The Best LLMs for Coding

The gap between “best coding model” and “best coding model for you” is wider than ever. As of October 2026, the closed frontier has split into a $10/$50 tier reserved for the hardest work and a $2/$10 tier that does most of the rest, while MIT-licensed models from China now report DeepSWE scores in the same range as the closed leaders. Which one you should use depends almost entirely on what your coding workload looks like.

This ranking covers the six models we’d actually recommend, why, and — just as important — the catch on each one.

How we picked

Three rules, consistent with everything on this site:

1. Benchmarks with provenance. We weight the hardest agentic benchmarks (Terminal-Bench 4.0, DeepSWE v1.1, CursorBench 4.0) over saturated ones. Almost every score below is vendor-reported, and different vendors run different harnesses, so we label the source and treat cross-vendor gaps of a few points as noise.
2. Real pricing, per task where possible. Sticker price per token misleads — models differ several-fold in how many tokens they spend on the same job, and cached reads now dominate agent bills. Where per-task data exists, it overrides the rate card.
3. Generally available only. Gemini 4 Argon (limited to trusted defenders for now) and Claude Mythos 5.1 (trusted-access programs) don’t qualify, and Mistral Large 4 is still a preview with weights due at the end of October. Rankings shift when they do.

The short answer

All six, side by side

The one benchmark most of this list shares is Terminal-Bench 4.0, a multi-hour terminal-work suite. Here is every pick that has published a score, plus two relevant references:

And the full decision matrix — pricing and context from each model’s page, benchmarks from the vendors’ published results:

*Reported by the vendor or in a competitor’s table on a different harness. Opus 5.5 is quoted at xhigh effort; treat cross-column comparisons as directional, not decimal-precise.

The reasoning, pick by pick

1. Claude Opus 5.5 — the default

Opus 5.5 arrived on September 22 at $4/$20 — 20% below Opus 5’s list price and, with fewer tokens per task, about 40% cheaper to run on typical workloads. On Anthropic’s own table it leads Fable 5.1 on every row shown: Terminal-Bench 4.0 66.4% vs 55.8%, CursorBench 4.0 57.8% vs 51.8%, OSWorld 2.1 81.8%. Anthropic itself says to start here for most workloads, and the price gap to Fable makes that easy advice to take.

Two details to plan for: thinking can no longer be switched off, and safeguards can reroute cyber and biology requests to an older model, so security-adjacent agents behave differently from the headline numbers.

Strengths: frontier results at $4/$20; 1M context and 128K output; fast mode at up to 2.5x speed ($8/$40); strong prompt-injection resistance. Watch out: trails GPT-6 Astra on agentic science and AutomationBench; safeguard fallbacks; always-on thinking; Anthropic cautions that benchmark margins are a weaker guide at this level.

2. Claude Sonnet 5.5 — the value pick

Sonnet 5.5 (September 28) keeps Sonnet 5’s $2/$10 price and changes almost everything else. Terminal-Bench 4.0 goes from 10.3% to 70.6%, a score that tops Opus 5.5’s 66.4% on Anthropic’s table, CursorBench 4.0 is 55.5% against Opus’s 57.8%, output is 30%+ faster, and Anthropic estimates up to 30% lower cost per task. Cache reads were halved to $0.10 on October 7, which makes long agent loops around 20% cheaper again.

Anthropic is candid about where it stops: Opus 5.5 is still clearly stronger on complex, open-ended work that needs sustained judgment. Use Sonnet for well-scoped tasks and as the subagent under an Opus orchestrator.

Strengths: the best Terminal-Bench 4.0 score on this list at $2/$10; faster and leaner than Sonnet 5; cache reads $0.10. Watch out: weaker on ambiguous, sprawling jobs; higher-risk cyber work falls back to Sonnet 5; Anthropic later restated some figures, so decimals vary between its pages.

3. GPT-6.1 Sol — the Codex pick

GPT-6.1 Sol (September 29) is OpenAI’s answer to the same $2/$10 slot. OpenAI says it matches the $10/$50 GPT-6 Astra on DeepSWE v1.1 at about a fifth of the cost, beats GPT-6 Sol’s best score by 6.4 points at lower effort, and sits 2.2 points above Opus 5.5 on AutomationBench at roughly a third of the cost. It has a 1.05M context, cached input at $0.10, and a faster Ultrafast variant for Codex was announced to follow.

The gap is evidence: OpenAI has not published a Terminal-Bench 4.0 score for it, and nearly every number is the vendor’s own. If your team lives in Codex or the OpenAI API, run it against Sonnet 5.5 on your own repo.

Strengths: $2/$10 with 1.05M context; near-Astra results at a fifth of the price (OpenAI-reported); a price ladder from Luna ($0.10/$0.50) to Astra ($10/$50). Watch out: confusing ladder (both GPT-6 Sol and GPT-6.1 Sol exist; target 6.1); no Terminal-Bench 4.0 score; text and image input only.

4. Claude Fable 5.1 — the long-horizon specialist

Fable 5.1 (September 1) is the generally available face of Anthropic’s Mythos-class models. Cache reads fell 75% to $0.25, which Anthropic estimates cuts a typical bill by about 25% and a highly agentic one by up to 45%; cyber false positives fell 60%; and customers describe unusually long, reliable runs — a 38-hour unattended ML job at Ramp, a crash nobody had explained in four to five years traced at Millennium.

