The Best AI Model for Coding Right Now, by the Numbers

If you write code with an assistant every day, the question is not "which model is smartest" in the abstract. It is which tier earns its keep in your actual loop: fast autocomplete, a chat window for tricky refactors, or a long agentic run that reads a big repo and edits dozens of files. So we took the LMArena coding category and joined it against our live price feed to find the best AI model for coding right now, scored on skill and on what it costs to run.

Data from our live pipeline, updated September 21, 2026. Prices sync every 6 hours.

We have not run our own coding benchmarks here. The capability numbers are LMArena's crowd-voted coding scores; the dollar figures come from our market sync. What we add is the join, the value math, and a developer's read on where each tier belongs. You can dig into any model on the models directory or check the raw value leaderboard.

The coding leaderboard right now

Here is the top ten by coding Arena score, with live per-million-token pricing and our Value Score. A higher Value Score means more coding capability per dollar.

Arena Scores © LMArena, licensed CC-BY-4.0, as of 2026-09-02. Prices are from our live market sync, refreshed every 6 hours.

A few things jump out. Anthropic owns seven of the ten slots, and the very top is a tight cluster: Claude Opus 4.6 at 1534 sits just two points above Opus 5, and the gap from #1 to #10 is only 35 points. On a crowd-voted scale, that is close enough that tie-breaking on price and behavior matters more than chasing the top row.

Top skill: the Claude Opus tier

Claude Opus 4.6 is the highest-scoring coding model in the list, and Opus 5 is essentially level with it. Both run at $5.00 per million input tokens and $25.00 per million output. That output price is the number to watch, because agentic coding tools generate a lot of tokens: reasoning, tool calls, diffs, re-reads. A single long run over a large repo can produce far more output than you expect.

So the Opus tier is where you go when the task is genuinely hard and correctness is expensive to get wrong: a gnarly migration, a subtle concurrency bug, or a refactor that spans modules. It is the frontier of the best AI model for coding on raw skill. It is also the tier that will show up loudest on your bill if you point it at every keystroke. Save it for the work that deserves it.

Note that Claude Fable 5 and Fable 5.1 score well (1519 and 1508) but carry double the price, $10.00 in and $50.00 out. That is why their Value Scores, 15.9 and 15.4, are the lowest in the table. On coding capability alone they do not clear the Opus models, so the premium is hard to justify for pure code work unless you need them for something outside this category.

Best value: Gemini 3.8 Flash

Here is the headline for anyone watching spend. Gemini 3.8 Flash scores 1516 on coding, which ties it with Opus 4.7 and lands it fifth overall, yet it costs $0.75 per million input and $3.75 per million output. That is a Value Score of 210.7, the best value-for-money in the entire coding category. Its sibling, Gemini 3.7 Flash, is right behind at 1504 and 202.7.

Read that again: you are giving up roughly 18 Arena points versus the #1 model while paying about a seventh of the output price. For high-volume, latency-sensitive work, that trade is lopsided in your favor. This is the tier for autocomplete, inline suggestions, quick chat questions, PR summaries, and the outer loop of an agent that fires thousands of small calls. When the token count is the problem, Flash is the answer.

We have not tested how either Gemini Flash model handles very large context windows in a long agentic run, so treat that as an open question for your own repo. But on the join of coding score and price, nothing else in the list is close on value. See the full ranking on the value leaderboard.

The middle ground: Sonnet and open-source Kimi

Two models sit in a useful middle. Claude Sonnet 4.6 scores 1504 at $3.00 in and $15.00 out, for a Value Score of 50.6. That makes it the sensible default for a lot of day-to-day agentic coding: strong enough to trust on real edits, priced well below the Opus tier. If you run a tool like Claude Code or Cursor and want a balance of skill and cost as your standing model, Sonnet is the obvious candidate.

Kimi K3 from Moonshot AI is the open-source entry, scoring 1511 at $1.70 in and $8.50 out, Value Score 91.4. That is the second-best value in the table behind the Gemini Flash pair, and it is open-source, which matters if you want the option to self-host or avoid lock-in. For teams that care about portability, it is worth a serious look.

How to pick by workflow

Match the tier to the loop, not to the leaderboard headline.

• Autocomplete and inline suggestions: you fire constant small requests, so price and latency dominate. Gemini 3.8 Flash or 3.7 Flash. The value math here is not subtle.
• Chat and mid-size refactors: Sonnet 4.6 balances skill and cost. Kimi K3 if you want open-source or cheaper output.
• Hard agentic runs over a big repo: Claude Opus 4.6 or Opus 5 when correctness is worth the $25.00 output rate. Budget for token volume, because agents are generous with it.
• Mixed stack: many teams route cheap traffic to Flash and reserve Opus for the hard calls. Our stacks and tools pages show how people wire this up.

If you are deciding between assistants rather than raw models, the Claude Code vs Cursor comparison and our coding hub go deeper on the tooling side. For the broader model debate, Claude vs ChatGPT and ChatGPT vs Gemini cover the trade-offs beyond coding.

The honest read

Right now the best AI model for coding on pure skill is Claude Opus 4.6, by a two-point margin over Opus 5. But the best model for most developers most of the time is a value question, and there the answer is Gemini 3.8 Flash: near-top coding skill at a fraction of the cost. The smart move is not one model, it is a routing policy that sends cheap work to Flash and expensive work to Opus. Prices move; we resync every 6 hours, so check pricing and trending before you commit a budget.

FAQ

What is the single best AI model for coding right now?

By LMArena coding score, Claude Opus 4.6 at 1534 is #1, just ahead of Claude Opus 5 at 1532. Both cost $5.00 per million input and $25.00 per million output.

Which coding model gives the best value?

Gemini 3.8 Flash, with a Value Score of 210.7. It scores 1516 on coding at $0.75 input and $3.75 output, the best skill-per-dollar in the coding category.

Is there a strong open-source option?

Yes. Kimi K3 from Moonshot AI scores 1511 at $1.70 input and $8.50 output, a Value Score of 91.4, and it is open-source if you want portability or self-hosting.

Milo, Scout AI Team