SU8 Codes
< Back to blog
Official article Updated 11 min read

Best LLM for Coding: Choose by Task

A task-type guide to choosing among the Claude, GPT, Gemini, DeepSeek, GLM and Kimi lineups, plus effort levels, pricing mechanics and how to read coding benchmarks. As of November 15, 2026.

#llm#coding#comparison
Best LLM for Coding: Choose by Task

Everyone searching for "the best LLM for coding" wants a name. The name keeps changing: Google shipped Gemini 4 Argon on September 30, 2026, Carbon followed on October 28, Hydrogen arrived on November 12, and the next round is already queued. There is a more durable way to answer the question: sort your coding work into a few task types, then match each type to the tier each lab officially designs its models for. This guide does that. All facts are as of November 15, 2026, with sources linked inline. If you haven't picked the agent tool itself yet, that decision has its own guide: Choosing an AI Coding Agent.

The short answer

  • Small, quick edits (a rename, a one-off script): use a lightweight tier, or your daily model at low effort.
  • Everyday feature work (new modules, well-scoped bug fixes, tests): use each lab's daily-driver tier at its default setting.
  • Hard problems (architecture decisions, mysterious bugs, large refactors, long multi-step runs): use the frontier tier and raise the effort level.
  • Huge context or batch jobs: pick by context window and by pricing mechanics (off-peak hours, batch APIs), not by leaderboard rank.

The rest of this article shows where each recommendation comes from.

What "coding work" actually breaks down into

Four task types cover most of what developers do with an AI coding agent:

  1. Small, quick edits. You review every result anyway; the fastest responding model wins.
  2. Everyday feature work. Clear scope, several files, tests to run. Most token budgets are spent here.
  3. Hard problems. Cross-module design, debugging the unknown, refactors, long agentic runs. A wrong step early costs everything after it.
  4. Huge context and batch jobs. Everything above leans on the context window, meaning how much text the model can read at once, measured in tokens. Batch work means many repetitive calls, like generating docs for a whole repository.

How each lab splits its own lineup

Vendors document what each model is for. Here are the official lineups (positions and versions per each vendor's own pages, as of November 15, 2026):

LabFrontier tierDaily tierLightweight / high-volume tierOfficial source
Anthropic (Claude)Fable ("your hardest and longest-running tasks"), Opus 5.5 (complex reasoning)Sonnet 5.5 ("daily coding tasks")Haiku 5.5 (simple tasks, fast and efficient)Model configuration docs
OpenAI (GPT)GPT-6 Astra ("the most demanding reasoning, coding, and professional work")GPT-6.1 Sol ("Near-Astra performance for complex work at a lower cost")GPT-6 Luna ("focused, high-volume tasks")Model-selection reference shipped with the Codex repo¹
Google (Gemini 4)Carbon ("our strongest model"), Argon (long-horizon professional tasks)Hydrogen (built for everyday execution; stands on its own)sameGemini models page, Hydrogen announcement
DeepSeekdeepseek-v4-pro (thinking mode)same (switch to non-thinking for speed)deepseek-flashModels & pricing
Zhipu (GLM)GLM-5.3 (flagship, complex software engineering)sameGLM-5.3-Flash / FlashXGLM-5.3 docs, overview
Moonshot (Kimi)Kimi K3.1 Pro (flagship)Kimi K3 (previous flagship, open-source)K3.1 Pro Highspeed (2–3× faster per the vendor)K3.1 Pro page, pricing

¹ OpenAI's file ships inside the codex repository and describes itself as a possibly outdated snapshot; the canonical docs live at developers.openai.com, which is unreachable from some networks.

Two official statements are worth reading in full. Google's Hydrogen announcement describes the division of labor better than any comparison table: teams use frontier models like Carbon and Argon for "deep architectural judgment when designing systems or untangling subtle bugs," and they still need "a fast, disciplined execution model for writing code, running tests, and verifying changes" (announcement, November 12, 2026). Anthropic's guidance compresses to one sentence: "Choose smaller models for more routine tasks and larger models for more complex or ambiguous tasks" (Claude Code blog, July 7, 2026).

Matching tasks to tiers

Small edits: lightweight tier, or low effort. Haiku 5.5, GPT-6 Luna, deepseek-flash and GLM-5.3-Flash are designed for exactly this. Z.ai's page claims GLM-5.3-Flash matches Claude Opus 4.8 on an intelligence score "at just 1/40 the price" (vendor's own claim, overview). Since you review every result yourself, the frontier tier adds little here.

Everyday work: daily tier at default effort. Sonnet 5.5, GPT-6.1 Sol, Gemini 4 Hydrogen, GLM-5.3 and the Kimi K3 series sit here. Anthropic describes medium effort (the default on its current models) as day-to-day engineering work "with a clear scope, such as implementing a new feature." This is where most token budgets go, so tier discipline pays off most here.

Hard problems: frontier tier, higher effort. Fable, Opus 5.5, GPT-6 Astra, Gemini 4 Carbon / Argon and Kimi K3.1 Pro occupy this tier. For a sense of how close the pack is, Moonshot's page reports K3.1 Pro at 65.7% on Terminal-Bench 4.0, citing GPT-6 Astra at 58.2% and Claude Opus 5.5 at 66.4% (vendor-reported numbers, page, retrieved November 15, 2026). The ordering changes every few weeks; test hard problems on your own repository rather than trusting any single score.

