Best LLM for Coding: Mythos Leads, Fable Ships

As of July 2026, Claude Mythos 5 (80.8, Supported) leads the coding ranking. Claude Fable 5 sits a quarter-point behind and is the practical public pick because Mythos access remains restricted. GPT-5.6 Sol ranks third, but its Estimated label matters: the score is useful, not equally well supported.

BenchAlign v5 calibrates comparable coding evidence instead of averaging raw percentages from benchmarks with different difficulty and coverage. SWE-Rebench, SWE-bench Pro, LiveCodeBench, SWE-bench Verified, agent evaluations, and standardized external signals can contribute when model identity and protocol match. Missing cells widen uncertainty; they do not become zeroes. The table rebuilds from the coding leaderboard on every site build.

One newer display benchmark worth watching is React Native Evals. It does not affect the weighted coding rank today, but it fills a real coverage gap by testing framework-specific mobile app implementation work that generic repository and competitive-programming benchmarks do not capture well. If React Native or Expo-style product work matters in your stack, read the React Native Evals explainer alongside the main coding leaderboard.

Top coding models, ranked

Table 1 Download CSV

Rank	Model	Type	License	Score	Evidence
1	Claude Mythos 5	Reasoning	Proprietary	80.8	Supported
2	Claude Fable 5	Reasoning	Proprietary	80.6	Supported
3	GPT-5.6 Sol	Reasoning	Proprietary	78.9	Supported
4	Kimi K3	Reasoning	Pending	78	Supported
5	GPT-5.6 Luna	Reasoning	Proprietary	72.6	Supported
6	GPT-5.5	Reasoning	Proprietary	71.4	Supported
7	Claude Opus 4.8	Reasoning	Proprietary	70.6	Supported
8	Claude Opus 4.7	Non-Reasoning	Proprietary	68.4	Supported
9	Claude Sonnet 5	Reasoning	Proprietary	68.4	Supported
10	GPT-5.6 Terra	Reasoning	Proprietary	66	Supported
11	Muse Spark 1.1	Reasoning	Proprietary	65.1	Supported
12	GLM-5.2	Reasoning	Open Weight	65.1	Supported

The score sets the order. Supported and Estimated describe the evidence behind each position. Estimated rows remain ranked, but should be held more loosely.

The open-weight gap reopened

The old verified-only table made the open-weight field look almost tied with the proprietary leader. The unified v5 ranking does not support that conclusion. GLM-5.2 and Kimi K2.7 Code are the first open rows, around 62, while Mythos and Fable sit near 82. That is a real gap, not a rounding error.

The economics follow. Claude Opus 4.8 runs $5/$25 per million tokens. DeepSeek V4 Pro runs $0.435/$0.87 via API (roughly 11x cheaper on input and 28x cheaper on output), and you can self-host it. For agent loops that burn hundreds of millions of tokens, that difference decides the architecture.

What the scores hide

BenchLM's public HumanEval page currently exposes 3 exact sourced rows:

Table 2 Download CSV

Rank	Model	Score	Evidence
1	DeepSeek V4 Pro Base	76.8%	Provider exact
2	Soofi S 30B-A3B	73.8%	Provider exact
3	DeepSeek V4 Flash Base	69.5%	Provider exact

That table cannot support a broad claim about the best coding model. The visible rows are DeepSeek base configurations, and missing frontier rows are missing evidence rather than failures. HumanEval also tests function-level Python generation, not repository navigation, multi-file changes, or an agent recovering from a failed tool call. Use the HumanEval page as a coverage receipt, then use repository benchmarks and the calibrated coding rank for the decision.

The ranked column hides one more thing: evidence strength is not capability. GPT-5.6 Sol ranks third with an Estimated label, while Claude Opus 4.8 ranks fourth with Supported evidence. Pick Sol when you are comfortable taking the newer estimate; pick Opus when you want the position backed by a broader direct record.

The trade-off is latency. Reasoning models think before they respond, which can add seconds to minutes of first-answer latency. For autocomplete and interactive assistants, a fast non-reasoning model or a light reasoning tier is still the right call; save the heavy reasoning rows for multi-file bug fixes and agent sessions where quality dominates.

Best for specific coding tasks

Code completion and autocomplete

Short completions (under 50 tokens) don't require SWE-bench-level capability. The latency and cost profile matter more than marginal benchmark differences.

Best options: a fast Gemini tier for cost-sensitive high-volume work, or DeepSeek V4 Flash where every millisecond and cent counts. The frontier rank is the wrong optimization target for a 30-token completion.

Multi-file bug fixing and refactors

This is exactly what the repository benchmarks measure, and where the Claude 5 rows earn their lead.

Best option: Claude Fable 5 when maximum quality matters and the price is acceptable; Claude Opus 4.8 when you want a Supported frontier position at half Fable's input and output price.

Agentic coding (AI coding agents, long sessions)

Agentic coding burns tokens fast, so the cost column matters as much as the score column. Claude Opus 4.8 at $5/$25 adds up quickly in agent loops making hundreds of calls.

Best option: Kimi K2.7 Code for an evidence-backed open row, or Claude Opus 4.8 for a stronger Supported proprietary position. Fable remains the quality pick when the agent loop can absorb $10/$50 pricing.

Competitive programming

LiveCodeBench pulls fresh competitive programming problems continuously, so it stays contamination-resistant. It is one input, not a substitute for the calibrated coding rank; check the LiveCodeBench benchmark page for current per-model scores.

SQL and data tasks

No dedicated SQL benchmark exists at frontier level yet. Do not turn the coding rank into an SQL claim. For batch data pipelines, start with a cheaper model and run a schema-specific execution eval before paying for a frontier row.

Test generation

Test generation is underrepresented in benchmarks. Repository work is relevant because fixing bugs often requires regression tests, but no current pillar isolates test quality. Run mutation tests or hidden failure cases on your own codebase before choosing.

Open-source coding models

Table 3 Download CSV

Rank	Model	Type	License	Score	Evidence
1	GLM-5.2	Reasoning	Open Weight	65.1	Supported
2	GLM-5	Non-Reasoning	Open Weight	59.4	Estimated
3	GLM-5.1	Reasoning	Open Weight	57.5	Supported
4	Kimi K2.5	Non-Reasoning	Open Weight	57	Estimated
5	Qwen3.5-27B	Reasoning	Open Weight	55.7	Estimated
6	Hy3	Reasoning	Open Weight	55.3	Estimated

GLM-5.2 is the highest open-weight row, but its evidence label is Estimated. Kimi K2.7 Code sits just behind it with Supported evidence. That makes Kimi the cleaner recommendation for teams that would rather give up half a point than defend a thinner position.

These are still budget and control choices, not substitutes for the top capability tier. The operational advantage remains substantial: open weights permit self-hosting, fine-tuning, and predictable data boundaries. The capability trade-off is now visible again.

How to choose

Need the highest measured capability: Claude Mythos 5, if you can access it. Otherwise, Claude Fable 5.

Running an AI coding agent at scale: Claude Opus 4.8 for Supported frontier quality, or Kimi K2.7 Code when open weights and token economics dominate.

Claude ecosystem: Fable 5 for peak quality, Opus 4.8 for a better price-capability balance, Sonnet for volume work.

Budget-first coding: DeepSeek V4 Flash ($0.14/$0.28), MiniMax M3 ($0.30/$1.20), or GLM-5.1 ($1.40/$4.40) depending on whether you care more about price, open weights, or context window.

→ See the full coding leaderboard · Compare SWE-bench scores · LiveCodeBench details · React Native Evals explainer

---