The honest framing is Anthropic’s own: start on Opus 5.5 and move to Fable only when evals on Opus at higher effort still fall short. At $10/$50, 2.5x Opus’s list price, Fable has to earn its slot on exactly the work that is hardest.

Strengths: the capability ceiling for multi-day autonomous work; Terminal-Bench-Science 0.1 52.6%; cheaper caching than Fable 5. Watch out: $10/$50; classifiers can reroute or zero out security tasks; behind Opus 5.5 on Anthropic’s own table; the least-restricted sibling (Mythos 5.1) is trusted-access only.

5. GLM-5.3 — the open flagship

GLM-5.3 (August 14) used the same 753B base as GLM-5.2 and changed only post-training — and DeepSWE v1.1 rose from 46.2 to 66.9, Terminal-Bench 3.0 from 4.6 to 28.3. It keeps the 1M context, ships weights on Hugging Face under a permissive MIT-style license, and costs about $1.40/$4.40 via API. The cost of ownership is literal: a 753B MoE needs a multi-GPU node to self-host.

It is also still behind the closed frontier on the hardest rows — 28.3 vs 34.6 for GPT-5.6 Sol on Terminal-Bench 3.0 in Z.ai’s own table — and text-only, so screenshot-to-code workflows need another model.

Strengths: a flagship-scale open-weight coder whose weights you can audit; MIT-style license; 1M context; a fraction of frontier pricing. Watch out: 753B MoE self-hosting is an ops project; text input only; mandatory thinking; vendor-reported scores.

6. DeepSeek V4.1-Flash — the budget agent

DeepSeek V4.1-Flash (September 10) is a 552B MoE with just 8B active for input and 16B for output, a 1M context, MIT license and a KV cache a quarter the size of the previous generation’s. DeepSeek reports DeepSWE v1.1 74.2 and Terminal-Bench 2.1 90.6%, at $0.15/$0.60 off-peak and $0.30/$1.20 at peak, with cache hits at $0.006 — the lowest price per point of coding competence on this list.

The caveat is the harness. DeepSeek’s 74.2 is on mini-SWE-agent; its own harness gives 72.6 and Claude Code 69.8. In Z.ai’s table, which uses the harder Terminal-Bench 4.0, it scores 31.2. Treat it as an excellent, cheap default for high-volume agent traffic — and test it on your tasks before betting a fleet on it.

Strengths: lowest price here; MIT license; 1M context with a small KV cache; images accepted as input. Watch out: the headline number is vendor-reported on one harness; the Terminal-Bench 4.0 gap to Claude is large; peak-hour pricing doubles the rate.

What didn’t make the list

• Grok 4.7 — $2/$6 with a 500K context and Terminal-Bench 4.0 of 37.6%. It’s a big jump over Grok 4.6, but well behind Sonnet 5.5’s 70.6% at the same input price (the harnesses differ), and the context is half the 1M windows that are now standard.
• MiMo-V2.6-Pro — open-weight 1.02T MoE under MIT with DeepSWE v1.1 71.9, audio and video input, and $0.435/$0.87 pricing. A strong open alternative; the same vendor-reported caveat applies.
• Mistral Large 4 — a 1.05T open-weight preview with strong cyber and legal results but coding scores behind the Chinese leaders (DeepSWE v1.1 61.7%). Weights are due at the end of October.
• Gemini 4 Argon — DeepSWE v1.1 77.9% and $2/$10 introductory pricing, but limited to trusted defenders. Revisit when it is generally available.
• Sakana Fugu — an orchestration system routing across a model pool rather than a model; worth a look if you’re open to that architecture.
• Kimi K3, Qwen3.8-Max — the 2.8T and 2.4T open-weight giants; practical adoption is limited by hardware and licenses.

A note on volatility: this list reflects benchmarks and pricing as of October 12, 2026.

Changelog

• 2026-10-12 — October update: Opus 5.5 takes the top slot from Fable 5; Sonnet 5.5 and GPT-6.1 Sol replace Sonnet 5 and GPT-5.6; Fable 5.1 moves to the long-horizon slot; GLM-5.3 replaces GLM-5.2; DeepSeek V4.1-Flash replaces Kimi K2.7 Code as the budget pick.
• 2026-07-29 — July update: Opus 5 replaces Opus 4.8; GPT-5.6 GA replaces GPT-5.5; added Kimi K3 and Inkling to honorable mentions.
• 2026-07-02 — Initial ranking published: Fable 5 · Opus 4.8 · Sonnet 5 · GPT-5.5 · GLM-5.2 · Kimi K2.7 Code.

Frequently asked questions

Sources

• Anthropic — Introducing Claude Opus 5.5
• Anthropic — Introducing Claude Sonnet 5.5
• Anthropic — Claude Fable 5.1 and Claude Mythos 5.1
• OpenAI — Introducing GPT-6.1 Sol
• Z.ai — GLM-5.3: Frontier Coding with Emergent Cyber Capabilities
• DeepSeek API Docs — DeepSeek-V4.1-Flash