Huge context and batch: pick by window and by pricing. Every lab now offers million-token windows: Claude's Sonnet 5.5 and Opus since 4.7 are native 1M (docs); DeepSeek lists 1M; GLM-5.3 lists 1M with a 128K output cap; the Kimi K3 series lists 1,048,576; Gemini 4 Hydrogen is priced flat across its entire 1M window at $1.50 per million input tokens and $7.50 per million output tokens (announcement, November 12, 2026). For batch economics, see the next section.

The second dial: effort and thinking levels

"Wrong answer" has two causes, and Anthropic's blog gives the cleanest triage: if the model had all the context and clearly tried but still failed, pick a more capable model; if it skipped a file, didn't run the tests, or abandoned a refactor halfway, raise the effort level instead (blog).

Every lab ships this dial. Anthropic uses low / medium / high / xhigh / max, with medium as the default on Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5; the vendor describes high as work "where verification matters or edge cases are likely, such as fixing a bug in an existing codebase," and warns that max can overthink. GLM-5.3 and Kimi K3.1 Pro expose low / high / max with max as the default and thinking always on (GLM-5.3, K3.1 Pro). DeepSeek splits it into thinking and non-thinking modes, with thinking as the default (pricing page). The dial moves real numbers: Anthropic's own tests found higher effort "tested more edge cases and verified more of its work before answering," and Google reports Hydrogen reaching 54 on the Artificial Analysis Intelligence Index at its highest xhigh setting versus 51 at its default high setting (vendor-reported). Keep daily work at default effort and save the raised settings for hard problems.

Pricing is part of the task decision

List price is only the first layer of the bill. Three mechanisms reward task-aware usage:

  • Off-peak pricing. DeepSeek prices off-peak hours at half of peak; peak covers only two UTC windows on weekdays (01:00–04:00 and 06:00–10:00, Monday through Friday, excluding Chinese public holidays), so weekends are entirely off-peak (pricing, November 15, 2026). Time-insensitive batch jobs belong there.
  • Batch APIs (queue large sets of requests for asynchronous processing). Moonshot bills kimi-k3.1-pro and kimi-k3 batch traffic at 60% of the standard price (pricing, November 15, 2026).
  • Lightweight tiers as a cost design. Hydrogen lists at $1.50 / $7.50 against $2 / $10 for Carbon and Argon (Carbon announcement, October 28, 2026), and Z.ai's 1/40-price claim for GLM-5.3-Flash is the most aggressive example of the same idea. Lightweight tiers exist to save money, not just latency.

The same model can also cost very different amounts depending on how you access it. We broke that problem down in Why the Same GPT Model Costs 7.5% of List Price at One API Provider and 95% at Another. Always check each vendor's live pricing page for current numbers.

How to read benchmarks without fooling yourself

Two kinds of numbers circulate. Vendor-reported scores — Google ranking Carbon #1 at 59 on Artificial Analysis Intelligence Index v5.0, Moonshot citing Terminal-Bench to position K3.1 Pro — are all verifiable on official pages, but each vendor picks its own benchmarks and builds its own test setup. Independent leaderboards are the other kind: SWE-bench is the most cited for coding, with Verified and other sub-tracks, and every entry records the model, the agent scaffold (the program harness driving the model) and the cost.

Three habits keep you honest. Compare scores within the same scaffold, because the harness changes results. Read vendor claims across their full benchmark set rather than their best cell. Treat every number as already aging — this article deliberately quotes no current rankings, because they will be stale within weeks. When a decision is on the line, check the leaderboard and official pages the same day, then run the task on your own repository.

One key, many models

Task-based selection in practice means running several models at once: one for hard problems, one for daily work, one for batch. SU8 Codes, the company behind this blog, is an API relay built on that idea — one key that reaches many models. Its model list has 32 models (as of November 16, 2026), covering five of the six labs above (Anthropic, OpenAI, DeepSeek, Zhipu and Moonshot); Google's three Gemini 4 models are not currently listed, so the Gemini 4 descriptions in this article follow Google's official pages. It accepts four protocols directly: OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and Gemini (su8.codes, model list).

FAQ

So which model is actually the smartest?

Smartest at what, measured when, by whom? Google calls Carbon its strongest model; Moonshot says K3.1 Pro approaches the strongest proprietary models; Anthropic reserves Fable for the hardest tasks. Each statement is true in its own frame. For a coding-specific outside reference, check the SWE-bench leaderboard and apply the three habits above.

Will a smaller model break my code?

Sometimes, and Anthropic's triage applies: a failure after full context and a genuine attempt means you need a stronger model; a failure from skipped files, unrun tests or an abandoned refactor means you need higher effort. Most "small model broke my code" reports are the second kind.

Are open-weight models good enough for real coding work?

DeepSeek, Zhipu and Moonshot all position their current flagships for software engineering: DeepSeek v4-pro ships a thinking mode, GLM-5.3 claims open-source state of the art on public benchmarks including Terminal Bench 3.0, and K3.1 Pro's reported public scores land within a few points of the strongest proprietary models (all vendor-reported). The reliable way to decide is the same in every case: run your real tasks on the candidates and compare both results and bills.

Should I switch models every time a new one ships?

No. New releases land every couple of months, and each switch costs re-testing and prompt rework. Protocol-level interfaces (OpenAI-compatible endpoints, Anthropic Messages) are far more stable than model names; build your workflow on the protocol, treat models as swappable parts, and trial each new release on small tasks before committing.