaves your machine, and there is no rate limit at 3am. What it does not buy you: frontier-tier reasoning. A local model that solves 70% of your tasks is a productivity win; pretend the other 30% does not exist and you will ship the wrong thing.
Which LLM is best for coding? The honest answer
The benchmark answer changes monthly and the leaderboard answer depends on who ran the test — read the benchmark guide before you trust any score. The pricing answer is stable: Fable 5.2 or Astra when correctness is the whole job, Opus 5.5 or Neon as the daily default, Qwen4-Max or Mistral Large 4 for volume, and open weights when the code cannot leave your network. Then run your own evaluation on about fifty of your real tasks — it takes an afternoon, and it beats every list on the internet, including this one. For the pure price question across all workloads, the cheapest LLM API breakdown prices the cheap tier end to end.
Frequently asked questions
Which LLM is best for coding in 2026?
At the top tier, Claude Fable 5.2 and GPT-6.2 Astra (both $8 / $40, prices checked 22 November 2026). For most work, Claude Opus 5.5 ($4 / $20) or Gemini 4 Neon ($3 / $15) is the better value. "Best" depends on whether your bottleneck is correctness, cost or privacy.
What is the best free or open-source LLM for coding?
Mistral Large 4 for overall strength with image input, DaisyRose 27B for multi-step coding on modest hardware (Apache 2.0), Inkling 2 Mini for volume and long context. Free to download is not free to run — you pay for GPUs instead of tokens.
Is a cheap LLM good enough for coding?
For tests, boilerplate, reviews and small diffs: usually yes. For architecture and debugging that has to be right the first time: use the cheap-first pattern and escalate failures to a frontier model instead of trusting the cheap tier with the whole job.
Does context window matter for coding?
Yes, more than for chat: the whole module — or the whole repo — has to fit in the window for a model to refactor rather than guess. Seven models in our index carry ~1M windows at prices from $0.50 to $40 output, so a long window no longer costs frontier money. The model index has the full table.